1 ============================ 2 LINUX KERNEL MEMORY BARRIERS 3 ============================ 4 5By: David Howells <dhowells@redhat.com> 6 Paul E. McKenney <paulmck@linux.ibm.com> 7 Will Deacon <will.deacon@arm.com> 8 Peter Zijlstra <peterz@infradead.org> 9 10========== 11DISCLAIMER 12========== 13 14This document is not a specification; it is intentionally (for the sake of 15brevity) and unintentionally (due to being human) incomplete. This document is 16meant as a guide to using the various memory barriers provided by Linux, but 17in case of any doubt (and there are many) please ask. Some doubts may be 18resolved by referring to the formal memory consistency model and related 19documentation at tools/memory-model/. Nevertheless, even this memory 20model should be viewed as the collective opinion of its maintainers rather 21than as an infallible oracle. 22 23To repeat, this document is not a specification of what Linux expects from 24hardware. 25 26The purpose of this document is twofold: 27 28 (1) to specify the minimum functionality that one can rely on for any 29 particular barrier, and 30 31 (2) to provide a guide as to how to use the barriers that are available. 32 33Note that an architecture can provide more than the minimum requirement 34for any particular barrier, but if the architecture provides less than 35that, that architecture is incorrect. 36 37Note also that it is possible that a barrier may be a no-op for an 38architecture because the way that arch works renders an explicit barrier 39unnecessary in that case. 40 41 42======== 43CONTENTS 44======== 45 46 (*) Abstract memory access model. 47 48 - Device operations. 49 - Guarantees. 50 51 (*) What are memory barriers? 52 53 - Varieties of memory barrier. 54 - What may not be assumed about memory barriers? 55 - Data dependency barriers (historical). 56 - Control dependencies. 57 - SMP barrier pairing. 58 - Examples of memory barrier sequences. 59 - Read memory barriers vs load speculation. 60 - Multicopy atomicity. 61 62 (*) Explicit kernel barriers. 63 64 - Compiler barrier. 65 - CPU memory barriers. 66 67 (*) Implicit kernel memory barriers. 68 69 - Lock acquisition functions. 70 - Interrupt disabling functions. 71 - Sleep and wake-up functions. 72 - Miscellaneous functions. 73 74 (*) Inter-CPU acquiring barrier effects. 75 76 - Acquires vs memory accesses. 77 78 (*) Where are memory barriers needed? 79 80 - Interprocessor interaction. 81 - Atomic operations. 82 - Accessing devices. 83 - Interrupts. 84 85 (*) Kernel I/O barrier effects. 86 87 (*) Assumed minimum execution ordering model. 88 89 (*) The effects of the cpu cache. 90 91 - Cache coherency. 92 - Cache coherency vs DMA. 93 - Cache coherency vs MMIO. 94 95 (*) The things CPUs get up to. 96 97 - And then there's the Alpha. 98 - Virtual Machine Guests. 99 100 (*) Example uses. 101 102 - Circular buffers. 103 104 (*) References. 105 106 107============================ 108ABSTRACT MEMORY ACCESS MODEL 109============================ 110 111Consider the following abstract model of the system: 112 113 : : 114 : : 115 : : 116 +-------+ : +--------+ : +-------+ 117 | | : | | : | | 118 | | : | | : | | 119 | CPU 1 |<----->| Memory |<----->| CPU 2 | 120 | | : | | : | | 121 | | : | | : | | 122 +-------+ : +--------+ : +-------+ 123 ^ : ^ : ^ 124 | : | : | 125 | : | : | 126 | : v : | 127 | : +--------+ : | 128 | : | | : | 129 | : | | : | 130 +---------->| Device |<----------+ 131 : | | : 132 : | | : 133 : +--------+ : 134 : : 135 136Each CPU executes a program that generates memory access operations. In the 137abstract CPU, memory operation ordering is very relaxed, and a CPU may actually 138perform the memory operations in any order it likes, provided program causality 139appears to be maintained. Similarly, the compiler may also arrange the 140instructions it emits in any order it likes, provided it doesn't affect the 141apparent operation of the program. 142 143So in the above diagram, the effects of the memory operations performed by a 144CPU are perceived by the rest of the system as the operations cross the 145interface between the CPU and rest of the system (the dotted lines). 146 147 148For example, consider the following sequence of events: 149 150 CPU 1 CPU 2 151 =============== =============== 152 { A == 1; B == 2 } 153 A = 3; x = B; 154 B = 4; y = A; 155 156The set of accesses as seen by the memory system in the middle can be arranged 157in 24 different combinations: 158 159 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4 160 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3 161 STORE A=3, y=LOAD A->3, STORE B=4, x=LOAD B->4 162 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4 163 STORE A=3, x=LOAD B->2, STORE B=4, y=LOAD A->3 164 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4 165 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4 166 STORE B=4, ... 167 ... 168 169and can thus result in four different combinations of values: 170 171 x == 2, y == 1 172 x == 2, y == 3 173 x == 4, y == 1 174 x == 4, y == 3 175 176 177Furthermore, the stores committed by a CPU to the memory system may not be 178perceived by the loads made by another CPU in the same order as the stores were 179committed. 180 181 182As a further example, consider this sequence of events: 183 184 CPU 1 CPU 2 185 =============== =============== 186 { A == 1, B == 2, C == 3, P == &A, Q == &C } 187 B = 4; Q = P; 188 P = &B D = *Q; 189 190There is an obvious data dependency here, as the value loaded into D depends on 191the address retrieved from P by CPU 2. At the end of the sequence, any of the 192following results are possible: 193 194 (Q == &A) and (D == 1) 195 (Q == &B) and (D == 2) 196 (Q == &B) and (D == 4) 197 198Note that CPU 2 will never try and load C into D because the CPU will load P 199into Q before issuing the load of *Q. 200 201 202DEVICE OPERATIONS 203----------------- 204 205Some devices present their control interfaces as collections of memory 206locations, but the order in which the control registers are accessed is very 207important. For instance, imagine an ethernet card with a set of internal 208registers that are accessed through an address port register (A) and a data 209port register (D). To read internal register 5, the following code might then 210be used: 211 212 *A = 5; 213 x = *D; 214 215but this might show up as either of the following two sequences: 216 217 STORE *A = 5, x = LOAD *D 218 x = LOAD *D, STORE *A = 5 219 220the second of which will almost certainly result in a malfunction, since it set 221the address _after_ attempting to read the register. 222 223 224GUARANTEES 225---------- 226 227There are some minimal guarantees that may be expected of a CPU: 228 229 (*) On any given CPU, dependent memory accesses will be issued in order, with 230 respect to itself. This means that for: 231 232 Q = READ_ONCE(P); D = READ_ONCE(*Q); 233 234 the CPU will issue the following memory operations: 235 236 Q = LOAD P, D = LOAD *Q 237 238 and always in that order. However, on DEC Alpha, READ_ONCE() also 239 emits a memory-barrier instruction, so that a DEC Alpha CPU will 240 instead issue the following memory operations: 241 242 Q = LOAD P, MEMORY_BARRIER, D = LOAD *Q, MEMORY_BARRIER 243 244 Whether on DEC Alpha or not, the READ_ONCE() also prevents compiler 245 mischief. 246 247 (*) Overlapping loads and stores within a particular CPU will appear to be 248 ordered within that CPU. This means that for: 249 250 a = READ_ONCE(*X); WRITE_ONCE(*X, b); 251 252 the CPU will only issue the following sequence of memory operations: 253 254 a = LOAD *X, STORE *X = b 255 256 And for: 257 258 WRITE_ONCE(*X, c); d = READ_ONCE(*X); 259 260 the CPU will only issue: 261 262 STORE *X = c, d = LOAD *X 263 264 (Loads and stores overlap if they are targeted at overlapping pieces of 265 memory). 266 267And there are a number of things that _must_ or _must_not_ be assumed: 268 269 (*) It _must_not_ be assumed that the compiler will do what you want 270 with memory references that are not protected by READ_ONCE() and 271 WRITE_ONCE(). Without them, the compiler is within its rights to 272 do all sorts of "creative" transformations, which are covered in 273 the COMPILER BARRIER section. 274 275 (*) It _must_not_ be assumed that independent loads and stores will be issued 276 in the order given. This means that for: 277 278 X = *A; Y = *B; *D = Z; 279 280 we may get any of the following sequences: 281 282 X = LOAD *A, Y = LOAD *B, STORE *D = Z 283 X = LOAD *A, STORE *D = Z, Y = LOAD *B 284 Y = LOAD *B, X = LOAD *A, STORE *D = Z 285 Y = LOAD *B, STORE *D = Z, X = LOAD *A 286 STORE *D = Z, X = LOAD *A, Y = LOAD *B 287 STORE *D = Z, Y = LOAD *B, X = LOAD *A 288 289 (*) It _must_ be assumed that overlapping memory accesses may be merged or 290 discarded. This means that for: 291 292 X = *A; Y = *(A + 4); 293 294 we may get any one of the following sequences: 295 296 X = LOAD *A; Y = LOAD *(A + 4); 297 Y = LOAD *(A + 4); X = LOAD *A; 298 {X, Y} = LOAD {*A, *(A + 4) }; 299 300 And for: 301 302 *A = X; *(A + 4) = Y; 303 304 we may get any of: 305 306 STORE *A = X; STORE *(A + 4) = Y; 307 STORE *(A + 4) = Y; STORE *A = X; 308 STORE {*A, *(A + 4) } = {X, Y}; 309 310And there are anti-guarantees: 311 312 (*) These guarantees do not apply to bitfields, because compilers often 313 generate code to modify these using non-atomic read-modify-write 314 sequences. Do not attempt to use bitfields to synchronize parallel 315 algorithms. 316 317 (*) Even in cases where bitfields are protected by locks, all fields 318 in a given bitfield must be protected by one lock. If two fields 319 in a given bitfield are protected by different locks, the compiler's 320 non-atomic read-modify-write sequences can cause an update to one 321 field to corrupt the value of an adjacent field. 322 323 (*) These guarantees apply only to properly aligned and sized scalar 324 variables. "Properly sized" currently means variables that are 325 the same size as "char", "short", "int" and "long". "Properly 326 aligned" means the natural alignment, thus no constraints for 327 "char", two-byte alignment for "short", four-byte alignment for 328 "int", and either four-byte or eight-byte alignment for "long", 329 on 32-bit and 64-bit systems, respectively. Note that these 330 guarantees were introduced into the C11 standard, so beware when 331 using older pre-C11 compilers (for example, gcc 4.6). The portion 332 of the standard containing this guarantee is Section 3.14, which 333 defines "memory location" as follows: 334 335 memory location 336 either an object of scalar type, or a maximal sequence 337 of adjacent bit-fields all having nonzero width 338 339 NOTE 1: Two threads of execution can update and access 340 separate memory locations without interfering with 341 each other. 342 343 NOTE 2: A bit-field and an adjacent non-bit-field member 344 are in separate memory locations. The same applies 345 to two bit-fields, if one is declared inside a nested 346 structure declaration and the other is not, or if the two 347 are separated by a zero-length bit-field declaration, 348 or if they are separated by a non-bit-field member 349 declaration. It is not safe to concurrently update two 350 bit-fields in the same structure if all members declared 351 between them are also bit-fields, no matter what the 352 sizes of those intervening bit-fields happen to be. 353 354 355========================= 356WHAT ARE MEMORY BARRIERS? 357========================= 358 359As can be seen above, independent memory operations are effectively performed 360in random order, but this can be a problem for CPU-CPU interaction and for I/O. 361What is required is some way of intervening to instruct the compiler and the 362CPU to restrict the order. 363 364Memory barriers are such interventions. They impose a perceived partial 365ordering over the memory operations on either side of the barrier. 366 367Such enforcement is important because the CPUs and other devices in a system 368can use a variety of tricks to improve performance, including reordering, 369deferral and combination of memory operations; speculative loads; speculative 370branch prediction and various types of caching. Memory barriers are used to 371override or suppress these tricks, allowing the code to sanely control the 372interaction of multiple CPUs and/or devices. 373 374 375VARIETIES OF MEMORY BARRIER 376--------------------------- 377 378Memory barriers come in four basic varieties: 379 380 (1) Write (or store) memory barriers. 381 382 A write memory barrier gives a guarantee that all the STORE operations 383 specified before the barrier will appear to happen before all the STORE 384 operations specified after the barrier with respect to the other 385 components of the system. 386 387 A write barrier is a partial ordering on stores only; it is not required 388 to have any effect on loads. 389 390 A CPU can be viewed as committing a sequence of store operations to the 391 memory system as time progresses. All stores _before_ a write barrier 392 will occur _before_ all the stores after the write barrier. 393 394 [!] Note that write barriers should normally be paired with read or data 395 dependency barriers; see the "SMP barrier pairing" subsection. 396 397 398 (2) Data dependency barriers. 399 400 A data dependency barrier is a weaker form of read barrier. In the case 401 where two loads are performed such that the second depends on the result 402 of the first (eg: the first load retrieves the address to which the second 403 load will be directed), a data dependency barrier would be required to 404 make sure that the target of the second load is updated after the address 405 obtained by the first load is accessed. 406 407 A data dependency barrier is a partial ordering on interdependent loads 408 only; it is not required to have any effect on stores, independent loads 409 or overlapping loads. 410 411 As mentioned in (1), the other CPUs in the system can be viewed as 412 committing sequences of stores to the memory system that the CPU being 413 considered can then perceive. A data dependency barrier issued by the CPU 414 under consideration guarantees that for any load preceding it, if that 415 load touches one of a sequence of stores from another CPU, then by the 416 time the barrier completes, the effects of all the stores prior to that 417 touched by the load will be perceptible to any loads issued after the data 418 dependency barrier. 419 420 See the "Examples of memory barrier sequences" subsection for diagrams 421 showing the ordering constraints. 422 423 [!] Note that the first load really has to have a _data_ dependency and 424 not a control dependency. If the address for the second load is dependent 425 on the first load, but the dependency is through a conditional rather than 426 actually loading the address itself, then it's a _control_ dependency and 427 a full read barrier or better is required. See the "Control dependencies" 428 subsection for more information. 429 430 [!] Note that data dependency barriers should normally be paired with 431 write barriers; see the "SMP barrier pairing" subsection. 432 433 434 (3) Read (or load) memory barriers. 435 436 A read barrier is a data dependency barrier plus a guarantee that all the 437 LOAD operations specified before the barrier will appear to happen before 438 all the LOAD operations specified after the barrier with respect to the 439 other components of the system. 440 441 A read barrier is a partial ordering on loads only; it is not required to 442 have any effect on stores. 443 444 Read memory barriers imply data dependency barriers, and so can substitute 445 for them. 446 447 [!] Note that read barriers should normally be paired with write barriers; 448 see the "SMP barrier pairing" subsection. 449 450 451 (4) General memory barriers. 452 453 A general memory barrier gives a guarantee that all the LOAD and STORE 454 operations specified before the barrier will appear to happen before all 455 the LOAD and STORE operations specified after the barrier with respect to 456 the other components of the system. 457 458 A general memory barrier is a partial ordering over both loads and stores. 459 460 General memory barriers imply both read and write memory barriers, and so 461 can substitute for either. 462 463 464And a couple of implicit varieties: 465 466 (5) ACQUIRE operations. 467 468 This acts as a one-way permeable barrier. It guarantees that all memory 469 operations after the ACQUIRE operation will appear to happen after the 470 ACQUIRE operation with respect to the other components of the system. 471 ACQUIRE operations include LOCK operations and both smp_load_acquire() 472 and smp_cond_load_acquire() operations. 473 474 Memory operations that occur before an ACQUIRE operation may appear to 475 happen after it completes. 476 477 An ACQUIRE operation should almost always be paired with a RELEASE 478 operation. 479 480 481 (6) RELEASE operations. 482 483 This also acts as a one-way permeable barrier. It guarantees that all 484 memory operations before the RELEASE operation will appear to happen 485 before the RELEASE operation with respect to the other components of the 486 system. RELEASE operations include UNLOCK operations and 487 smp_store_release() operations. 488 489 Memory operations that occur after a RELEASE operation may appear to 490 happen before it completes. 491 492 The use of ACQUIRE and RELEASE operations generally precludes the need 493 for other sorts of memory barrier. In addition, a RELEASE+ACQUIRE pair is 494 -not- guaranteed to act as a full memory barrier. However, after an 495 ACQUIRE on a given variable, all memory accesses preceding any prior 496 RELEASE on that same variable are guaranteed to be visible. In other 497 words, within a given variable's critical section, all accesses of all 498 previous critical sections for that variable are guaranteed to have 499 completed. 500 501 This means that ACQUIRE acts as a minimal "acquire" operation and 502 RELEASE acts as a minimal "release" operation. 503 504A subset of the atomic operations described in atomic_t.txt have ACQUIRE and 505RELEASE variants in addition to fully-ordered and relaxed (no barrier 506semantics) definitions. For compound atomics performing both a load and a 507store, ACQUIRE semantics apply only to the load and RELEASE semantics apply 508only to the store portion of the operation. 509 510Memory barriers are only required where there's a possibility of interaction 511between two CPUs or between a CPU and a device. If it can be guaranteed that 512there won't be any such interaction in any particular piece of code, then 513memory barriers are unnecessary in that piece of code. 514 515 516Note that these are the _minimum_ guarantees. Different architectures may give 517more substantial guarantees, but they may _not_ be relied upon outside of arch 518specific code. 519 520 521WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS? 522---------------------------------------------- 523 524There are certain things that the Linux kernel memory barriers do not guarantee: 525 526 (*) There is no guarantee that any of the memory accesses specified before a 527 memory barrier will be _complete_ by the completion of a memory barrier 528 instruction; the barrier can be considered to draw a line in that CPU's 529 access queue that accesses of the appropriate type may not cross. 530 531 (*) There is no guarantee that issuing a memory barrier on one CPU will have 532 any direct effect on another CPU or any other hardware in the system. The 533 indirect effect will be the order in which the second CPU sees the effects 534 of the first CPU's accesses occur, but see the next point: 535 536 (*) There is no guarantee that a CPU will see the correct order of effects 537 from a second CPU's accesses, even _if_ the second CPU uses a memory 538 barrier, unless the first CPU _also_ uses a matching memory barrier (see 539 the subsection on "SMP Barrier Pairing"). 540 541 (*) There is no guarantee that some intervening piece of off-the-CPU 542 hardware[*] will not reorder the memory accesses. CPU cache coherency 543 mechanisms should propagate the indirect effects of a memory barrier 544 between CPUs, but might not do so in order. 545 546 [*] For information on bus mastering DMA and coherency please read: 547 548 Documentation/driver-api/pci/pci.rst 549 Documentation/DMA-API-HOWTO.txt 550 Documentation/DMA-API.txt 551 552 553DATA DEPENDENCY BARRIERS (HISTORICAL) 554------------------------------------- 555 556As of v4.15 of the Linux kernel, an smp_read_barrier_depends() was 557added to READ_ONCE(), which means that about the only people who 558need to pay attention to this section are those working on DEC Alpha 559architecture-specific code and those working on READ_ONCE() itself. 560For those who need it, and for those who are interested in the history, 561here is the story of data-dependency barriers. 562 563The usage requirements of data dependency barriers are a little subtle, and 564it's not always obvious that they're needed. To illustrate, consider the 565following sequence of events: 566 567 CPU 1 CPU 2 568 =============== =============== 569 { A == 1, B == 2, C == 3, P == &A, Q == &C } 570 B = 4; 571 <write barrier> 572 WRITE_ONCE(P, &B) 573 Q = READ_ONCE(P); 574 D = *Q; 575 576There's a clear data dependency here, and it would seem that by the end of the 577sequence, Q must be either &A or &B, and that: 578 579 (Q == &A) implies (D == 1) 580 (Q == &B) implies (D == 4) 581 582But! CPU 2's perception of P may be updated _before_ its perception of B, thus 583leading to the following situation: 584 585 (Q == &B) and (D == 2) ???? 586 587While this may seem like a failure of coherency or causality maintenance, it 588isn't, and this behaviour can be observed on certain real CPUs (such as the DEC 589Alpha). 590 591To deal with this, a data dependency barrier or better must be inserted 592between the address load and the data load: 593 594 CPU 1 CPU 2 595 =============== =============== 596 { A == 1, B == 2, C == 3, P == &A, Q == &C } 597 B = 4; 598 <write barrier> 599 WRITE_ONCE(P, &B); 600 Q = READ_ONCE(P); 601 <data dependency barrier> 602 D = *Q; 603 604This enforces the occurrence of one of the two implications, and prevents the 605third possibility from arising. 606 607 608[!] Note that this extremely counterintuitive situation arises most easily on 609machines with split caches, so that, for example, one cache bank processes 610even-numbered cache lines and the other bank processes odd-numbered cache 611lines. The pointer P might be stored in an odd-numbered cache line, and the 612variable B might be stored in an even-numbered cache line. Then, if the 613even-numbered bank of the reading CPU's cache is extremely busy while the 614odd-numbered bank is idle, one can see the new value of the pointer P (&B), 615but the old value of the variable B (2). 616 617 618A data-dependency barrier is not required to order dependent writes 619because the CPUs that the Linux kernel supports don't do writes 620until they are certain (1) that the write will actually happen, (2) 621of the location of the write, and (3) of the value to be written. 622But please carefully read the "CONTROL DEPENDENCIES" section and the 623Documentation/RCU/rcu_dereference.txt file: The compiler can and does 624break dependencies in a great many highly creative ways. 625 626 CPU 1 CPU 2 627 =============== =============== 628 { A == 1, B == 2, C = 3, P == &A, Q == &C } 629 B = 4; 630 <write barrier> 631 WRITE_ONCE(P, &B); 632 Q = READ_ONCE(P); 633 WRITE_ONCE(*Q, 5); 634 635Therefore, no data-dependency barrier is required to order the read into 636Q with the store into *Q. In other words, this outcome is prohibited, 637even without a data-dependency barrier: 638 639 (Q == &B) && (B == 4) 640 641Please note that this pattern should be rare. After all, the whole point 642of dependency ordering is to -prevent- writes to the data structure, along 643with the expensive cache misses associated with those writes. This pattern 644can be used to record rare error conditions and the like, and the CPUs' 645naturally occurring ordering prevents such records from being lost. 646 647 648Note well that the ordering provided by a data dependency is local to 649the CPU containing it. See the section on "Multicopy atomicity" for 650more information. 651 652 653The data dependency barrier is very important to the RCU system, 654for example. See rcu_assign_pointer() and rcu_dereference() in 655include/linux/rcupdate.h. This permits the current target of an RCU'd 656pointer to be replaced with a new modified target, without the replacement 657target appearing to be incompletely initialised. 658 659See also the subsection on "Cache Coherency" for a more thorough example. 660 661 662CONTROL DEPENDENCIES 663-------------------- 664 665Control dependencies can be a bit tricky because current compilers do 666not understand them. The purpose of this section is to help you prevent 667the compiler's ignorance from breaking your code. 668 669A load-load control dependency requires a full read memory barrier, not 670simply a data dependency barrier to make it work correctly. Consider the 671following bit of code: 672 673 q = READ_ONCE(a); 674 if (q) { 675 <data dependency barrier> /* BUG: No data dependency!!! */ 676 p = READ_ONCE(b); 677 } 678 679This will not have the desired effect because there is no actual data 680dependency, but rather a control dependency that the CPU may short-circuit 681by attempting to predict the outcome in advance, so that other CPUs see 682the load from b as having happened before the load from a. In such a 683case what's actually required is: 684 685 q = READ_ONCE(a); 686 if (q) { 687 <read barrier> 688 p = READ_ONCE(b); 689 } 690 691However, stores are not speculated. This means that ordering -is- provided 692for load-store control dependencies, as in the following example: 693 694 q = READ_ONCE(a); 695 if (q) { 696 WRITE_ONCE(b, 1); 697 } 698 699Control dependencies pair normally with other types of barriers. 700That said, please note that neither READ_ONCE() nor WRITE_ONCE() 701are optional! Without the READ_ONCE(), the compiler might combine the 702load from 'a' with other loads from 'a'. Without the WRITE_ONCE(), 703the compiler might combine the store to 'b' with other stores to 'b'. 704Either can result in highly counterintuitive effects on ordering. 705 706Worse yet, if the compiler is able to prove (say) that the value of 707variable 'a' is always non-zero, it would be well within its rights 708to optimize the original example by eliminating the "if" statement 709as follows: 710 711 q = a; 712 b = 1; /* BUG: Compiler and CPU can both reorder!!! */ 713 714So don't leave out the READ_ONCE(). 715 716It is tempting to try to enforce ordering on identical stores on both 717branches of the "if" statement as follows: 718 719 q = READ_ONCE(a); 720 if (q) { 721 barrier(); 722 WRITE_ONCE(b, 1); 723 do_something(); 724 } else { 725 barrier(); 726 WRITE_ONCE(b, 1); 727 do_something_else(); 728 } 729 730Unfortunately, current compilers will transform this as follows at high 731optimization levels: 732 733 q = READ_ONCE(a); 734 barrier(); 735 WRITE_ONCE(b, 1); /* BUG: No ordering vs. load from a!!! */ 736 if (q) { 737 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */ 738 do_something(); 739 } else { 740 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */ 741 do_something_else(); 742 } 743 744Now there is no conditional between the load from 'a' and the store to 745'b', which means that the CPU is within its rights to reorder them: 746The conditional is absolutely required, and must be present in the 747assembly code even after all compiler optimizations have been applied. 748Therefore, if you need ordering in this example, you need explicit 749memory barriers, for example, smp_store_release(): 750 751 q = READ_ONCE(a); 752 if (q) { 753 smp_store_release(&b, 1); 754 do_something(); 755 } else { 756 smp_store_release(&b, 1); 757 do_something_else(); 758 } 759 760In contrast, without explicit memory barriers, two-legged-if control 761ordering is guaranteed only when the stores differ, for example: 762 763 q = READ_ONCE(a); 764 if (q) { 765 WRITE_ONCE(b, 1); 766 do_something(); 767 } else { 768 WRITE_ONCE(b, 2); 769 do_something_else(); 770 } 771 772The initial READ_ONCE() is still required to prevent the compiler from 773proving the value of 'a'. 774 775In addition, you need to be careful what you do with the local variable 'q', 776otherwise the compiler might be able to guess the value and again remove 777the needed conditional. For example: 778 779 q = READ_ONCE(a); 780 if (q % MAX) { 781 WRITE_ONCE(b, 1); 782 do_something(); 783 } else { 784 WRITE_ONCE(b, 2); 785 do_something_else(); 786 } 787 788If MAX is defined to be 1, then the compiler knows that (q % MAX) is 789equal to zero, in which case the compiler is within its rights to 790transform the above code into the following: 791 792 q = READ_ONCE(a); 793 WRITE_ONCE(b, 2); 794 do_something_else(); 795 796Given this transformation, the CPU is not required to respect the ordering 797between the load from variable 'a' and the store to variable 'b'. It is 798tempting to add a barrier(), but this does not help. The conditional 799is gone, and the barrier won't bring it back. Therefore, if you are 800relying on this ordering, you should make sure that MAX is greater than 801one, perhaps as follows: 802 803 q = READ_ONCE(a); 804 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */ 805 if (q % MAX) { 806 WRITE_ONCE(b, 1); 807 do_something(); 808 } else { 809 WRITE_ONCE(b, 2); 810 do_something_else(); 811 } 812 813Please note once again that the stores to 'b' differ. If they were 814identical, as noted earlier, the compiler could pull this store outside 815of the 'if' statement. 816 817You must also be careful not to rely too much on boolean short-circuit 818evaluation. Consider this example: 819 820 q = READ_ONCE(a); 821 if (q || 1 > 0) 822 WRITE_ONCE(b, 1); 823 824Because the first condition cannot fault and the second condition is 825always true, the compiler can transform this example as following, 826defeating control dependency: 827 828 q = READ_ONCE(a); 829 WRITE_ONCE(b, 1); 830 831This example underscores the need to ensure that the compiler cannot 832out-guess your code. More generally, although READ_ONCE() does force 833the compiler to actually emit code for a given load, it does not force 834the compiler to use the results. 835 836In addition, control dependencies apply only to the then-clause and 837else-clause of the if-statement in question. In particular, it does 838not necessarily apply to code following the if-statement: 839 840 q = READ_ONCE(a); 841 if (q) { 842 WRITE_ONCE(b, 1); 843 } else { 844 WRITE_ONCE(b, 2); 845 } 846 WRITE_ONCE(c, 1); /* BUG: No ordering against the read from 'a'. */ 847 848It is tempting to argue that there in fact is ordering because the 849compiler cannot reorder volatile accesses and also cannot reorder 850the writes to 'b' with the condition. Unfortunately for this line 851of reasoning, the compiler might compile the two writes to 'b' as 852conditional-move instructions, as in this fanciful pseudo-assembly 853language: 854 855 ld r1,a 856 cmp r1,$0 857 cmov,ne r4,$1 858 cmov,eq r4,$2 859 st r4,b 860 st $1,c 861 862A weakly ordered CPU would have no dependency of any sort between the load 863from 'a' and the store to 'c'. The control dependencies would extend 864only to the pair of cmov instructions and the store depending on them. 865In short, control dependencies apply only to the stores in the then-clause 866and else-clause of the if-statement in question (including functions 867invoked by those two clauses), not to code following that if-statement. 868 869 870Note well that the ordering provided by a control dependency is local 871to the CPU containing it. See the section on "Multicopy atomicity" 872for more information. 873 874 875In summary: 876 877 (*) Control dependencies can order prior loads against later stores. 878 However, they do -not- guarantee any other sort of ordering: 879 Not prior loads against later loads, nor prior stores against 880 later anything. If you need these other forms of ordering, 881 use smp_rmb(), smp_wmb(), or, in the case of prior stores and 882 later loads, smp_mb(). 883 884 (*) If both legs of the "if" statement begin with identical stores to 885 the same variable, then those stores must be ordered, either by 886 preceding both of them with smp_mb() or by using smp_store_release() 887 to carry out the stores. Please note that it is -not- sufficient 888 to use barrier() at beginning of each leg of the "if" statement 889 because, as shown by the example above, optimizing compilers can 890 destroy the control dependency while respecting the letter of the 891 barrier() law. 892 893 (*) Control dependencies require at least one run-time conditional 894 between the prior load and the subsequent store, and this 895 conditional must involve the prior load. If the compiler is able 896 to optimize the conditional away, it will have also optimized 897 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE() 898 can help to preserve the needed conditional. 899 900 (*) Control dependencies require that the compiler avoid reordering the 901 dependency into nonexistence. Careful use of READ_ONCE() or 902 atomic{,64}_read() can help to preserve your control dependency. 903 Please see the COMPILER BARRIER section for more information. 904 905 (*) Control dependencies apply only to the then-clause and else-clause 906 of the if-statement containing the control dependency, including 907 any functions that these two clauses call. Control dependencies 908 do -not- apply to code following the if-statement containing the 909 control dependency. 910 911 (*) Control dependencies pair normally with other types of barriers. 912 913 (*) Control dependencies do -not- provide multicopy atomicity. If you 914 need all the CPUs to see a given store at the same time, use smp_mb(). 915 916 (*) Compilers do not understand control dependencies. It is therefore 917 your job to ensure that they do not break your code. 918 919 920SMP BARRIER PAIRING 921------------------- 922 923When dealing with CPU-CPU interactions, certain types of memory barrier should 924always be paired. A lack of appropriate pairing is almost certainly an error. 925 926General barriers pair with each other, though they also pair with most 927other types of barriers, albeit without multicopy atomicity. An acquire 928barrier pairs with a release barrier, but both may also pair with other 929barriers, including of course general barriers. A write barrier pairs 930with a data dependency barrier, a control dependency, an acquire barrier, 931a release barrier, a read barrier, or a general barrier. Similarly a 932read barrier, control dependency, or a data dependency barrier pairs 933with a write barrier, an acquire barrier, a release barrier, or a 934general barrier: 935 936 CPU 1 CPU 2 937 =============== =============== 938 WRITE_ONCE(a, 1); 939 <write barrier> 940 WRITE_ONCE(b, 2); x = READ_ONCE(b); 941 <read barrier> 942 y = READ_ONCE(a); 943 944Or: 945 946 CPU 1 CPU 2 947 =============== =============================== 948 a = 1; 949 <write barrier> 950 WRITE_ONCE(b, &a); x = READ_ONCE(b); 951 <data dependency barrier> 952 y = *x; 953 954Or even: 955 956 CPU 1 CPU 2 957 =============== =============================== 958 r1 = READ_ONCE(y); 959 <general barrier> 960 WRITE_ONCE(x, 1); if (r2 = READ_ONCE(x)) { 961 <implicit control dependency> 962 WRITE_ONCE(y, 1); 963 } 964 965 assert(r1 == 0 || r2 == 0); 966 967Basically, the read barrier always has to be there, even though it can be of 968the "weaker" type. 969 970[!] Note that the stores before the write barrier would normally be expected to 971match the loads after the read barrier or the data dependency barrier, and vice 972versa: 973 974 CPU 1 CPU 2 975 =================== =================== 976 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c); 977 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d); 978 <write barrier> \ <read barrier> 979 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a); 980 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b); 981 982 983EXAMPLES OF MEMORY BARRIER SEQUENCES 984------------------------------------ 985 986Firstly, write barriers act as partial orderings on store operations. 987Consider the following sequence of events: 988 989 CPU 1 990 ======================= 991 STORE A = 1 992 STORE B = 2 993 STORE C = 3 994 <write barrier> 995 STORE D = 4 996 STORE E = 5 997 998This sequence of events is committed to the memory coherence system in an order 999that the rest of the system might perceive as the unordered set of { STORE A, 1000STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E 1001}: 1002 1003 +-------+ : : 1004 | | +------+ 1005 | |------>| C=3 | } /\ 1006 | | : +------+ }----- \ -----> Events perceptible to 1007 | | : | A=1 | } \/ the rest of the system 1008 | | : +------+ } 1009 | CPU 1 | : | B=2 | } 1010 | | +------+ } 1011 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier 1012 | | +------+ } requires all stores prior to the 1013 | | : | E=5 | } barrier to be committed before 1014 | | : +------+ } further stores may take place 1015 | |------>| D=4 | } 1016 | | +------+ 1017 +-------+ : : 1018 | 1019 | Sequence in which stores are committed to the 1020 | memory system by CPU 1 1021 V 1022 1023 1024Secondly, data dependency barriers act as partial orderings on data-dependent 1025loads. Consider the following sequence of events: 1026 1027 CPU 1 CPU 2 1028 ======================= ======================= 1029 { B = 7; X = 9; Y = 8; C = &Y } 1030 STORE A = 1 1031 STORE B = 2 1032 <write barrier> 1033 STORE C = &B LOAD X 1034 STORE D = 4 LOAD C (gets &B) 1035 LOAD *C (reads B) 1036 1037Without intervention, CPU 2 may perceive the events on CPU 1 in some 1038effectively random order, despite the write barrier issued by CPU 1: 1039 1040 +-------+ : : : : 1041 | | +------+ +-------+ | Sequence of update 1042 | |------>| B=2 |----- --->| Y->8 | | of perception on 1043 | | : +------+ \ +-------+ | CPU 2 1044 | CPU 1 | : | A=1 | \ --->| C->&Y | V 1045 | | +------+ | +-------+ 1046 | | wwwwwwwwwwwwwwww | : : 1047 | | +------+ | : : 1048 | | : | C=&B |--- | : : +-------+ 1049 | | : +------+ \ | +-------+ | | 1050 | |------>| D=4 | ----------->| C->&B |------>| | 1051 | | +------+ | +-------+ | | 1052 +-------+ : : | : : | | 1053 | : : | | 1054 | : : | CPU 2 | 1055 | +-------+ | | 1056 Apparently incorrect ---> | | B->7 |------>| | 1057 perception of B (!) | +-------+ | | 1058 | : : | | 1059 | +-------+ | | 1060 The load of X holds ---> \ | X->9 |------>| | 1061 up the maintenance \ +-------+ | | 1062 of coherence of B ----->| B->2 | +-------+ 1063 +-------+ 1064 : : 1065 1066 1067In the above example, CPU 2 perceives that B is 7, despite the load of *C 1068(which would be B) coming after the LOAD of C. 1069 1070If, however, a data dependency barrier were to be placed between the load of C 1071and the load of *C (ie: B) on CPU 2: 1072 1073 CPU 1 CPU 2 1074 ======================= ======================= 1075 { B = 7; X = 9; Y = 8; C = &Y } 1076 STORE A = 1 1077 STORE B = 2 1078 <write barrier> 1079 STORE C = &B LOAD X 1080 STORE D = 4 LOAD C (gets &B) 1081 <data dependency barrier> 1082 LOAD *C (reads B) 1083 1084then the following will occur: 1085 1086 +-------+ : : : : 1087 | | +------+ +-------+ 1088 | |------>| B=2 |----- --->| Y->8 | 1089 | | : +------+ \ +-------+ 1090 | CPU 1 | : | A=1 | \ --->| C->&Y | 1091 | | +------+ | +-------+ 1092 | | wwwwwwwwwwwwwwww | : : 1093 | | +------+ | : : 1094 | | : | C=&B |--- | : : +-------+ 1095 | | : +------+ \ | +-------+ | | 1096 | |------>| D=4 | ----------->| C->&B |------>| | 1097 | | +------+ | +-------+ | | 1098 +-------+ : : | : : | | 1099 | : : | | 1100 | : : | CPU 2 | 1101 | +-------+ | | 1102 | | X->9 |------>| | 1103 | +-------+ | | 1104 Makes sure all effects ---> \ ddddddddddddddddd | | 1105 prior to the store of C \ +-------+ | | 1106 are perceptible to ----->| B->2 |------>| | 1107 subsequent loads +-------+ | | 1108 : : +-------+ 1109 1110 1111And thirdly, a read barrier acts as a partial order on loads. Consider the 1112following sequence of events: 1113 1114 CPU 1 CPU 2 1115 ======================= ======================= 1116 { A = 0, B = 9 } 1117 STORE A=1 1118 <write barrier> 1119 STORE B=2 1120 LOAD B 1121 LOAD A 1122 1123Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in 1124some effectively random order, despite the write barrier issued by CPU 1: 1125 1126 +-------+ : : : : 1127 | | +------+ +-------+ 1128 | |------>| A=1 |------ --->| A->0 | 1129 | | +------+ \ +-------+ 1130 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1131 | | +------+ | +-------+ 1132 | |------>| B=2 |--- | : : 1133 | | +------+ \ | : : +-------+ 1134 +-------+ : : \ | +-------+ | | 1135 ---------->| B->2 |------>| | 1136 | +-------+ | CPU 2 | 1137 | | A->0 |------>| | 1138 | +-------+ | | 1139 | : : +-------+ 1140 \ : : 1141 \ +-------+ 1142 ---->| A->1 | 1143 +-------+ 1144 : : 1145 1146 1147If, however, a read barrier were to be placed between the load of B and the 1148load of A on CPU 2: 1149 1150 CPU 1 CPU 2 1151 ======================= ======================= 1152 { A = 0, B = 9 } 1153 STORE A=1 1154 <write barrier> 1155 STORE B=2 1156 LOAD B 1157 <read barrier> 1158 LOAD A 1159 1160then the partial ordering imposed by CPU 1 will be perceived correctly by CPU 11612: 1162 1163 +-------+ : : : : 1164 | | +------+ +-------+ 1165 | |------>| A=1 |------ --->| A->0 | 1166 | | +------+ \ +-------+ 1167 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1168 | | +------+ | +-------+ 1169 | |------>| B=2 |--- | : : 1170 | | +------+ \ | : : +-------+ 1171 +-------+ : : \ | +-------+ | | 1172 ---------->| B->2 |------>| | 1173 | +-------+ | CPU 2 | 1174 | : : | | 1175 | : : | | 1176 At this point the read ----> \ rrrrrrrrrrrrrrrrr | | 1177 barrier causes all effects \ +-------+ | | 1178 prior to the storage of B ---->| A->1 |------>| | 1179 to be perceptible to CPU 2 +-------+ | | 1180 : : +-------+ 1181 1182 1183To illustrate this more completely, consider what could happen if the code 1184contained a load of A either side of the read barrier: 1185 1186 CPU 1 CPU 2 1187 ======================= ======================= 1188 { A = 0, B = 9 } 1189 STORE A=1 1190 <write barrier> 1191 STORE B=2 1192 LOAD B 1193 LOAD A [first load of A] 1194 <read barrier> 1195 LOAD A [second load of A] 1196 1197Even though the two loads of A both occur after the load of B, they may both 1198come up with different values: 1199 1200 +-------+ : : : : 1201 | | +------+ +-------+ 1202 | |------>| A=1 |------ --->| A->0 | 1203 | | +------+ \ +-------+ 1204 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1205 | | +------+ | +-------+ 1206 | |------>| B=2 |--- | : : 1207 | | +------+ \ | : : +-------+ 1208 +-------+ : : \ | +-------+ | | 1209 ---------->| B->2 |------>| | 1210 | +-------+ | CPU 2 | 1211 | : : | | 1212 | : : | | 1213 | +-------+ | | 1214 | | A->0 |------>| 1st | 1215 | +-------+ | | 1216 At this point the read ----> \ rrrrrrrrrrrrrrrrr | | 1217 barrier causes all effects \ +-------+ | | 1218 prior to the storage of B ---->| A->1 |------>| 2nd | 1219 to be perceptible to CPU 2 +-------+ | | 1220 : : +-------+ 1221 1222 1223But it may be that the update to A from CPU 1 becomes perceptible to CPU 2 1224before the read barrier completes anyway: 1225 1226 +-------+ : : : : 1227 | | +------+ +-------+ 1228 | |------>| A=1 |------ --->| A->0 | 1229 | | +------+ \ +-------+ 1230 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1231 | | +------+ | +-------+ 1232 | |------>| B=2 |--- | : : 1233 | | +------+ \ | : : +-------+ 1234 +-------+ : : \ | +-------+ | | 1235 ---------->| B->2 |------>| | 1236 | +-------+ | CPU 2 | 1237 | : : | | 1238 \ : : | | 1239 \ +-------+ | | 1240 ---->| A->1 |------>| 1st | 1241 +-------+ | | 1242 rrrrrrrrrrrrrrrrr | | 1243 +-------+ | | 1244 | A->1 |------>| 2nd | 1245 +-------+ | | 1246 : : +-------+ 1247 1248 1249The guarantee is that the second load will always come up with A == 1 if the 1250load of B came up with B == 2. No such guarantee exists for the first load of 1251A; that may come up with either A == 0 or A == 1. 1252 1253 1254READ MEMORY BARRIERS VS LOAD SPECULATION 1255---------------------------------------- 1256 1257Many CPUs speculate with loads: that is they see that they will need to load an 1258item from memory, and they find a time where they're not using the bus for any 1259other loads, and so do the load in advance - even though they haven't actually 1260got to that point in the instruction execution flow yet. This permits the 1261actual load instruction to potentially complete immediately because the CPU 1262already has the value to hand. 1263 1264It may turn out that the CPU didn't actually need the value - perhaps because a 1265branch circumvented the load - in which case it can discard the value or just 1266cache it for later use. 1267 1268Consider: 1269 1270 CPU 1 CPU 2 1271 ======================= ======================= 1272 LOAD B 1273 DIVIDE } Divide instructions generally 1274 DIVIDE } take a long time to perform 1275 LOAD A 1276 1277Which might appear as this: 1278 1279 : : +-------+ 1280 +-------+ | | 1281 --->| B->2 |------>| | 1282 +-------+ | CPU 2 | 1283 : :DIVIDE | | 1284 +-------+ | | 1285 The CPU being busy doing a ---> --->| A->0 |~~~~ | | 1286 division speculates on the +-------+ ~ | | 1287 LOAD of A : : ~ | | 1288 : :DIVIDE | | 1289 : : ~ | | 1290 Once the divisions are complete --> : : ~-->| | 1291 the CPU can then perform the : : | | 1292 LOAD with immediate effect : : +-------+ 1293 1294 1295Placing a read barrier or a data dependency barrier just before the second 1296load: 1297 1298 CPU 1 CPU 2 1299 ======================= ======================= 1300 LOAD B 1301 DIVIDE 1302 DIVIDE 1303 <read barrier> 1304 LOAD A 1305 1306will force any value speculatively obtained to be reconsidered to an extent 1307dependent on the type of barrier used. If there was no change made to the 1308speculated memory location, then the speculated value will just be used: 1309 1310 : : +-------+ 1311 +-------+ | | 1312 --->| B->2 |------>| | 1313 +-------+ | CPU 2 | 1314 : :DIVIDE | | 1315 +-------+ | | 1316 The CPU being busy doing a ---> --->| A->0 |~~~~ | | 1317 division speculates on the +-------+ ~ | | 1318 LOAD of A : : ~ | | 1319 : :DIVIDE | | 1320 : : ~ | | 1321 : : ~ | | 1322 rrrrrrrrrrrrrrrr~ | | 1323 : : ~ | | 1324 : : ~-->| | 1325 : : | | 1326 : : +-------+ 1327 1328 1329but if there was an update or an invalidation from another CPU pending, then 1330the speculation will be cancelled and the value reloaded: 1331 1332 : : +-------+ 1333 +-------+ | | 1334 --->| B->2 |------>| | 1335 +-------+ | CPU 2 | 1336 : :DIVIDE | | 1337 +-------+ | | 1338 The CPU being busy doing a ---> --->| A->0 |~~~~ | | 1339 division speculates on the +-------+ ~ | | 1340 LOAD of A : : ~ | | 1341 : :DIVIDE | | 1342 : : ~ | | 1343 : : ~ | | 1344 rrrrrrrrrrrrrrrrr | | 1345 +-------+ | | 1346 The speculation is discarded ---> --->| A->1 |------>| | 1347 and an updated value is +-------+ | | 1348 retrieved : : +-------+ 1349 1350 1351MULTICOPY ATOMICITY 1352-------------------- 1353 1354Multicopy atomicity is a deeply intuitive notion about ordering that is 1355not always provided by real computer systems, namely that a given store 1356becomes visible at the same time to all CPUs, or, alternatively, that all 1357CPUs agree on the order in which all stores become visible. However, 1358support of full multicopy atomicity would rule out valuable hardware 1359optimizations, so a weaker form called ``other multicopy atomicity'' 1360instead guarantees only that a given store becomes visible at the same 1361time to all -other- CPUs. The remainder of this document discusses this 1362weaker form, but for brevity will call it simply ``multicopy atomicity''. 1363 1364The following example demonstrates multicopy atomicity: 1365 1366 CPU 1 CPU 2 CPU 3 1367 ======================= ======================= ======================= 1368 { X = 0, Y = 0 } 1369 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1) 1370 <general barrier> <read barrier> 1371 STORE Y=r1 LOAD X 1372 1373Suppose that CPU 2's load from X returns 1, which it then stores to Y, 1374and CPU 3's load from Y returns 1. This indicates that CPU 1's store 1375to X precedes CPU 2's load from X and that CPU 2's store to Y precedes 1376CPU 3's load from Y. In addition, the memory barriers guarantee that 1377CPU 2 executes its load before its store, and CPU 3 loads from Y before 1378it loads from X. The question is then "Can CPU 3's load from X return 0?" 1379 1380Because CPU 3's load from X in some sense comes after CPU 2's load, it 1381is natural to expect that CPU 3's load from X must therefore return 1. 1382This expectation follows from multicopy atomicity: if a load executing 1383on CPU B follows a load from the same variable executing on CPU A (and 1384CPU A did not originally store the value which it read), then on 1385multicopy-atomic systems, CPU B's load must return either the same value 1386that CPU A's load did or some later value. However, the Linux kernel 1387does not require systems to be multicopy atomic. 1388 1389The use of a general memory barrier in the example above compensates 1390for any lack of multicopy atomicity. In the example, if CPU 2's load 1391from X returns 1 and CPU 3's load from Y returns 1, then CPU 3's load 1392from X must indeed also return 1. 1393 1394However, dependencies, read barriers, and write barriers are not always 1395able to compensate for non-multicopy atomicity. For example, suppose 1396that CPU 2's general barrier is removed from the above example, leaving 1397only the data dependency shown below: 1398 1399 CPU 1 CPU 2 CPU 3 1400 ======================= ======================= ======================= 1401 { X = 0, Y = 0 } 1402 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1) 1403 <data dependency> <read barrier> 1404 STORE Y=r1 LOAD X (reads 0) 1405 1406This substitution allows non-multicopy atomicity to run rampant: in 1407this example, it is perfectly legal for CPU 2's load from X to return 1, 1408CPU 3's load from Y to return 1, and its load from X to return 0. 1409 1410The key point is that although CPU 2's data dependency orders its load 1411and store, it does not guarantee to order CPU 1's store. Thus, if this 1412example runs on a non-multicopy-atomic system where CPUs 1 and 2 share a 1413store buffer or a level of cache, CPU 2 might have early access to CPU 1's 1414writes. General barriers are therefore required to ensure that all CPUs 1415agree on the combined order of multiple accesses. 1416 1417General barriers can compensate not only for non-multicopy atomicity, 1418but can also generate additional ordering that can ensure that -all- 1419CPUs will perceive the same order of -all- operations. In contrast, a 1420chain of release-acquire pairs do not provide this additional ordering, 1421which means that only those CPUs on the chain are guaranteed to agree 1422on the combined order of the accesses. For example, switching to C code 1423in deference to the ghost of Herman Hollerith: 1424 1425 int u, v, x, y, z; 1426 1427 void cpu0(void) 1428 { 1429 r0 = smp_load_acquire(&x); 1430 WRITE_ONCE(u, 1); 1431 smp_store_release(&y, 1); 1432 } 1433 1434 void cpu1(void) 1435 { 1436 r1 = smp_load_acquire(&y); 1437 r4 = READ_ONCE(v); 1438 r5 = READ_ONCE(u); 1439 smp_store_release(&z, 1); 1440 } 1441 1442 void cpu2(void) 1443 { 1444 r2 = smp_load_acquire(&z); 1445 smp_store_release(&x, 1); 1446 } 1447 1448 void cpu3(void) 1449 { 1450 WRITE_ONCE(v, 1); 1451 smp_mb(); 1452 r3 = READ_ONCE(u); 1453 } 1454 1455Because cpu0(), cpu1(), and cpu2() participate in a chain of 1456smp_store_release()/smp_load_acquire() pairs, the following outcome 1457is prohibited: 1458 1459 r0 == 1 && r1 == 1 && r2 == 1 1460 1461Furthermore, because of the release-acquire relationship between cpu0() 1462and cpu1(), cpu1() must see cpu0()'s writes, so that the following 1463outcome is prohibited: 1464 1465 r1 == 1 && r5 == 0 1466 1467However, the ordering provided by a release-acquire chain is local 1468to the CPUs participating in that chain and does not apply to cpu3(), 1469at least aside from stores. Therefore, the following outcome is possible: 1470 1471 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 1472 1473As an aside, the following outcome is also possible: 1474 1475 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 && r5 == 1 1476 1477Although cpu0(), cpu1(), and cpu2() will see their respective reads and 1478writes in order, CPUs not involved in the release-acquire chain might 1479well disagree on the order. This disagreement stems from the fact that 1480the weak memory-barrier instructions used to implement smp_load_acquire() 1481and smp_store_release() are not required to order prior stores against 1482subsequent loads in all cases. This means that cpu3() can see cpu0()'s 1483store to u as happening -after- cpu1()'s load from v, even though 1484both cpu0() and cpu1() agree that these two operations occurred in the 1485intended order. 1486 1487However, please keep in mind that smp_load_acquire() is not magic. 1488In particular, it simply reads from its argument with ordering. It does 1489-not- ensure that any particular value will be read. Therefore, the 1490following outcome is possible: 1491 1492 r0 == 0 && r1 == 0 && r2 == 0 && r5 == 0 1493 1494Note that this outcome can happen even on a mythical sequentially 1495consistent system where nothing is ever reordered. 1496 1497To reiterate, if your code requires full ordering of all operations, 1498use general barriers throughout. 1499 1500 1501======================== 1502EXPLICIT KERNEL BARRIERS 1503======================== 1504 1505The Linux kernel has a variety of different barriers that act at different 1506levels: 1507 1508 (*) Compiler barrier. 1509 1510 (*) CPU memory barriers. 1511 1512 1513COMPILER BARRIER 1514---------------- 1515 1516The Linux kernel has an explicit compiler barrier function that prevents the 1517compiler from moving the memory accesses either side of it to the other side: 1518 1519 barrier(); 1520 1521This is a general barrier -- there are no read-read or write-write 1522variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be 1523thought of as weak forms of barrier() that affect only the specific 1524accesses flagged by the READ_ONCE() or WRITE_ONCE(). 1525 1526The barrier() function has the following effects: 1527 1528 (*) Prevents the compiler from reordering accesses following the 1529 barrier() to precede any accesses preceding the barrier(). 1530 One example use for this property is to ease communication between 1531 interrupt-handler code and the code that was interrupted. 1532 1533 (*) Within a loop, forces the compiler to load the variables used 1534 in that loop's conditional on each pass through that loop. 1535 1536The READ_ONCE() and WRITE_ONCE() functions can prevent any number of 1537optimizations that, while perfectly safe in single-threaded code, can 1538be fatal in concurrent code. Here are some examples of these sorts 1539of optimizations: 1540 1541 (*) The compiler is within its rights to reorder loads and stores 1542 to the same variable, and in some cases, the CPU is within its 1543 rights to reorder loads to the same variable. This means that 1544 the following code: 1545 1546 a[0] = x; 1547 a[1] = x; 1548 1549 Might result in an older value of x stored in a[1] than in a[0]. 1550 Prevent both the compiler and the CPU from doing this as follows: 1551 1552 a[0] = READ_ONCE(x); 1553 a[1] = READ_ONCE(x); 1554 1555 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for 1556 accesses from multiple CPUs to a single variable. 1557 1558 (*) The compiler is within its rights to merge successive loads from 1559 the same variable. Such merging can cause the compiler to "optimize" 1560 the following code: 1561 1562 while (tmp = a) 1563 do_something_with(tmp); 1564 1565 into the following code, which, although in some sense legitimate 1566 for single-threaded code, is almost certainly not what the developer 1567 intended: 1568 1569 if (tmp = a) 1570 for (;;) 1571 do_something_with(tmp); 1572 1573 Use READ_ONCE() to prevent the compiler from doing this to you: 1574 1575 while (tmp = READ_ONCE(a)) 1576 do_something_with(tmp); 1577 1578 (*) The compiler is within its rights to reload a variable, for example, 1579 in cases where high register pressure prevents the compiler from 1580 keeping all data of interest in registers. The compiler might 1581 therefore optimize the variable 'tmp' out of our previous example: 1582 1583 while (tmp = a) 1584 do_something_with(tmp); 1585 1586 This could result in the following code, which is perfectly safe in 1587 single-threaded code, but can be fatal in concurrent code: 1588 1589 while (a) 1590 do_something_with(a); 1591 1592 For example, the optimized version of this code could result in 1593 passing a zero to do_something_with() in the case where the variable 1594 a was modified by some other CPU between the "while" statement and 1595 the call to do_something_with(). 1596 1597 Again, use READ_ONCE() to prevent the compiler from doing this: 1598 1599 while (tmp = READ_ONCE(a)) 1600 do_something_with(tmp); 1601 1602 Note that if the compiler runs short of registers, it might save 1603 tmp onto the stack. The overhead of this saving and later restoring 1604 is why compilers reload variables. Doing so is perfectly safe for 1605 single-threaded code, so you need to tell the compiler about cases 1606 where it is not safe. 1607 1608 (*) The compiler is within its rights to omit a load entirely if it knows 1609 what the value will be. For example, if the compiler can prove that 1610 the value of variable 'a' is always zero, it can optimize this code: 1611 1612 while (tmp = a) 1613 do_something_with(tmp); 1614 1615 Into this: 1616 1617 do { } while (0); 1618 1619 This transformation is a win for single-threaded code because it 1620 gets rid of a load and a branch. The problem is that the compiler 1621 will carry out its proof assuming that the current CPU is the only 1622 one updating variable 'a'. If variable 'a' is shared, then the 1623 compiler's proof will be erroneous. Use READ_ONCE() to tell the 1624 compiler that it doesn't know as much as it thinks it does: 1625 1626 while (tmp = READ_ONCE(a)) 1627 do_something_with(tmp); 1628 1629 But please note that the compiler is also closely watching what you 1630 do with the value after the READ_ONCE(). For example, suppose you 1631 do the following and MAX is a preprocessor macro with the value 1: 1632 1633 while ((tmp = READ_ONCE(a)) % MAX) 1634 do_something_with(tmp); 1635 1636 Then the compiler knows that the result of the "%" operator applied 1637 to MAX will always be zero, again allowing the compiler to optimize 1638 the code into near-nonexistence. (It will still load from the 1639 variable 'a'.) 1640 1641 (*) Similarly, the compiler is within its rights to omit a store entirely 1642 if it knows that the variable already has the value being stored. 1643 Again, the compiler assumes that the current CPU is the only one 1644 storing into the variable, which can cause the compiler to do the 1645 wrong thing for shared variables. For example, suppose you have 1646 the following: 1647 1648 a = 0; 1649 ... Code that does not store to variable a ... 1650 a = 0; 1651 1652 The compiler sees that the value of variable 'a' is already zero, so 1653 it might well omit the second store. This would come as a fatal 1654 surprise if some other CPU might have stored to variable 'a' in the 1655 meantime. 1656 1657 Use WRITE_ONCE() to prevent the compiler from making this sort of 1658 wrong guess: 1659 1660 WRITE_ONCE(a, 0); 1661 ... Code that does not store to variable a ... 1662 WRITE_ONCE(a, 0); 1663 1664 (*) The compiler is within its rights to reorder memory accesses unless 1665 you tell it not to. For example, consider the following interaction 1666 between process-level code and an interrupt handler: 1667 1668 void process_level(void) 1669 { 1670 msg = get_message(); 1671 flag = true; 1672 } 1673 1674 void interrupt_handler(void) 1675 { 1676 if (flag) 1677 process_message(msg); 1678 } 1679 1680 There is nothing to prevent the compiler from transforming 1681 process_level() to the following, in fact, this might well be a 1682 win for single-threaded code: 1683 1684 void process_level(void) 1685 { 1686 flag = true; 1687 msg = get_message(); 1688 } 1689 1690 If the interrupt occurs between these two statement, then 1691 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE() 1692 to prevent this as follows: 1693 1694 void process_level(void) 1695 { 1696 WRITE_ONCE(msg, get_message()); 1697 WRITE_ONCE(flag, true); 1698 } 1699 1700 void interrupt_handler(void) 1701 { 1702 if (READ_ONCE(flag)) 1703 process_message(READ_ONCE(msg)); 1704 } 1705 1706 Note that the READ_ONCE() and WRITE_ONCE() wrappers in 1707 interrupt_handler() are needed if this interrupt handler can itself 1708 be interrupted by something that also accesses 'flag' and 'msg', 1709 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE() 1710 and WRITE_ONCE() are not needed in interrupt_handler() other than 1711 for documentation purposes. (Note also that nested interrupts 1712 do not typically occur in modern Linux kernels, in fact, if an 1713 interrupt handler returns with interrupts enabled, you will get a 1714 WARN_ONCE() splat.) 1715 1716 You should assume that the compiler can move READ_ONCE() and 1717 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(), 1718 barrier(), or similar primitives. 1719 1720 This effect could also be achieved using barrier(), but READ_ONCE() 1721 and WRITE_ONCE() are more selective: With READ_ONCE() and 1722 WRITE_ONCE(), the compiler need only forget the contents of the 1723 indicated memory locations, while with barrier() the compiler must 1724 discard the value of all memory locations that it has currented 1725 cached in any machine registers. Of course, the compiler must also 1726 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur, 1727 though the CPU of course need not do so. 1728 1729 (*) The compiler is within its rights to invent stores to a variable, 1730 as in the following example: 1731 1732 if (a) 1733 b = a; 1734 else 1735 b = 42; 1736 1737 The compiler might save a branch by optimizing this as follows: 1738 1739 b = 42; 1740 if (a) 1741 b = a; 1742 1743 In single-threaded code, this is not only safe, but also saves 1744 a branch. Unfortunately, in concurrent code, this optimization 1745 could cause some other CPU to see a spurious value of 42 -- even 1746 if variable 'a' was never zero -- when loading variable 'b'. 1747 Use WRITE_ONCE() to prevent this as follows: 1748 1749 if (a) 1750 WRITE_ONCE(b, a); 1751 else 1752 WRITE_ONCE(b, 42); 1753 1754 The compiler can also invent loads. These are usually less 1755 damaging, but they can result in cache-line bouncing and thus in 1756 poor performance and scalability. Use READ_ONCE() to prevent 1757 invented loads. 1758 1759 (*) For aligned memory locations whose size allows them to be accessed 1760 with a single memory-reference instruction, prevents "load tearing" 1761 and "store tearing," in which a single large access is replaced by 1762 multiple smaller accesses. For example, given an architecture having 1763 16-bit store instructions with 7-bit immediate fields, the compiler 1764 might be tempted to use two 16-bit store-immediate instructions to 1765 implement the following 32-bit store: 1766 1767 p = 0x00010002; 1768 1769 Please note that GCC really does use this sort of optimization, 1770 which is not surprising given that it would likely take more 1771 than two instructions to build the constant and then store it. 1772 This optimization can therefore be a win in single-threaded code. 1773 In fact, a recent bug (since fixed) caused GCC to incorrectly use 1774 this optimization in a volatile store. In the absence of such bugs, 1775 use of WRITE_ONCE() prevents store tearing in the following example: 1776 1777 WRITE_ONCE(p, 0x00010002); 1778 1779 Use of packed structures can also result in load and store tearing, 1780 as in this example: 1781 1782 struct __attribute__((__packed__)) foo { 1783 short a; 1784 int b; 1785 short c; 1786 }; 1787 struct foo foo1, foo2; 1788 ... 1789 1790 foo2.a = foo1.a; 1791 foo2.b = foo1.b; 1792 foo2.c = foo1.c; 1793 1794 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no 1795 volatile markings, the compiler would be well within its rights to 1796 implement these three assignment statements as a pair of 32-bit 1797 loads followed by a pair of 32-bit stores. This would result in 1798 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE() 1799 and WRITE_ONCE() again prevent tearing in this example: 1800 1801 foo2.a = foo1.a; 1802 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b)); 1803 foo2.c = foo1.c; 1804 1805All that aside, it is never necessary to use READ_ONCE() and 1806WRITE_ONCE() on a variable that has been marked volatile. For example, 1807because 'jiffies' is marked volatile, it is never necessary to 1808say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and 1809WRITE_ONCE() are implemented as volatile casts, which has no effect when 1810its argument is already marked volatile. 1811 1812Please note that these compiler barriers have no direct effect on the CPU, 1813which may then reorder things however it wishes. 1814 1815 1816CPU MEMORY BARRIERS 1817------------------- 1818 1819The Linux kernel has eight basic CPU memory barriers: 1820 1821 TYPE MANDATORY SMP CONDITIONAL 1822 =============== ======================= =========================== 1823 GENERAL mb() smp_mb() 1824 WRITE wmb() smp_wmb() 1825 READ rmb() smp_rmb() 1826 DATA DEPENDENCY READ_ONCE() 1827 1828 1829All memory barriers except the data dependency barriers imply a compiler 1830barrier. Data dependencies do not impose any additional compiler ordering. 1831 1832Aside: In the case of data dependencies, the compiler would be expected 1833to issue the loads in the correct order (eg. `a[b]` would have to load 1834the value of b before loading a[b]), however there is no guarantee in 1835the C specification that the compiler may not speculate the value of b 1836(eg. is equal to 1) and load a before b (eg. tmp = a[1]; if (b != 1) 1837tmp = a[b]; ). There is also the problem of a compiler reloading b after 1838having loaded a[b], thus having a newer copy of b than a[b]. A consensus 1839has not yet been reached about these problems, however the READ_ONCE() 1840macro is a good place to start looking. 1841 1842SMP memory barriers are reduced to compiler barriers on uniprocessor compiled 1843systems because it is assumed that a CPU will appear to be self-consistent, 1844and will order overlapping accesses correctly with respect to itself. 1845However, see the subsection on "Virtual Machine Guests" below. 1846 1847[!] Note that SMP memory barriers _must_ be used to control the ordering of 1848references to shared memory on SMP systems, though the use of locking instead 1849is sufficient. 1850 1851Mandatory barriers should not be used to control SMP effects, since mandatory 1852barriers impose unnecessary overhead on both SMP and UP systems. They may, 1853however, be used to control MMIO effects on accesses through relaxed memory I/O 1854windows. These barriers are required even on non-SMP systems as they affect 1855the order in which memory operations appear to a device by prohibiting both the 1856compiler and the CPU from reordering them. 1857 1858 1859There are some more advanced barrier functions: 1860 1861 (*) smp_store_mb(var, value) 1862 1863 This assigns the value to the variable and then inserts a full memory 1864 barrier after it. It isn't guaranteed to insert anything more than a 1865 compiler barrier in a UP compilation. 1866 1867 1868 (*) smp_mb__before_atomic(); 1869 (*) smp_mb__after_atomic(); 1870 1871 These are for use with atomic RMW functions that do not imply memory 1872 barriers, but where the code needs a memory barrier. Examples for atomic 1873 RMW functions that do not imply are memory barrier are e.g. add, 1874 subtract, (failed) conditional operations, _relaxed functions, 1875 but not atomic_read or atomic_set. A common example where a memory 1876 barrier may be required is when atomic ops are used for reference 1877 counting. 1878 1879 These are also used for atomic RMW bitop functions that do not imply a 1880 memory barrier (such as set_bit and clear_bit). 1881 1882 As an example, consider a piece of code that marks an object as being dead 1883 and then decrements the object's reference count: 1884 1885 obj->dead = 1; 1886 smp_mb__before_atomic(); 1887 atomic_dec(&obj->ref_count); 1888 1889 This makes sure that the death mark on the object is perceived to be set 1890 *before* the reference counter is decremented. 1891 1892 See Documentation/atomic_{t,bitops}.txt for more information. 1893 1894 1895 (*) dma_wmb(); 1896 (*) dma_rmb(); 1897 1898 These are for use with consistent memory to guarantee the ordering 1899 of writes or reads of shared memory accessible to both the CPU and a 1900 DMA capable device. 1901 1902 For example, consider a device driver that shares memory with a device 1903 and uses a descriptor status value to indicate if the descriptor belongs 1904 to the device or the CPU, and a doorbell to notify it when new 1905 descriptors are available: 1906 1907 if (desc->status != DEVICE_OWN) { 1908 /* do not read data until we own descriptor */ 1909 dma_rmb(); 1910 1911 /* read/modify data */ 1912 read_data = desc->data; 1913 desc->data = write_data; 1914 1915 /* flush modifications before status update */ 1916 dma_wmb(); 1917 1918 /* assign ownership */ 1919 desc->status = DEVICE_OWN; 1920 1921 /* notify device of new descriptors */ 1922 writel(DESC_NOTIFY, doorbell); 1923 } 1924 1925 The dma_rmb() allows us guarantee the device has released ownership 1926 before we read the data from the descriptor, and the dma_wmb() allows 1927 us to guarantee the data is written to the descriptor before the device 1928 can see it now has ownership. Note that, when using writel(), a prior 1929 wmb() is not needed to guarantee that the cache coherent memory writes 1930 have completed before writing to the MMIO region. The cheaper 1931 writel_relaxed() does not provide this guarantee and must not be used 1932 here. 1933 1934 See the subsection "Kernel I/O barrier effects" for more information on 1935 relaxed I/O accessors and the Documentation/DMA-API.txt file for more 1936 information on consistent memory. 1937 1938 1939=============================== 1940IMPLICIT KERNEL MEMORY BARRIERS 1941=============================== 1942 1943Some of the other functions in the linux kernel imply memory barriers, amongst 1944which are locking and scheduling functions. 1945 1946This specification is a _minimum_ guarantee; any particular architecture may 1947provide more substantial guarantees, but these may not be relied upon outside 1948of arch specific code. 1949 1950 1951LOCK ACQUISITION FUNCTIONS 1952-------------------------- 1953 1954The Linux kernel has a number of locking constructs: 1955 1956 (*) spin locks 1957 (*) R/W spin locks 1958 (*) mutexes 1959 (*) semaphores 1960 (*) R/W semaphores 1961 1962In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations 1963for each construct. These operations all imply certain barriers: 1964 1965 (1) ACQUIRE operation implication: 1966 1967 Memory operations issued after the ACQUIRE will be completed after the 1968 ACQUIRE operation has completed. 1969 1970 Memory operations issued before the ACQUIRE may be completed after 1971 the ACQUIRE operation has completed. 1972 1973 (2) RELEASE operation implication: 1974 1975 Memory operations issued before the RELEASE will be completed before the 1976 RELEASE operation has completed. 1977 1978 Memory operations issued after the RELEASE may be completed before the 1979 RELEASE operation has completed. 1980 1981 (3) ACQUIRE vs ACQUIRE implication: 1982 1983 All ACQUIRE operations issued before another ACQUIRE operation will be 1984 completed before that ACQUIRE operation. 1985 1986 (4) ACQUIRE vs RELEASE implication: 1987 1988 All ACQUIRE operations issued before a RELEASE operation will be 1989 completed before the RELEASE operation. 1990 1991 (5) Failed conditional ACQUIRE implication: 1992 1993 Certain locking variants of the ACQUIRE operation may fail, either due to 1994 being unable to get the lock immediately, or due to receiving an unblocked 1995 signal while asleep waiting for the lock to become available. Failed 1996 locks do not imply any sort of barrier. 1997 1998[!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only 1999one-way barriers is that the effects of instructions outside of a critical 2000section may seep into the inside of the critical section. 2001 2002An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier 2003because it is possible for an access preceding the ACQUIRE to happen after the 2004ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and 2005the two accesses can themselves then cross: 2006 2007 *A = a; 2008 ACQUIRE M 2009 RELEASE M 2010 *B = b; 2011 2012may occur as: 2013 2014 ACQUIRE M, STORE *B, STORE *A, RELEASE M 2015 2016When the ACQUIRE and RELEASE are a lock acquisition and release, 2017respectively, this same reordering can occur if the lock's ACQUIRE and 2018RELEASE are to the same lock variable, but only from the perspective of 2019another CPU not holding that lock. In short, a ACQUIRE followed by an 2020RELEASE may -not- be assumed to be a full memory barrier. 2021 2022Similarly, the reverse case of a RELEASE followed by an ACQUIRE does 2023not imply a full memory barrier. Therefore, the CPU's execution of the 2024critical sections corresponding to the RELEASE and the ACQUIRE can cross, 2025so that: 2026 2027 *A = a; 2028 RELEASE M 2029 ACQUIRE N 2030 *B = b; 2031 2032could occur as: 2033 2034 ACQUIRE N, STORE *B, STORE *A, RELEASE M 2035 2036It might appear that this reordering could introduce a deadlock. 2037However, this cannot happen because if such a deadlock threatened, 2038the RELEASE would simply complete, thereby avoiding the deadlock. 2039 2040 Why does this work? 2041 2042 One key point is that we are only talking about the CPU doing 2043 the reordering, not the compiler. If the compiler (or, for 2044 that matter, the developer) switched the operations, deadlock 2045 -could- occur. 2046 2047 But suppose the CPU reordered the operations. In this case, 2048 the unlock precedes the lock in the assembly code. The CPU 2049 simply elected to try executing the later lock operation first. 2050 If there is a deadlock, this lock operation will simply spin (or 2051 try to sleep, but more on that later). The CPU will eventually 2052 execute the unlock operation (which preceded the lock operation 2053 in the assembly code), which will unravel the potential deadlock, 2054 allowing the lock operation to succeed. 2055 2056 But what if the lock is a sleeplock? In that case, the code will 2057 try to enter the scheduler, where it will eventually encounter 2058 a memory barrier, which will force the earlier unlock operation 2059 to complete, again unraveling the deadlock. There might be 2060 a sleep-unlock race, but the locking primitive needs to resolve 2061 such races properly in any case. 2062 2063Locks and semaphores may not provide any guarantee of ordering on UP compiled 2064systems, and so cannot be counted on in such a situation to actually achieve 2065anything at all - especially with respect to I/O accesses - unless combined 2066with interrupt disabling operations. 2067 2068See also the section on "Inter-CPU acquiring barrier effects". 2069 2070 2071As an example, consider the following: 2072 2073 *A = a; 2074 *B = b; 2075 ACQUIRE 2076 *C = c; 2077 *D = d; 2078 RELEASE 2079 *E = e; 2080 *F = f; 2081 2082The following sequence of events is acceptable: 2083 2084 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE 2085 2086 [+] Note that {*F,*A} indicates a combined access. 2087 2088But none of the following are: 2089 2090 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E 2091 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F 2092 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F 2093 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E 2094 2095 2096 2097INTERRUPT DISABLING FUNCTIONS 2098----------------------------- 2099 2100Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts 2101(RELEASE equivalent) will act as compiler barriers only. So if memory or I/O 2102barriers are required in such a situation, they must be provided from some 2103other means. 2104 2105 2106SLEEP AND WAKE-UP FUNCTIONS 2107--------------------------- 2108 2109Sleeping and waking on an event flagged in global data can be viewed as an 2110interaction between two pieces of data: the task state of the task waiting for 2111the event and the global data used to indicate the event. To make sure that 2112these appear to happen in the right order, the primitives to begin the process 2113of going to sleep, and the primitives to initiate a wake up imply certain 2114barriers. 2115 2116Firstly, the sleeper normally follows something like this sequence of events: 2117 2118 for (;;) { 2119 set_current_state(TASK_UNINTERRUPTIBLE); 2120 if (event_indicated) 2121 break; 2122 schedule(); 2123 } 2124 2125A general memory barrier is interpolated automatically by set_current_state() 2126after it has altered the task state: 2127 2128 CPU 1 2129 =============================== 2130 set_current_state(); 2131 smp_store_mb(); 2132 STORE current->state 2133 <general barrier> 2134 LOAD event_indicated 2135 2136set_current_state() may be wrapped by: 2137 2138 prepare_to_wait(); 2139 prepare_to_wait_exclusive(); 2140 2141which therefore also imply a general memory barrier after setting the state. 2142The whole sequence above is available in various canned forms, all of which 2143interpolate the memory barrier in the right place: 2144 2145 wait_event(); 2146 wait_event_interruptible(); 2147 wait_event_interruptible_exclusive(); 2148 wait_event_interruptible_timeout(); 2149 wait_event_killable(); 2150 wait_event_timeout(); 2151 wait_on_bit(); 2152 wait_on_bit_lock(); 2153 2154 2155Secondly, code that performs a wake up normally follows something like this: 2156 2157 event_indicated = 1; 2158 wake_up(&event_wait_queue); 2159 2160or: 2161 2162 event_indicated = 1; 2163 wake_up_process(event_daemon); 2164 2165A general memory barrier is executed by wake_up() if it wakes something up. 2166If it doesn't wake anything up then a memory barrier may or may not be 2167executed; you must not rely on it. The barrier occurs before the task state 2168is accessed, in particular, it sits between the STORE to indicate the event 2169and the STORE to set TASK_RUNNING: 2170 2171 CPU 1 (Sleeper) CPU 2 (Waker) 2172 =============================== =============================== 2173 set_current_state(); STORE event_indicated 2174 smp_store_mb(); wake_up(); 2175 STORE current->state ... 2176 <general barrier> <general barrier> 2177 LOAD event_indicated if ((LOAD task->state) & TASK_NORMAL) 2178 STORE task->state 2179 2180where "task" is the thread being woken up and it equals CPU 1's "current". 2181 2182To repeat, a general memory barrier is guaranteed to be executed by wake_up() 2183if something is actually awakened, but otherwise there is no such guarantee. 2184To see this, consider the following sequence of events, where X and Y are both 2185initially zero: 2186 2187 CPU 1 CPU 2 2188 =============================== =============================== 2189 X = 1; Y = 1; 2190 smp_mb(); wake_up(); 2191 LOAD Y LOAD X 2192 2193If a wakeup does occur, one (at least) of the two loads must see 1. If, on 2194the other hand, a wakeup does not occur, both loads might see 0. 2195 2196wake_up_process() always executes a general memory barrier. The barrier again 2197occurs before the task state is accessed. In particular, if the wake_up() in 2198the previous snippet were replaced by a call to wake_up_process() then one of 2199the two loads would be guaranteed to see 1. 2200 2201The available waker functions include: 2202 2203 complete(); 2204 wake_up(); 2205 wake_up_all(); 2206 wake_up_bit(); 2207 wake_up_interruptible(); 2208 wake_up_interruptible_all(); 2209 wake_up_interruptible_nr(); 2210 wake_up_interruptible_poll(); 2211 wake_up_interruptible_sync(); 2212 wake_up_interruptible_sync_poll(); 2213 wake_up_locked(); 2214 wake_up_locked_poll(); 2215 wake_up_nr(); 2216 wake_up_poll(); 2217 wake_up_process(); 2218 2219In terms of memory ordering, these functions all provide the same guarantees of 2220a wake_up() (or stronger). 2221 2222[!] Note that the memory barriers implied by the sleeper and the waker do _not_ 2223order multiple stores before the wake-up with respect to loads of those stored 2224values after the sleeper has called set_current_state(). For instance, if the 2225sleeper does: 2226 2227 set_current_state(TASK_INTERRUPTIBLE); 2228 if (event_indicated) 2229 break; 2230 __set_current_state(TASK_RUNNING); 2231 do_something(my_data); 2232 2233and the waker does: 2234 2235 my_data = value; 2236 event_indicated = 1; 2237 wake_up(&event_wait_queue); 2238 2239there's no guarantee that the change to event_indicated will be perceived by 2240the sleeper as coming after the change to my_data. In such a circumstance, the 2241code on both sides must interpolate its own memory barriers between the 2242separate data accesses. Thus the above sleeper ought to do: 2243 2244 set_current_state(TASK_INTERRUPTIBLE); 2245 if (event_indicated) { 2246 smp_rmb(); 2247 do_something(my_data); 2248 } 2249 2250and the waker should do: 2251 2252 my_data = value; 2253 smp_wmb(); 2254 event_indicated = 1; 2255 wake_up(&event_wait_queue); 2256 2257 2258MISCELLANEOUS FUNCTIONS 2259----------------------- 2260 2261Other functions that imply barriers: 2262 2263 (*) schedule() and similar imply full memory barriers. 2264 2265 2266=================================== 2267INTER-CPU ACQUIRING BARRIER EFFECTS 2268=================================== 2269 2270On SMP systems locking primitives give a more substantial form of barrier: one 2271that does affect memory access ordering on other CPUs, within the context of 2272conflict on any particular lock. 2273 2274 2275ACQUIRES VS MEMORY ACCESSES 2276--------------------------- 2277 2278Consider the following: the system has a pair of spinlocks (M) and (Q), and 2279three CPUs; then should the following sequence of events occur: 2280 2281 CPU 1 CPU 2 2282 =============================== =============================== 2283 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e); 2284 ACQUIRE M ACQUIRE Q 2285 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f); 2286 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g); 2287 RELEASE M RELEASE Q 2288 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h); 2289 2290Then there is no guarantee as to what order CPU 3 will see the accesses to *A 2291through *H occur in, other than the constraints imposed by the separate locks 2292on the separate CPUs. It might, for example, see: 2293 2294 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M 2295 2296But it won't see any of: 2297 2298 *B, *C or *D preceding ACQUIRE M 2299 *A, *B or *C following RELEASE M 2300 *F, *G or *H preceding ACQUIRE Q 2301 *E, *F or *G following RELEASE Q 2302 2303 2304================================= 2305WHERE ARE MEMORY BARRIERS NEEDED? 2306================================= 2307 2308Under normal operation, memory operation reordering is generally not going to 2309be a problem as a single-threaded linear piece of code will still appear to 2310work correctly, even if it's in an SMP kernel. There are, however, four 2311circumstances in which reordering definitely _could_ be a problem: 2312 2313 (*) Interprocessor interaction. 2314 2315 (*) Atomic operations. 2316 2317 (*) Accessing devices. 2318 2319 (*) Interrupts. 2320 2321 2322INTERPROCESSOR INTERACTION 2323-------------------------- 2324 2325When there's a system with more than one processor, more than one CPU in the 2326system may be working on the same data set at the same time. This can cause 2327synchronisation problems, and the usual way of dealing with them is to use 2328locks. Locks, however, are quite expensive, and so it may be preferable to 2329operate without the use of a lock if at all possible. In such a case 2330operations that affect both CPUs may have to be carefully ordered to prevent 2331a malfunction. 2332 2333Consider, for example, the R/W semaphore slow path. Here a waiting process is 2334queued on the semaphore, by virtue of it having a piece of its stack linked to 2335the semaphore's list of waiting processes: 2336 2337 struct rw_semaphore { 2338 ... 2339 spinlock_t lock; 2340 struct list_head waiters; 2341 }; 2342 2343 struct rwsem_waiter { 2344 struct list_head list; 2345 struct task_struct *task; 2346 }; 2347 2348To wake up a particular waiter, the up_read() or up_write() functions have to: 2349 2350 (1) read the next pointer from this waiter's record to know as to where the 2351 next waiter record is; 2352 2353 (2) read the pointer to the waiter's task structure; 2354 2355 (3) clear the task pointer to tell the waiter it has been given the semaphore; 2356 2357 (4) call wake_up_process() on the task; and 2358 2359 (5) release the reference held on the waiter's task struct. 2360 2361In other words, it has to perform this sequence of events: 2362 2363 LOAD waiter->list.next; 2364 LOAD waiter->task; 2365 STORE waiter->task; 2366 CALL wakeup 2367 RELEASE task 2368 2369and if any of these steps occur out of order, then the whole thing may 2370malfunction. 2371 2372Once it has queued itself and dropped the semaphore lock, the waiter does not 2373get the lock again; it instead just waits for its task pointer to be cleared 2374before proceeding. Since the record is on the waiter's stack, this means that 2375if the task pointer is cleared _before_ the next pointer in the list is read, 2376another CPU might start processing the waiter and might clobber the waiter's 2377stack before the up*() function has a chance to read the next pointer. 2378 2379Consider then what might happen to the above sequence of events: 2380 2381 CPU 1 CPU 2 2382 =============================== =============================== 2383 down_xxx() 2384 Queue waiter 2385 Sleep 2386 up_yyy() 2387 LOAD waiter->task; 2388 STORE waiter->task; 2389 Woken up by other event 2390 <preempt> 2391 Resume processing 2392 down_xxx() returns 2393 call foo() 2394 foo() clobbers *waiter 2395 </preempt> 2396 LOAD waiter->list.next; 2397 --- OOPS --- 2398 2399This could be dealt with using the semaphore lock, but then the down_xxx() 2400function has to needlessly get the spinlock again after being woken up. 2401 2402The way to deal with this is to insert a general SMP memory barrier: 2403 2404 LOAD waiter->list.next; 2405 LOAD waiter->task; 2406 smp_mb(); 2407 STORE waiter->task; 2408 CALL wakeup 2409 RELEASE task 2410 2411In this case, the barrier makes a guarantee that all memory accesses before the 2412barrier will appear to happen before all the memory accesses after the barrier 2413with respect to the other CPUs on the system. It does _not_ guarantee that all 2414the memory accesses before the barrier will be complete by the time the barrier 2415instruction itself is complete. 2416 2417On a UP system - where this wouldn't be a problem - the smp_mb() is just a 2418compiler barrier, thus making sure the compiler emits the instructions in the 2419right order without actually intervening in the CPU. Since there's only one 2420CPU, that CPU's dependency ordering logic will take care of everything else. 2421 2422 2423ATOMIC OPERATIONS 2424----------------- 2425 2426While they are technically interprocessor interaction considerations, atomic 2427operations are noted specially as some of them imply full memory barriers and 2428some don't, but they're very heavily relied on as a group throughout the 2429kernel. 2430 2431See Documentation/atomic_t.txt for more information. 2432 2433 2434ACCESSING DEVICES 2435----------------- 2436 2437Many devices can be memory mapped, and so appear to the CPU as if they're just 2438a set of memory locations. To control such a device, the driver usually has to 2439make the right memory accesses in exactly the right order. 2440 2441However, having a clever CPU or a clever compiler creates a potential problem 2442in that the carefully sequenced accesses in the driver code won't reach the 2443device in the requisite order if the CPU or the compiler thinks it is more 2444efficient to reorder, combine or merge accesses - something that would cause 2445the device to malfunction. 2446 2447Inside of the Linux kernel, I/O should be done through the appropriate accessor 2448routines - such as inb() or writel() - which know how to make such accesses 2449appropriately sequential. While this, for the most part, renders the explicit 2450use of memory barriers unnecessary, if the accessor functions are used to refer 2451to an I/O memory window with relaxed memory access properties, then _mandatory_ 2452memory barriers are required to enforce ordering. 2453 2454See Documentation/driver-api/device-io.rst for more information. 2455 2456 2457INTERRUPTS 2458---------- 2459 2460A driver may be interrupted by its own interrupt service routine, and thus the 2461two parts of the driver may interfere with each other's attempts to control or 2462access the device. 2463 2464This may be alleviated - at least in part - by disabling local interrupts (a 2465form of locking), such that the critical operations are all contained within 2466the interrupt-disabled section in the driver. While the driver's interrupt 2467routine is executing, the driver's core may not run on the same CPU, and its 2468interrupt is not permitted to happen again until the current interrupt has been 2469handled, thus the interrupt handler does not need to lock against that. 2470 2471However, consider a driver that was talking to an ethernet card that sports an 2472address register and a data register. If that driver's core talks to the card 2473under interrupt-disablement and then the driver's interrupt handler is invoked: 2474 2475 LOCAL IRQ DISABLE 2476 writew(ADDR, 3); 2477 writew(DATA, y); 2478 LOCAL IRQ ENABLE 2479 <interrupt> 2480 writew(ADDR, 4); 2481 q = readw(DATA); 2482 </interrupt> 2483 2484The store to the data register might happen after the second store to the 2485address register if ordering rules are sufficiently relaxed: 2486 2487 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA 2488 2489 2490If ordering rules are relaxed, it must be assumed that accesses done inside an 2491interrupt disabled section may leak outside of it and may interleave with 2492accesses performed in an interrupt - and vice versa - unless implicit or 2493explicit barriers are used. 2494 2495Normally this won't be a problem because the I/O accesses done inside such 2496sections will include synchronous load operations on strictly ordered I/O 2497registers that form implicit I/O barriers. 2498 2499 2500A similar situation may occur between an interrupt routine and two routines 2501running on separate CPUs that communicate with each other. If such a case is 2502likely, then interrupt-disabling locks should be used to guarantee ordering. 2503 2504 2505========================== 2506KERNEL I/O BARRIER EFFECTS 2507========================== 2508 2509Interfacing with peripherals via I/O accesses is deeply architecture and device 2510specific. Therefore, drivers which are inherently non-portable may rely on 2511specific behaviours of their target systems in order to achieve synchronization 2512in the most lightweight manner possible. For drivers intending to be portable 2513between multiple architectures and bus implementations, the kernel offers a 2514series of accessor functions that provide various degrees of ordering 2515guarantees: 2516 2517 (*) readX(), writeX(): 2518 2519 The readX() and writeX() MMIO accessors take a pointer to the 2520 peripheral being accessed as an __iomem * parameter. For pointers 2521 mapped with the default I/O attributes (e.g. those returned by 2522 ioremap()), the ordering guarantees are as follows: 2523 2524 1. All readX() and writeX() accesses to the same peripheral are ordered 2525 with respect to each other. This ensures that MMIO register accesses 2526 by the same CPU thread to a particular device will arrive in program 2527 order. 2528 2529 2. A writeX() issued by a CPU thread holding a spinlock is ordered 2530 before a writeX() to the same peripheral from another CPU thread 2531 issued after a later acquisition of the same spinlock. This ensures 2532 that MMIO register writes to a particular device issued while holding 2533 a spinlock will arrive in an order consistent with acquisitions of 2534 the lock. 2535 2536 3. A writeX() by a CPU thread to the peripheral will first wait for the 2537 completion of all prior writes to memory either issued by, or 2538 propagated to, the same thread. This ensures that writes by the CPU 2539 to an outbound DMA buffer allocated by dma_alloc_coherent() will be 2540 visible to a DMA engine when the CPU writes to its MMIO control 2541 register to trigger the transfer. 2542 2543 4. A readX() by a CPU thread from the peripheral will complete before 2544 any subsequent reads from memory by the same thread can begin. This 2545 ensures that reads by the CPU from an incoming DMA buffer allocated 2546 by dma_alloc_coherent() will not see stale data after reading from 2547 the DMA engine's MMIO status register to establish that the DMA 2548 transfer has completed. 2549 2550 5. A readX() by a CPU thread from the peripheral will complete before 2551 any subsequent delay() loop can begin execution on the same thread. 2552 This ensures that two MMIO register writes by the CPU to a peripheral 2553 will arrive at least 1us apart if the first write is immediately read 2554 back with readX() and udelay(1) is called prior to the second 2555 writeX(): 2556 2557 writel(42, DEVICE_REGISTER_0); // Arrives at the device... 2558 readl(DEVICE_REGISTER_0); 2559 udelay(1); 2560 writel(42, DEVICE_REGISTER_1); // ...at least 1us before this. 2561 2562 The ordering properties of __iomem pointers obtained with non-default 2563 attributes (e.g. those returned by ioremap_wc()) are specific to the 2564 underlying architecture and therefore the guarantees listed above cannot 2565 generally be relied upon for accesses to these types of mappings. 2566 2567 (*) readX_relaxed(), writeX_relaxed(): 2568 2569 These are similar to readX() and writeX(), but provide weaker memory 2570 ordering guarantees. Specifically, they do not guarantee ordering with 2571 respect to locking, normal memory accesses or delay() loops (i.e. 2572 bullets 2-5 above) but they are still guaranteed to be ordered with 2573 respect to other accesses from the same CPU thread to the same 2574 peripheral when operating on __iomem pointers mapped with the default 2575 I/O attributes. 2576 2577 (*) readsX(), writesX(): 2578 2579 The readsX() and writesX() MMIO accessors are designed for accessing 2580 register-based, memory-mapped FIFOs residing on peripherals that are not 2581 capable of performing DMA. Consequently, they provide only the ordering 2582 guarantees of readX_relaxed() and writeX_relaxed(), as documented above. 2583 2584 (*) inX(), outX(): 2585 2586 The inX() and outX() accessors are intended to access legacy port-mapped 2587 I/O peripherals, which may require special instructions on some 2588 architectures (notably x86). The port number of the peripheral being 2589 accessed is passed as an argument. 2590 2591 Since many CPU architectures ultimately access these peripherals via an 2592 internal virtual memory mapping, the portable ordering guarantees 2593 provided by inX() and outX() are the same as those provided by readX() 2594 and writeX() respectively when accessing a mapping with the default I/O 2595 attributes. 2596 2597 Device drivers may expect outX() to emit a non-posted write transaction 2598 that waits for a completion response from the I/O peripheral before 2599 returning. This is not guaranteed by all architectures and is therefore 2600 not part of the portable ordering semantics. 2601 2602 (*) insX(), outsX(): 2603 2604 As above, the insX() and outsX() accessors provide the same ordering 2605 guarantees as readsX() and writesX() respectively when accessing a 2606 mapping with the default I/O attributes. 2607 2608 (*) ioreadX(), iowriteX(): 2609 2610 These will perform appropriately for the type of access they're actually 2611 doing, be it inX()/outX() or readX()/writeX(). 2612 2613With the exception of the string accessors (insX(), outsX(), readsX() and 2614writesX()), all of the above assume that the underlying peripheral is 2615little-endian and will therefore perform byte-swapping operations on big-endian 2616architectures. 2617 2618 2619======================================== 2620ASSUMED MINIMUM EXECUTION ORDERING MODEL 2621======================================== 2622 2623It has to be assumed that the conceptual CPU is weakly-ordered but that it will 2624maintain the appearance of program causality with respect to itself. Some CPUs 2625(such as i386 or x86_64) are more constrained than others (such as powerpc or 2626frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside 2627of arch-specific code. 2628 2629This means that it must be considered that the CPU will execute its instruction 2630stream in any order it feels like - or even in parallel - provided that if an 2631instruction in the stream depends on an earlier instruction, then that 2632earlier instruction must be sufficiently complete[*] before the later 2633instruction may proceed; in other words: provided that the appearance of 2634causality is maintained. 2635 2636 [*] Some instructions have more than one effect - such as changing the 2637 condition codes, changing registers or changing memory - and different 2638 instructions may depend on different effects. 2639 2640A CPU may also discard any instruction sequence that winds up having no 2641ultimate effect. For example, if two adjacent instructions both load an 2642immediate value into the same register, the first may be discarded. 2643 2644 2645Similarly, it has to be assumed that compiler might reorder the instruction 2646stream in any way it sees fit, again provided the appearance of causality is 2647maintained. 2648 2649 2650============================ 2651THE EFFECTS OF THE CPU CACHE 2652============================ 2653 2654The way cached memory operations are perceived across the system is affected to 2655a certain extent by the caches that lie between CPUs and memory, and by the 2656memory coherence system that maintains the consistency of state in the system. 2657 2658As far as the way a CPU interacts with another part of the system through the 2659caches goes, the memory system has to include the CPU's caches, and memory 2660barriers for the most part act at the interface between the CPU and its cache 2661(memory barriers logically act on the dotted line in the following diagram): 2662 2663 <--- CPU ---> : <----------- Memory -----------> 2664 : 2665 +--------+ +--------+ : +--------+ +-----------+ 2666 | | | | : | | | | +--------+ 2667 | CPU | | Memory | : | CPU | | | | | 2668 | Core |--->| Access |----->| Cache |<-->| | | | 2669 | | | Queue | : | | | |--->| Memory | 2670 | | | | : | | | | | | 2671 +--------+ +--------+ : +--------+ | | | | 2672 : | Cache | +--------+ 2673 : | Coherency | 2674 : | Mechanism | +--------+ 2675 +--------+ +--------+ : +--------+ | | | | 2676 | | | | : | | | | | | 2677 | CPU | | Memory | : | CPU | | |--->| Device | 2678 | Core |--->| Access |----->| Cache |<-->| | | | 2679 | | | Queue | : | | | | | | 2680 | | | | : | | | | +--------+ 2681 +--------+ +--------+ : +--------+ +-----------+ 2682 : 2683 : 2684 2685Although any particular load or store may not actually appear outside of the 2686CPU that issued it since it may have been satisfied within the CPU's own cache, 2687it will still appear as if the full memory access had taken place as far as the 2688other CPUs are concerned since the cache coherency mechanisms will migrate the 2689cacheline over to the accessing CPU and propagate the effects upon conflict. 2690 2691The CPU core may execute instructions in any order it deems fit, provided the 2692expected program causality appears to be maintained. Some of the instructions 2693generate load and store operations which then go into the queue of memory 2694accesses to be performed. The core may place these in the queue in any order 2695it wishes, and continue execution until it is forced to wait for an instruction 2696to complete. 2697 2698What memory barriers are concerned with is controlling the order in which 2699accesses cross from the CPU side of things to the memory side of things, and 2700the order in which the effects are perceived to happen by the other observers 2701in the system. 2702 2703[!] Memory barriers are _not_ needed within a given CPU, as CPUs always see 2704their own loads and stores as if they had happened in program order. 2705 2706[!] MMIO or other device accesses may bypass the cache system. This depends on 2707the properties of the memory window through which devices are accessed and/or 2708the use of any special device communication instructions the CPU may have. 2709 2710 2711CACHE COHERENCY 2712--------------- 2713 2714Life isn't quite as simple as it may appear above, however: for while the 2715caches are expected to be coherent, there's no guarantee that that coherency 2716will be ordered. This means that while changes made on one CPU will 2717eventually become visible on all CPUs, there's no guarantee that they will 2718become apparent in the same order on those other CPUs. 2719 2720 2721Consider dealing with a system that has a pair of CPUs (1 & 2), each of which 2722has a pair of parallel data caches (CPU 1 has A/B, and CPU 2 has C/D): 2723 2724 : 2725 : +--------+ 2726 : +---------+ | | 2727 +--------+ : +--->| Cache A |<------->| | 2728 | | : | +---------+ | | 2729 | CPU 1 |<---+ | | 2730 | | : | +---------+ | | 2731 +--------+ : +--->| Cache B |<------->| | 2732 : +---------+ | | 2733 : | Memory | 2734 : +---------+ | System | 2735 +--------+ : +--->| Cache C |<------->| | 2736 | | : | +---------+ | | 2737 | CPU 2 |<---+ | | 2738 | | : | +---------+ | | 2739 +--------+ : +--->| Cache D |<------->| | 2740 : +---------+ | | 2741 : +--------+ 2742 : 2743 2744Imagine the system has the following properties: 2745 2746 (*) an odd-numbered cache line may be in cache A, cache C or it may still be 2747 resident in memory; 2748 2749 (*) an even-numbered cache line may be in cache B, cache D or it may still be 2750 resident in memory; 2751 2752 (*) while the CPU core is interrogating one cache, the other cache may be 2753 making use of the bus to access the rest of the system - perhaps to 2754 displace a dirty cacheline or to do a speculative load; 2755 2756 (*) each cache has a queue of operations that need to be applied to that cache 2757 to maintain coherency with the rest of the system; 2758 2759 (*) the coherency queue is not flushed by normal loads to lines already 2760 present in the cache, even though the contents of the queue may 2761 potentially affect those loads. 2762 2763Imagine, then, that two writes are made on the first CPU, with a write barrier 2764between them to guarantee that they will appear to reach that CPU's caches in 2765the requisite order: 2766 2767 CPU 1 CPU 2 COMMENT 2768 =============== =============== ======================================= 2769 u == 0, v == 1 and p == &u, q == &u 2770 v = 2; 2771 smp_wmb(); Make sure change to v is visible before 2772 change to p 2773 <A:modify v=2> v is now in cache A exclusively 2774 p = &v; 2775 <B:modify p=&v> p is now in cache B exclusively 2776 2777The write memory barrier forces the other CPUs in the system to perceive that 2778the local CPU's caches have apparently been updated in the correct order. But 2779now imagine that the second CPU wants to read those values: 2780 2781 CPU 1 CPU 2 COMMENT 2782 =============== =============== ======================================= 2783 ... 2784 q = p; 2785 x = *q; 2786 2787The above pair of reads may then fail to happen in the expected order, as the 2788cacheline holding p may get updated in one of the second CPU's caches while 2789the update to the cacheline holding v is delayed in the other of the second 2790CPU's caches by some other cache event: 2791 2792 CPU 1 CPU 2 COMMENT 2793 =============== =============== ======================================= 2794 u == 0, v == 1 and p == &u, q == &u 2795 v = 2; 2796 smp_wmb(); 2797 <A:modify v=2> <C:busy> 2798 <C:queue v=2> 2799 p = &v; q = p; 2800 <D:request p> 2801 <B:modify p=&v> <D:commit p=&v> 2802 <D:read p> 2803 x = *q; 2804 <C:read *q> Reads from v before v updated in cache 2805 <C:unbusy> 2806 <C:commit v=2> 2807 2808Basically, while both cachelines will be updated on CPU 2 eventually, there's 2809no guarantee that, without intervention, the order of update will be the same 2810as that committed on CPU 1. 2811 2812 2813To intervene, we need to interpolate a data dependency barrier or a read 2814barrier between the loads (which as of v4.15 is supplied unconditionally 2815by the READ_ONCE() macro). This will force the cache to commit its 2816coherency queue before processing any further requests: 2817 2818 CPU 1 CPU 2 COMMENT 2819 =============== =============== ======================================= 2820 u == 0, v == 1 and p == &u, q == &u 2821 v = 2; 2822 smp_wmb(); 2823 <A:modify v=2> <C:busy> 2824 <C:queue v=2> 2825 p = &v; q = p; 2826 <D:request p> 2827 <B:modify p=&v> <D:commit p=&v> 2828 <D:read p> 2829 smp_read_barrier_depends() 2830 <C:unbusy> 2831 <C:commit v=2> 2832 x = *q; 2833 <C:read *q> Reads from v after v updated in cache 2834 2835 2836This sort of problem can be encountered on DEC Alpha processors as they have a 2837split cache that improves performance by making better use of the data bus. 2838While most CPUs do imply a data dependency barrier on the read when a memory 2839access depends on a read, not all do, so it may not be relied on. 2840 2841Other CPUs may also have split caches, but must coordinate between the various 2842cachelets for normal memory accesses. The semantics of the Alpha removes the 2843need for hardware coordination in the absence of memory barriers, which 2844permitted Alpha to sport higher CPU clock rates back in the day. However, 2845please note that (again, as of v4.15) smp_read_barrier_depends() should not 2846be used except in Alpha arch-specific code and within the READ_ONCE() macro. 2847 2848 2849CACHE COHERENCY VS DMA 2850---------------------- 2851 2852Not all systems maintain cache coherency with respect to devices doing DMA. In 2853such cases, a device attempting DMA may obtain stale data from RAM because 2854dirty cache lines may be resident in the caches of various CPUs, and may not 2855have been written back to RAM yet. To deal with this, the appropriate part of 2856the kernel must flush the overlapping bits of cache on each CPU (and maybe 2857invalidate them as well). 2858 2859In addition, the data DMA'd to RAM by a device may be overwritten by dirty 2860cache lines being written back to RAM from a CPU's cache after the device has 2861installed its own data, or cache lines present in the CPU's cache may simply 2862obscure the fact that RAM has been updated, until at such time as the cacheline 2863is discarded from the CPU's cache and reloaded. To deal with this, the 2864appropriate part of the kernel must invalidate the overlapping bits of the 2865cache on each CPU. 2866 2867See Documentation/core-api/cachetlb.rst for more information on cache management. 2868 2869 2870CACHE COHERENCY VS MMIO 2871----------------------- 2872 2873Memory mapped I/O usually takes place through memory locations that are part of 2874a window in the CPU's memory space that has different properties assigned than 2875the usual RAM directed window. 2876 2877Amongst these properties is usually the fact that such accesses bypass the 2878caching entirely and go directly to the device buses. This means MMIO accesses 2879may, in effect, overtake accesses to cached memory that were emitted earlier. 2880A memory barrier isn't sufficient in such a case, but rather the cache must be 2881flushed between the cached memory write and the MMIO access if the two are in 2882any way dependent. 2883 2884 2885========================= 2886THE THINGS CPUS GET UP TO 2887========================= 2888 2889A programmer might take it for granted that the CPU will perform memory 2890operations in exactly the order specified, so that if the CPU is, for example, 2891given the following piece of code to execute: 2892 2893 a = READ_ONCE(*A); 2894 WRITE_ONCE(*B, b); 2895 c = READ_ONCE(*C); 2896 d = READ_ONCE(*D); 2897 WRITE_ONCE(*E, e); 2898 2899they would then expect that the CPU will complete the memory operation for each 2900instruction before moving on to the next one, leading to a definite sequence of 2901operations as seen by external observers in the system: 2902 2903 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E. 2904 2905 2906Reality is, of course, much messier. With many CPUs and compilers, the above 2907assumption doesn't hold because: 2908 2909 (*) loads are more likely to need to be completed immediately to permit 2910 execution progress, whereas stores can often be deferred without a 2911 problem; 2912 2913 (*) loads may be done speculatively, and the result discarded should it prove 2914 to have been unnecessary; 2915 2916 (*) loads may be done speculatively, leading to the result having been fetched 2917 at the wrong time in the expected sequence of events; 2918 2919 (*) the order of the memory accesses may be rearranged to promote better use 2920 of the CPU buses and caches; 2921 2922 (*) loads and stores may be combined to improve performance when talking to 2923 memory or I/O hardware that can do batched accesses of adjacent locations, 2924 thus cutting down on transaction setup costs (memory and PCI devices may 2925 both be able to do this); and 2926 2927 (*) the CPU's data cache may affect the ordering, and while cache-coherency 2928 mechanisms may alleviate this - once the store has actually hit the cache 2929 - there's no guarantee that the coherency management will be propagated in 2930 order to other CPUs. 2931 2932So what another CPU, say, might actually observe from the above piece of code 2933is: 2934 2935 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B 2936 2937 (Where "LOAD {*C,*D}" is a combined load) 2938 2939 2940However, it is guaranteed that a CPU will be self-consistent: it will see its 2941_own_ accesses appear to be correctly ordered, without the need for a memory 2942barrier. For instance with the following code: 2943 2944 U = READ_ONCE(*A); 2945 WRITE_ONCE(*A, V); 2946 WRITE_ONCE(*A, W); 2947 X = READ_ONCE(*A); 2948 WRITE_ONCE(*A, Y); 2949 Z = READ_ONCE(*A); 2950 2951and assuming no intervention by an external influence, it can be assumed that 2952the final result will appear to be: 2953 2954 U == the original value of *A 2955 X == W 2956 Z == Y 2957 *A == Y 2958 2959The code above may cause the CPU to generate the full sequence of memory 2960accesses: 2961 2962 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A 2963 2964in that order, but, without intervention, the sequence may have almost any 2965combination of elements combined or discarded, provided the program's view 2966of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE() 2967are -not- optional in the above example, as there are architectures 2968where a given CPU might reorder successive loads to the same location. 2969On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is 2970necessary to prevent this, for example, on Itanium the volatile casts 2971used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq 2972and st.rel instructions (respectively) that prevent such reordering. 2973 2974The compiler may also combine, discard or defer elements of the sequence before 2975the CPU even sees them. 2976 2977For instance: 2978 2979 *A = V; 2980 *A = W; 2981 2982may be reduced to: 2983 2984 *A = W; 2985 2986since, without either a write barrier or an WRITE_ONCE(), it can be 2987assumed that the effect of the storage of V to *A is lost. Similarly: 2988 2989 *A = Y; 2990 Z = *A; 2991 2992may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be 2993reduced to: 2994 2995 *A = Y; 2996 Z = Y; 2997 2998and the LOAD operation never appear outside of the CPU. 2999 3000 3001AND THEN THERE'S THE ALPHA 3002-------------------------- 3003 3004The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that, 3005some versions of the Alpha CPU have a split data cache, permitting them to have 3006two semantically-related cache lines updated at separate times. This is where 3007the data dependency barrier really becomes necessary as this synchronises both 3008caches with the memory coherence system, thus making it seem like pointer 3009changes vs new data occur in the right order. 3010 3011The Alpha defines the Linux kernel's memory model, although as of v4.15 3012the Linux kernel's addition of smp_read_barrier_depends() to READ_ONCE() 3013greatly reduced Alpha's impact on the memory model. 3014 3015See the subsection on "Cache Coherency" above. 3016 3017 3018VIRTUAL MACHINE GUESTS 3019---------------------- 3020 3021Guests running within virtual machines might be affected by SMP effects even if 3022the guest itself is compiled without SMP support. This is an artifact of 3023interfacing with an SMP host while running an UP kernel. Using mandatory 3024barriers for this use-case would be possible but is often suboptimal. 3025 3026To handle this case optimally, low-level virt_mb() etc macros are available. 3027These have the same effect as smp_mb() etc when SMP is enabled, but generate 3028identical code for SMP and non-SMP systems. For example, virtual machine guests 3029should use virt_mb() rather than smp_mb() when synchronizing against a 3030(possibly SMP) host. 3031 3032These are equivalent to smp_mb() etc counterparts in all other respects, 3033in particular, they do not control MMIO effects: to control 3034MMIO effects, use mandatory barriers. 3035 3036 3037============ 3038EXAMPLE USES 3039============ 3040 3041CIRCULAR BUFFERS 3042---------------- 3043 3044Memory barriers can be used to implement circular buffering without the need 3045of a lock to serialise the producer with the consumer. See: 3046 3047 Documentation/core-api/circular-buffers.rst 3048 3049for details. 3050 3051 3052========== 3053REFERENCES 3054========== 3055 3056Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek, 3057Digital Press) 3058 Chapter 5.2: Physical Address Space Characteristics 3059 Chapter 5.4: Caches and Write Buffers 3060 Chapter 5.5: Data Sharing 3061 Chapter 5.6: Read/Write Ordering 3062 3063AMD64 Architecture Programmer's Manual Volume 2: System Programming 3064 Chapter 7.1: Memory-Access Ordering 3065 Chapter 7.4: Buffering and Combining Memory Writes 3066 3067ARM Architecture Reference Manual (ARMv8, for ARMv8-A architecture profile) 3068 Chapter B2: The AArch64 Application Level Memory Model 3069 3070IA-32 Intel Architecture Software Developer's Manual, Volume 3: 3071System Programming Guide 3072 Chapter 7.1: Locked Atomic Operations 3073 Chapter 7.2: Memory Ordering 3074 Chapter 7.4: Serializing Instructions 3075 3076The SPARC Architecture Manual, Version 9 3077 Chapter 8: Memory Models 3078 Appendix D: Formal Specification of the Memory Models 3079 Appendix J: Programming with the Memory Models 3080 3081Storage in the PowerPC (Stone and Fitzgerald) 3082 3083UltraSPARC Programmer Reference Manual 3084 Chapter 5: Memory Accesses and Cacheability 3085 Chapter 15: Sparc-V9 Memory Models 3086 3087UltraSPARC III Cu User's Manual 3088 Chapter 9: Memory Models 3089 3090UltraSPARC IIIi Processor User's Manual 3091 Chapter 8: Memory Models 3092 3093UltraSPARC Architecture 2005 3094 Chapter 9: Memory 3095 Appendix D: Formal Specifications of the Memory Models 3096 3097UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005 3098 Chapter 8: Memory Models 3099 Appendix F: Caches and Cache Coherency 3100 3101Solaris Internals, Core Kernel Architecture, p63-68: 3102 Chapter 3.3: Hardware Considerations for Locks and 3103 Synchronization 3104 3105Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching 3106for Kernel Programmers: 3107 Chapter 13: Other Memory Models 3108 3109Intel Itanium Architecture Software Developer's Manual: Volume 1: 3110 Section 2.6: Speculation 3111 Section 4.4: Memory Access 3112