1 ===================== 1 ============================ 2 LINUX KERNEL MEMORY B 2 LINUX KERNEL MEMORY BARRIERS 3 ===================== 3 ============================ 4 4 5 By: David Howells <dhowells@redhat.com> 5 By: David Howells <dhowells@redhat.com> 6 Paul E. McKenney <paulmck@linux.ibm.com> 6 Paul E. McKenney <paulmck@linux.ibm.com> 7 Will Deacon <will.deacon@arm.com> 7 Will Deacon <will.deacon@arm.com> 8 Peter Zijlstra <peterz@infradead.org> 8 Peter Zijlstra <peterz@infradead.org> 9 9 10 ========== 10 ========== 11 DISCLAIMER 11 DISCLAIMER 12 ========== 12 ========== 13 13 14 This document is not a specification; it is in 14 This document is not a specification; it is intentionally (for the sake of 15 brevity) and unintentionally (due to being hum 15 brevity) and unintentionally (due to being human) incomplete. This document is 16 meant as a guide to using the various memory b 16 meant as a guide to using the various memory barriers provided by Linux, but 17 in case of any doubt (and there are many) plea 17 in case of any doubt (and there are many) please ask. Some doubts may be 18 resolved by referring to the formal memory con 18 resolved by referring to the formal memory consistency model and related 19 documentation at tools/memory-model/. Neverth 19 documentation at tools/memory-model/. Nevertheless, even this memory 20 model should be viewed as the collective opini 20 model should be viewed as the collective opinion of its maintainers rather 21 than as an infallible oracle. 21 than as an infallible oracle. 22 22 23 To repeat, this document is not a specificatio 23 To repeat, this document is not a specification of what Linux expects from 24 hardware. 24 hardware. 25 25 26 The purpose of this document is twofold: 26 The purpose of this document is twofold: 27 27 28 (1) to specify the minimum functionality that 28 (1) to specify the minimum functionality that one can rely on for any 29 particular barrier, and 29 particular barrier, and 30 30 31 (2) to provide a guide as to how to use the b 31 (2) to provide a guide as to how to use the barriers that are available. 32 32 33 Note that an architecture can provide more tha 33 Note that an architecture can provide more than the minimum requirement 34 for any particular barrier, but if the archite 34 for any particular barrier, but if the architecture provides less than 35 that, that architecture is incorrect. 35 that, that architecture is incorrect. 36 36 37 Note also that it is possible that a barrier m 37 Note also that it is possible that a barrier may be a no-op for an 38 architecture because the way that arch works r 38 architecture because the way that arch works renders an explicit barrier 39 unnecessary in that case. 39 unnecessary in that case. 40 40 41 41 42 ======== 42 ======== 43 CONTENTS 43 CONTENTS 44 ======== 44 ======== 45 45 46 (*) Abstract memory access model. 46 (*) Abstract memory access model. 47 47 48 - Device operations. 48 - Device operations. 49 - Guarantees. 49 - Guarantees. 50 50 51 (*) What are memory barriers? 51 (*) What are memory barriers? 52 52 53 - Varieties of memory barrier. 53 - Varieties of memory barrier. 54 - What may not be assumed about memory ba 54 - What may not be assumed about memory barriers? 55 - Address-dependency barriers (historical !! 55 - Data dependency barriers (historical). 56 - Control dependencies. 56 - Control dependencies. 57 - SMP barrier pairing. 57 - SMP barrier pairing. 58 - Examples of memory barrier sequences. 58 - Examples of memory barrier sequences. 59 - Read memory barriers vs load speculatio 59 - Read memory barriers vs load speculation. 60 - Multicopy atomicity. 60 - Multicopy atomicity. 61 61 62 (*) Explicit kernel barriers. 62 (*) Explicit kernel barriers. 63 63 64 - Compiler barrier. 64 - Compiler barrier. 65 - CPU memory barriers. 65 - CPU memory barriers. 66 66 67 (*) Implicit kernel memory barriers. 67 (*) Implicit kernel memory barriers. 68 68 69 - Lock acquisition functions. 69 - Lock acquisition functions. 70 - Interrupt disabling functions. 70 - Interrupt disabling functions. 71 - Sleep and wake-up functions. 71 - Sleep and wake-up functions. 72 - Miscellaneous functions. 72 - Miscellaneous functions. 73 73 74 (*) Inter-CPU acquiring barrier effects. 74 (*) Inter-CPU acquiring barrier effects. 75 75 76 - Acquires vs memory accesses. 76 - Acquires vs memory accesses. 77 77 78 (*) Where are memory barriers needed? 78 (*) Where are memory barriers needed? 79 79 80 - Interprocessor interaction. 80 - Interprocessor interaction. 81 - Atomic operations. 81 - Atomic operations. 82 - Accessing devices. 82 - Accessing devices. 83 - Interrupts. 83 - Interrupts. 84 84 85 (*) Kernel I/O barrier effects. 85 (*) Kernel I/O barrier effects. 86 86 87 (*) Assumed minimum execution ordering model. 87 (*) Assumed minimum execution ordering model. 88 88 89 (*) The effects of the cpu cache. 89 (*) The effects of the cpu cache. 90 90 91 - Cache coherency. 91 - Cache coherency. 92 - Cache coherency vs DMA. 92 - Cache coherency vs DMA. 93 - Cache coherency vs MMIO. 93 - Cache coherency vs MMIO. 94 94 95 (*) The things CPUs get up to. 95 (*) The things CPUs get up to. 96 96 97 - And then there's the Alpha. 97 - And then there's the Alpha. 98 - Virtual Machine Guests. 98 - Virtual Machine Guests. 99 99 100 (*) Example uses. 100 (*) Example uses. 101 101 102 - Circular buffers. 102 - Circular buffers. 103 103 104 (*) References. 104 (*) References. 105 105 106 106 107 ============================ 107 ============================ 108 ABSTRACT MEMORY ACCESS MODEL 108 ABSTRACT MEMORY ACCESS MODEL 109 ============================ 109 ============================ 110 110 111 Consider the following abstract model of the s 111 Consider the following abstract model of the system: 112 112 113 : : 113 : : 114 : : 114 : : 115 : : 115 : : 116 +-------+ : +--------+ : 116 +-------+ : +--------+ : +-------+ 117 | | : | | : 117 | | : | | : | | 118 | | : | | : 118 | | : | | : | | 119 | CPU 1 |<----->| Memory |<--- 119 | CPU 1 |<----->| Memory |<----->| CPU 2 | 120 | | : | | : 120 | | : | | : | | 121 | | : | | : 121 | | : | | : | | 122 +-------+ : +--------+ : 122 +-------+ : +--------+ : +-------+ 123 ^ : ^ : 123 ^ : ^ : ^ 124 | : | : 124 | : | : | 125 | : | : 125 | : | : | 126 | : v : 126 | : v : | 127 | : +--------+ : 127 | : +--------+ : | 128 | : | | : 128 | : | | : | 129 | : | | : 129 | : | | : | 130 +---------->| Device |<--- 130 +---------->| Device |<----------+ 131 : | | : 131 : | | : 132 : | | : 132 : | | : 133 : +--------+ : 133 : +--------+ : 134 : : 134 : : 135 135 136 Each CPU executes a program that generates mem 136 Each CPU executes a program that generates memory access operations. In the 137 abstract CPU, memory operation ordering is ver 137 abstract CPU, memory operation ordering is very relaxed, and a CPU may actually 138 perform the memory operations in any order it 138 perform the memory operations in any order it likes, provided program causality 139 appears to be maintained. Similarly, the comp 139 appears to be maintained. Similarly, the compiler may also arrange the 140 instructions it emits in any order it likes, p 140 instructions it emits in any order it likes, provided it doesn't affect the 141 apparent operation of the program. 141 apparent operation of the program. 142 142 143 So in the above diagram, the effects of the me 143 So in the above diagram, the effects of the memory operations performed by a 144 CPU are perceived by the rest of the system as 144 CPU are perceived by the rest of the system as the operations cross the 145 interface between the CPU and rest of the syst 145 interface between the CPU and rest of the system (the dotted lines). 146 146 147 147 148 For example, consider the following sequence o 148 For example, consider the following sequence of events: 149 149 150 CPU 1 CPU 2 150 CPU 1 CPU 2 151 =============== =============== 151 =============== =============== 152 { A == 1; B == 2 } 152 { A == 1; B == 2 } 153 A = 3; x = B; 153 A = 3; x = B; 154 B = 4; y = A; 154 B = 4; y = A; 155 155 156 The set of accesses as seen by the memory syst 156 The set of accesses as seen by the memory system in the middle can be arranged 157 in 24 different combinations: 157 in 24 different combinations: 158 158 159 STORE A=3, STORE B=4, y=LOAD 159 STORE A=3, STORE B=4, y=LOAD A->3, x=LOAD B->4 160 STORE A=3, STORE B=4, x=LOAD 160 STORE A=3, STORE B=4, x=LOAD B->4, y=LOAD A->3 161 STORE A=3, y=LOAD A->3, STORE 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 162 STORE A=3, y=LOAD A->3, x=LOAD B->2, STORE B=4 163 STORE A=3, x=LOAD B->2, STORE 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 164 STORE A=3, x=LOAD B->2, y=LOAD A->3, STORE B=4 165 STORE B=4, STORE A=3, y=LOAD 165 STORE B=4, STORE A=3, y=LOAD A->3, x=LOAD B->4 166 STORE B=4, ... 166 STORE B=4, ... 167 ... 167 ... 168 168 169 and can thus result in four different combinat 169 and can thus result in four different combinations of values: 170 170 171 x == 2, y == 1 171 x == 2, y == 1 172 x == 2, y == 3 172 x == 2, y == 3 173 x == 4, y == 1 173 x == 4, y == 1 174 x == 4, y == 3 174 x == 4, y == 3 175 175 176 176 177 Furthermore, the stores committed by a CPU to 177 Furthermore, the stores committed by a CPU to the memory system may not be 178 perceived by the loads made by another CPU in 178 perceived by the loads made by another CPU in the same order as the stores were 179 committed. 179 committed. 180 180 181 181 182 As a further example, consider this sequence o 182 As a further example, consider this sequence of events: 183 183 184 CPU 1 CPU 2 184 CPU 1 CPU 2 185 =============== =============== 185 =============== =============== 186 { A == 1, B == 2, C == 3, P == &A, Q = 186 { A == 1, B == 2, C == 3, P == &A, Q == &C } 187 B = 4; Q = P; 187 B = 4; Q = P; 188 P = &B; D = *Q; 188 P = &B; D = *Q; 189 189 190 There is an obvious address dependency here, a !! 190 There is an obvious data dependency here, as the value loaded into D depends on 191 on the address retrieved from P by CPU 2. At !! 191 the address retrieved from P by CPU 2. At the end of the sequence, any of the 192 the following results are possible: !! 192 following results are possible: 193 193 194 (Q == &A) and (D == 1) 194 (Q == &A) and (D == 1) 195 (Q == &B) and (D == 2) 195 (Q == &B) and (D == 2) 196 (Q == &B) and (D == 4) 196 (Q == &B) and (D == 4) 197 197 198 Note that CPU 2 will never try and load C into 198 Note that CPU 2 will never try and load C into D because the CPU will load P 199 into Q before issuing the load of *Q. 199 into Q before issuing the load of *Q. 200 200 201 201 202 DEVICE OPERATIONS 202 DEVICE OPERATIONS 203 ----------------- 203 ----------------- 204 204 205 Some devices present their control interfaces 205 Some devices present their control interfaces as collections of memory 206 locations, but the order in which the control 206 locations, but the order in which the control registers are accessed is very 207 important. For instance, imagine an ethernet 207 important. For instance, imagine an ethernet card with a set of internal 208 registers that are accessed through an address 208 registers that are accessed through an address port register (A) and a data 209 port register (D). To read internal register 209 port register (D). To read internal register 5, the following code might then 210 be used: 210 be used: 211 211 212 *A = 5; 212 *A = 5; 213 x = *D; 213 x = *D; 214 214 215 but this might show up as either of the follow 215 but this might show up as either of the following two sequences: 216 216 217 STORE *A = 5, x = LOAD *D 217 STORE *A = 5, x = LOAD *D 218 x = LOAD *D, STORE *A = 5 218 x = LOAD *D, STORE *A = 5 219 219 220 the second of which will almost certainly resu 220 the second of which will almost certainly result in a malfunction, since it set 221 the address _after_ attempting to read the reg 221 the address _after_ attempting to read the register. 222 222 223 223 224 GUARANTEES 224 GUARANTEES 225 ---------- 225 ---------- 226 226 227 There are some minimal guarantees that may be 227 There are some minimal guarantees that may be expected of a CPU: 228 228 229 (*) On any given CPU, dependent memory access 229 (*) On any given CPU, dependent memory accesses will be issued in order, with 230 respect to itself. This means that for: 230 respect to itself. This means that for: 231 231 232 Q = READ_ONCE(P); D = READ_ONCE(*Q); 232 Q = READ_ONCE(P); D = READ_ONCE(*Q); 233 233 234 the CPU will issue the following memory o 234 the CPU will issue the following memory operations: 235 235 236 Q = LOAD P, D = LOAD *Q 236 Q = LOAD P, D = LOAD *Q 237 237 238 and always in that order. However, on DE 238 and always in that order. However, on DEC Alpha, READ_ONCE() also 239 emits a memory-barrier instruction, so th 239 emits a memory-barrier instruction, so that a DEC Alpha CPU will 240 instead issue the following memory operat 240 instead issue the following memory operations: 241 241 242 Q = LOAD P, MEMORY_BARRIER, D = LOAD * 242 Q = LOAD P, MEMORY_BARRIER, D = LOAD *Q, MEMORY_BARRIER 243 243 244 Whether on DEC Alpha or not, the READ_ONC 244 Whether on DEC Alpha or not, the READ_ONCE() also prevents compiler 245 mischief. 245 mischief. 246 246 247 (*) Overlapping loads and stores within a par 247 (*) Overlapping loads and stores within a particular CPU will appear to be 248 ordered within that CPU. This means that 248 ordered within that CPU. This means that for: 249 249 250 a = READ_ONCE(*X); WRITE_ONCE(*X, b); 250 a = READ_ONCE(*X); WRITE_ONCE(*X, b); 251 251 252 the CPU will only issue the following seq 252 the CPU will only issue the following sequence of memory operations: 253 253 254 a = LOAD *X, STORE *X = b 254 a = LOAD *X, STORE *X = b 255 255 256 And for: 256 And for: 257 257 258 WRITE_ONCE(*X, c); d = READ_ONCE(*X); 258 WRITE_ONCE(*X, c); d = READ_ONCE(*X); 259 259 260 the CPU will only issue: 260 the CPU will only issue: 261 261 262 STORE *X = c, d = LOAD *X 262 STORE *X = c, d = LOAD *X 263 263 264 (Loads and stores overlap if they are tar 264 (Loads and stores overlap if they are targeted at overlapping pieces of 265 memory). 265 memory). 266 266 267 And there are a number of things that _must_ o 267 And there are a number of things that _must_ or _must_not_ be assumed: 268 268 269 (*) It _must_not_ be assumed that the compile 269 (*) It _must_not_ be assumed that the compiler will do what you want 270 with memory references that are not prote 270 with memory references that are not protected by READ_ONCE() and 271 WRITE_ONCE(). Without them, the compiler 271 WRITE_ONCE(). Without them, the compiler is within its rights to 272 do all sorts of "creative" transformation 272 do all sorts of "creative" transformations, which are covered in 273 the COMPILER BARRIER section. 273 the COMPILER BARRIER section. 274 274 275 (*) It _must_not_ be assumed that independent 275 (*) It _must_not_ be assumed that independent loads and stores will be issued 276 in the order given. This means that for: 276 in the order given. This means that for: 277 277 278 X = *A; Y = *B; *D = Z; 278 X = *A; Y = *B; *D = Z; 279 279 280 we may get any of the following sequences 280 we may get any of the following sequences: 281 281 282 X = LOAD *A, Y = LOAD *B, STORE *D = 282 X = LOAD *A, Y = LOAD *B, STORE *D = Z 283 X = LOAD *A, STORE *D = Z, Y = LOAD * 283 X = LOAD *A, STORE *D = Z, Y = LOAD *B 284 Y = LOAD *B, X = LOAD *A, STORE *D = 284 Y = LOAD *B, X = LOAD *A, STORE *D = Z 285 Y = LOAD *B, STORE *D = Z, X = LOAD * 285 Y = LOAD *B, STORE *D = Z, X = LOAD *A 286 STORE *D = Z, X = LOAD *A, Y = LOAD * 286 STORE *D = Z, X = LOAD *A, Y = LOAD *B 287 STORE *D = Z, Y = LOAD *B, X = LOAD * 287 STORE *D = Z, Y = LOAD *B, X = LOAD *A 288 288 289 (*) It _must_ be assumed that overlapping mem 289 (*) It _must_ be assumed that overlapping memory accesses may be merged or 290 discarded. This means that for: 290 discarded. This means that for: 291 291 292 X = *A; Y = *(A + 4); 292 X = *A; Y = *(A + 4); 293 293 294 we may get any one of the following seque 294 we may get any one of the following sequences: 295 295 296 X = LOAD *A; Y = LOAD *(A + 4); 296 X = LOAD *A; Y = LOAD *(A + 4); 297 Y = LOAD *(A + 4); X = LOAD *A; 297 Y = LOAD *(A + 4); X = LOAD *A; 298 {X, Y} = LOAD {*A, *(A + 4) }; 298 {X, Y} = LOAD {*A, *(A + 4) }; 299 299 300 And for: 300 And for: 301 301 302 *A = X; *(A + 4) = Y; 302 *A = X; *(A + 4) = Y; 303 303 304 we may get any of: 304 we may get any of: 305 305 306 STORE *A = X; STORE *(A + 4) = Y; 306 STORE *A = X; STORE *(A + 4) = Y; 307 STORE *(A + 4) = Y; STORE *A = X; 307 STORE *(A + 4) = Y; STORE *A = X; 308 STORE {*A, *(A + 4) } = {X, Y}; 308 STORE {*A, *(A + 4) } = {X, Y}; 309 309 310 And there are anti-guarantees: 310 And there are anti-guarantees: 311 311 312 (*) These guarantees do not apply to bitfield 312 (*) These guarantees do not apply to bitfields, because compilers often 313 generate code to modify these using non-a 313 generate code to modify these using non-atomic read-modify-write 314 sequences. Do not attempt to use bitfiel 314 sequences. Do not attempt to use bitfields to synchronize parallel 315 algorithms. 315 algorithms. 316 316 317 (*) Even in cases where bitfields are protect 317 (*) Even in cases where bitfields are protected by locks, all fields 318 in a given bitfield must be protected by 318 in a given bitfield must be protected by one lock. If two fields 319 in a given bitfield are protected by diff 319 in a given bitfield are protected by different locks, the compiler's 320 non-atomic read-modify-write sequences ca 320 non-atomic read-modify-write sequences can cause an update to one 321 field to corrupt the value of an adjacent 321 field to corrupt the value of an adjacent field. 322 322 323 (*) These guarantees apply only to properly a 323 (*) These guarantees apply only to properly aligned and sized scalar 324 variables. "Properly sized" currently me 324 variables. "Properly sized" currently means variables that are 325 the same size as "char", "short", "int" a 325 the same size as "char", "short", "int" and "long". "Properly 326 aligned" means the natural alignment, thu 326 aligned" means the natural alignment, thus no constraints for 327 "char", two-byte alignment for "short", f 327 "char", two-byte alignment for "short", four-byte alignment for 328 "int", and either four-byte or eight-byte 328 "int", and either four-byte or eight-byte alignment for "long", 329 on 32-bit and 64-bit systems, respectivel 329 on 32-bit and 64-bit systems, respectively. Note that these 330 guarantees were introduced into the C11 s 330 guarantees were introduced into the C11 standard, so beware when 331 using older pre-C11 compilers (for exampl 331 using older pre-C11 compilers (for example, gcc 4.6). The portion 332 of the standard containing this guarantee 332 of the standard containing this guarantee is Section 3.14, which 333 defines "memory location" as follows: 333 defines "memory location" as follows: 334 334 335 memory location 335 memory location 336 either an object of scalar typ 336 either an object of scalar type, or a maximal sequence 337 of adjacent bit-fields all hav 337 of adjacent bit-fields all having nonzero width 338 338 339 NOTE 1: Two threads of executi 339 NOTE 1: Two threads of execution can update and access 340 separate memory locations with 340 separate memory locations without interfering with 341 each other. 341 each other. 342 342 343 NOTE 2: A bit-field and an adj 343 NOTE 2: A bit-field and an adjacent non-bit-field member 344 are in separate memory locatio 344 are in separate memory locations. The same applies 345 to two bit-fields, if one is d 345 to two bit-fields, if one is declared inside a nested 346 structure declaration and the 346 structure declaration and the other is not, or if the two 347 are separated by a zero-length 347 are separated by a zero-length bit-field declaration, 348 or if they are separated by a 348 or if they are separated by a non-bit-field member 349 declaration. It is not safe to 349 declaration. It is not safe to concurrently update two 350 bit-fields in the same structu 350 bit-fields in the same structure if all members declared 351 between them are also bit-fiel 351 between them are also bit-fields, no matter what the 352 sizes of those intervening bit 352 sizes of those intervening bit-fields happen to be. 353 353 354 354 355 ========================= 355 ========================= 356 WHAT ARE MEMORY BARRIERS? 356 WHAT ARE MEMORY BARRIERS? 357 ========================= 357 ========================= 358 358 359 As can be seen above, independent memory opera 359 As can be seen above, independent memory operations are effectively performed 360 in random order, but this can be a problem for 360 in random order, but this can be a problem for CPU-CPU interaction and for I/O. 361 What is required is some way of intervening to 361 What is required is some way of intervening to instruct the compiler and the 362 CPU to restrict the order. 362 CPU to restrict the order. 363 363 364 Memory barriers are such interventions. They 364 Memory barriers are such interventions. They impose a perceived partial 365 ordering over the memory operations on either 365 ordering over the memory operations on either side of the barrier. 366 366 367 Such enforcement is important because the CPUs 367 Such enforcement is important because the CPUs and other devices in a system 368 can use a variety of tricks to improve perform 368 can use a variety of tricks to improve performance, including reordering, 369 deferral and combination of memory operations; 369 deferral and combination of memory operations; speculative loads; speculative 370 branch prediction and various types of caching 370 branch prediction and various types of caching. Memory barriers are used to 371 override or suppress these tricks, allowing th 371 override or suppress these tricks, allowing the code to sanely control the 372 interaction of multiple CPUs and/or devices. 372 interaction of multiple CPUs and/or devices. 373 373 374 374 375 VARIETIES OF MEMORY BARRIER 375 VARIETIES OF MEMORY BARRIER 376 --------------------------- 376 --------------------------- 377 377 378 Memory barriers come in four basic varieties: 378 Memory barriers come in four basic varieties: 379 379 380 (1) Write (or store) memory barriers. 380 (1) Write (or store) memory barriers. 381 381 382 A write memory barrier gives a guarantee 382 A write memory barrier gives a guarantee that all the STORE operations 383 specified before the barrier will appear 383 specified before the barrier will appear to happen before all the STORE 384 operations specified after the barrier wi 384 operations specified after the barrier with respect to the other 385 components of the system. 385 components of the system. 386 386 387 A write barrier is a partial ordering on 387 A write barrier is a partial ordering on stores only; it is not required 388 to have any effect on loads. 388 to have any effect on loads. 389 389 390 A CPU can be viewed as committing a seque 390 A CPU can be viewed as committing a sequence of store operations to the 391 memory system as time progresses. All st 391 memory system as time progresses. All stores _before_ a write barrier 392 will occur _before_ all the stores after 392 will occur _before_ all the stores after the write barrier. 393 393 394 [!] Note that write barriers should norma !! 394 [!] Note that write barriers should normally be paired with read or data 395 address-dependency barriers; see the "SMP !! 395 dependency barriers; see the "SMP barrier pairing" subsection. 396 396 397 397 398 (2) Address-dependency barriers (historical). !! 398 (2) Data dependency barriers. 399 [!] This section is marked as HISTORICAL: !! 399 400 smp_read_barrier_depends() macro, the sem !! 400 A data dependency barrier is a weaker form of read barrier. In the case 401 implicit in all marked accesses. For mor !! 401 where two loads are performed such that the second depends on the result 402 including how compiler transformations ca !! 402 of the first (eg: the first load retrieves the address to which the second 403 dependencies, see Documentation/RCU/rcu_d !! 403 load will be directed), a data dependency barrier would be required to 404 !! 404 make sure that the target of the second load is updated after the address 405 An address-dependency barrier is a weaker !! 405 obtained by the first load is accessed. 406 case where two loads are performed such t !! 406 407 result of the first (eg: the first load r !! 407 A data dependency barrier is a partial ordering on interdependent loads 408 the second load will be directed), an add !! 408 only; it is not required to have any effect on stores, independent loads 409 be required to make sure that the target !! 409 or overlapping loads. 410 after the address obtained by the first l << 411 << 412 An address-dependency barrier is a partia << 413 loads only; it is not required to have an << 414 loads or overlapping loads. << 415 410 416 As mentioned in (1), the other CPUs in th 411 As mentioned in (1), the other CPUs in the system can be viewed as 417 committing sequences of stores to the mem 412 committing sequences of stores to the memory system that the CPU being 418 considered can then perceive. An address !! 413 considered can then perceive. A data dependency barrier issued by the CPU 419 the CPU under consideration guarantees th !! 414 under consideration guarantees that for any load preceding it, if that 420 if that load touches one of a sequence of !! 415 load touches one of a sequence of stores from another CPU, then by the 421 by the time the barrier completes, the ef !! 416 time the barrier completes, the effects of all the stores prior to that 422 that touched by the load will be percepti !! 417 touched by the load will be perceptible to any loads issued after the data 423 the address-dependency barrier. !! 418 dependency barrier. 424 419 425 See the "Examples of memory barrier seque 420 See the "Examples of memory barrier sequences" subsection for diagrams 426 showing the ordering constraints. 421 showing the ordering constraints. 427 422 428 [!] Note that the first load really has t !! 423 [!] Note that the first load really has to have a _data_ dependency and 429 not a control dependency. If the address 424 not a control dependency. If the address for the second load is dependent 430 on the first load, but the dependency is 425 on the first load, but the dependency is through a conditional rather than 431 actually loading the address itself, then 426 actually loading the address itself, then it's a _control_ dependency and 432 a full read barrier or better is required 427 a full read barrier or better is required. See the "Control dependencies" 433 subsection for more information. 428 subsection for more information. 434 429 435 [!] Note that address-dependency barriers !! 430 [!] Note that data dependency barriers should normally be paired with 436 write barriers; see the "SMP barrier pair 431 write barriers; see the "SMP barrier pairing" subsection. 437 432 438 [!] Kernel release v5.9 removed kernel AP << 439 dependency barriers. Nowadays, APIs for << 440 variables such as READ_ONCE() and rcu_der << 441 address-dependency barriers. << 442 433 443 (3) Read (or load) memory barriers. 434 (3) Read (or load) memory barriers. 444 435 445 A read barrier is an address-dependency b !! 436 A read barrier is a data dependency barrier plus a guarantee that all the 446 the LOAD operations specified before the !! 437 LOAD operations specified before the barrier will appear to happen before 447 before all the LOAD operations specified !! 438 all the LOAD operations specified after the barrier with respect to the 448 the other components of the system. !! 439 other components of the system. 449 440 450 A read barrier is a partial ordering on l 441 A read barrier is a partial ordering on loads only; it is not required to 451 have any effect on stores. 442 have any effect on stores. 452 443 453 Read memory barriers imply address-depend !! 444 Read memory barriers imply data dependency barriers, and so can substitute 454 substitute for them. !! 445 for them. 455 446 456 [!] Note that read barriers should normal 447 [!] Note that read barriers should normally be paired with write barriers; 457 see the "SMP barrier pairing" subsection. 448 see the "SMP barrier pairing" subsection. 458 449 459 450 460 (4) General memory barriers. 451 (4) General memory barriers. 461 452 462 A general memory barrier gives a guarante 453 A general memory barrier gives a guarantee that all the LOAD and STORE 463 operations specified before the barrier w 454 operations specified before the barrier will appear to happen before all 464 the LOAD and STORE operations specified a 455 the LOAD and STORE operations specified after the barrier with respect to 465 the other components of the system. 456 the other components of the system. 466 457 467 A general memory barrier is a partial ord 458 A general memory barrier is a partial ordering over both loads and stores. 468 459 469 General memory barriers imply both read a 460 General memory barriers imply both read and write memory barriers, and so 470 can substitute for either. 461 can substitute for either. 471 462 472 463 473 And a couple of implicit varieties: 464 And a couple of implicit varieties: 474 465 475 (5) ACQUIRE operations. 466 (5) ACQUIRE operations. 476 467 477 This acts as a one-way permeable barrier. 468 This acts as a one-way permeable barrier. It guarantees that all memory 478 operations after the ACQUIRE operation wi 469 operations after the ACQUIRE operation will appear to happen after the 479 ACQUIRE operation with respect to the oth 470 ACQUIRE operation with respect to the other components of the system. 480 ACQUIRE operations include LOCK operation 471 ACQUIRE operations include LOCK operations and both smp_load_acquire() 481 and smp_cond_load_acquire() operations. 472 and smp_cond_load_acquire() operations. 482 473 483 Memory operations that occur before an AC 474 Memory operations that occur before an ACQUIRE operation may appear to 484 happen after it completes. 475 happen after it completes. 485 476 486 An ACQUIRE operation should almost always 477 An ACQUIRE operation should almost always be paired with a RELEASE 487 operation. 478 operation. 488 479 489 480 490 (6) RELEASE operations. 481 (6) RELEASE operations. 491 482 492 This also acts as a one-way permeable bar 483 This also acts as a one-way permeable barrier. It guarantees that all 493 memory operations before the RELEASE oper 484 memory operations before the RELEASE operation will appear to happen 494 before the RELEASE operation with respect 485 before the RELEASE operation with respect to the other components of the 495 system. RELEASE operations include UNLOCK 486 system. RELEASE operations include UNLOCK operations and 496 smp_store_release() operations. 487 smp_store_release() operations. 497 488 498 Memory operations that occur after a RELE 489 Memory operations that occur after a RELEASE operation may appear to 499 happen before it completes. 490 happen before it completes. 500 491 501 The use of ACQUIRE and RELEASE operations 492 The use of ACQUIRE and RELEASE operations generally precludes the need 502 for other sorts of memory barrier. In ad 493 for other sorts of memory barrier. In addition, a RELEASE+ACQUIRE pair is 503 -not- guaranteed to act as a full memory 494 -not- guaranteed to act as a full memory barrier. However, after an 504 ACQUIRE on a given variable, all memory a 495 ACQUIRE on a given variable, all memory accesses preceding any prior 505 RELEASE on that same variable are guarant 496 RELEASE on that same variable are guaranteed to be visible. In other 506 words, within a given variable's critical 497 words, within a given variable's critical section, all accesses of all 507 previous critical sections for that varia 498 previous critical sections for that variable are guaranteed to have 508 completed. 499 completed. 509 500 510 This means that ACQUIRE acts as a minimal 501 This means that ACQUIRE acts as a minimal "acquire" operation and 511 RELEASE acts as a minimal "release" opera 502 RELEASE acts as a minimal "release" operation. 512 503 513 A subset of the atomic operations described in 504 A subset of the atomic operations described in atomic_t.txt have ACQUIRE and 514 RELEASE variants in addition to fully-ordered 505 RELEASE variants in addition to fully-ordered and relaxed (no barrier 515 semantics) definitions. For compound atomics 506 semantics) definitions. For compound atomics performing both a load and a 516 store, ACQUIRE semantics apply only to the loa 507 store, ACQUIRE semantics apply only to the load and RELEASE semantics apply 517 only to the store portion of the operation. 508 only to the store portion of the operation. 518 509 519 Memory barriers are only required where there' 510 Memory barriers are only required where there's a possibility of interaction 520 between two CPUs or between a CPU and a device 511 between two CPUs or between a CPU and a device. If it can be guaranteed that 521 there won't be any such interaction in any par 512 there won't be any such interaction in any particular piece of code, then 522 memory barriers are unnecessary in that piece 513 memory barriers are unnecessary in that piece of code. 523 514 524 515 525 Note that these are the _minimum_ guarantees. 516 Note that these are the _minimum_ guarantees. Different architectures may give 526 more substantial guarantees, but they may _not 517 more substantial guarantees, but they may _not_ be relied upon outside of arch 527 specific code. 518 specific code. 528 519 529 520 530 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS? 521 WHAT MAY NOT BE ASSUMED ABOUT MEMORY BARRIERS? 531 ---------------------------------------------- 522 ---------------------------------------------- 532 523 533 There are certain things that the Linux kernel 524 There are certain things that the Linux kernel memory barriers do not guarantee: 534 525 535 (*) There is no guarantee that any of the mem 526 (*) There is no guarantee that any of the memory accesses specified before a 536 memory barrier will be _complete_ by the 527 memory barrier will be _complete_ by the completion of a memory barrier 537 instruction; the barrier can be considere 528 instruction; the barrier can be considered to draw a line in that CPU's 538 access queue that accesses of the appropr 529 access queue that accesses of the appropriate type may not cross. 539 530 540 (*) There is no guarantee that issuing a memo 531 (*) There is no guarantee that issuing a memory barrier on one CPU will have 541 any direct effect on another CPU or any o 532 any direct effect on another CPU or any other hardware in the system. The 542 indirect effect will be the order in whic 533 indirect effect will be the order in which the second CPU sees the effects 543 of the first CPU's accesses occur, but se 534 of the first CPU's accesses occur, but see the next point: 544 535 545 (*) There is no guarantee that a CPU will see 536 (*) There is no guarantee that a CPU will see the correct order of effects 546 from a second CPU's accesses, even _if_ t 537 from a second CPU's accesses, even _if_ the second CPU uses a memory 547 barrier, unless the first CPU _also_ uses 538 barrier, unless the first CPU _also_ uses a matching memory barrier (see 548 the subsection on "SMP Barrier Pairing"). 539 the subsection on "SMP Barrier Pairing"). 549 540 550 (*) There is no guarantee that some interveni 541 (*) There is no guarantee that some intervening piece of off-the-CPU 551 hardware[*] will not reorder the memory a 542 hardware[*] will not reorder the memory accesses. CPU cache coherency 552 mechanisms should propagate the indirect 543 mechanisms should propagate the indirect effects of a memory barrier 553 between CPUs, but might not do so in orde 544 between CPUs, but might not do so in order. 554 545 555 [*] For information on bus mastering D 546 [*] For information on bus mastering DMA and coherency please read: 556 547 557 Documentation/driver-api/pci/pci.r 548 Documentation/driver-api/pci/pci.rst 558 Documentation/core-api/dma-api-how 549 Documentation/core-api/dma-api-howto.rst 559 Documentation/core-api/dma-api.rst 550 Documentation/core-api/dma-api.rst 560 551 561 552 562 ADDRESS-DEPENDENCY BARRIERS (HISTORICAL) !! 553 DATA DEPENDENCY BARRIERS (HISTORICAL) 563 ---------------------------------------- !! 554 ------------------------------------- 564 [!] This section is marked as HISTORICAL: it c << 565 smp_read_barrier_depends() macro, the semantic << 566 in all marked accesses. For more up-to-date i << 567 how compiler transformations can sometimes bre << 568 see Documentation/RCU/rcu_dereference.rst. << 569 555 570 As of v4.15 of the Linux kernel, an smp_mb() w 556 As of v4.15 of the Linux kernel, an smp_mb() was added to READ_ONCE() for 571 DEC Alpha, which means that about the only peo 557 DEC Alpha, which means that about the only people who need to pay attention 572 to this section are those working on DEC Alpha 558 to this section are those working on DEC Alpha architecture-specific code 573 and those working on READ_ONCE() itself. For 559 and those working on READ_ONCE() itself. For those who need it, and for 574 those who are interested in the history, here 560 those who are interested in the history, here is the story of 575 address-dependency barriers. !! 561 data-dependency barriers. 576 562 577 [!] While address dependencies are observed in !! 563 The usage requirements of data dependency barriers are a little subtle, and 578 load-to-store relations, address-dependency ba << 579 for load-to-store situations. << 580 << 581 The requirement of address-dependency barriers << 582 it's not always obvious that they're needed. 564 it's not always obvious that they're needed. To illustrate, consider the 583 following sequence of events: 565 following sequence of events: 584 566 585 CPU 1 CPU 2 567 CPU 1 CPU 2 586 =============== =============== 568 =============== =============== 587 { A == 1, B == 2, C == 3, P == &A, Q = 569 { A == 1, B == 2, C == 3, P == &A, Q == &C } 588 B = 4; 570 B = 4; 589 <write barrier> 571 <write barrier> 590 WRITE_ONCE(P, &B); 572 WRITE_ONCE(P, &B); 591 Q = READ_ONCE_OL !! 573 Q = READ_ONCE(P); 592 D = *Q; 574 D = *Q; 593 575 594 [!] READ_ONCE_OLD() corresponds to READ_ONCE() !! 576 There's a clear data dependency here, and it would seem that by the end of the 595 doesn't imply an address-dependency barrier. !! 577 sequence, Q must be either &A or &B, and that: 596 << 597 There's a clear address dependency here, and i << 598 the sequence, Q must be either &A or &B, and t << 599 578 600 (Q == &A) implies (D == 1) 579 (Q == &A) implies (D == 1) 601 (Q == &B) implies (D == 4) 580 (Q == &B) implies (D == 4) 602 581 603 But! CPU 2's perception of P may be updated _ 582 But! CPU 2's perception of P may be updated _before_ its perception of B, thus 604 leading to the following situation: 583 leading to the following situation: 605 584 606 (Q == &B) and (D == 2) ???? 585 (Q == &B) and (D == 2) ???? 607 586 608 While this may seem like a failure of coherenc 587 While this may seem like a failure of coherency or causality maintenance, it 609 isn't, and this behaviour can be observed on c 588 isn't, and this behaviour can be observed on certain real CPUs (such as the DEC 610 Alpha). 589 Alpha). 611 590 612 To deal with this, READ_ONCE() provides an imp !! 591 To deal with this, a data dependency barrier or better must be inserted 613 since kernel release v4.15: !! 592 between the address load and the data load: 614 593 615 CPU 1 CPU 2 594 CPU 1 CPU 2 616 =============== =============== 595 =============== =============== 617 { A == 1, B == 2, C == 3, P == &A, Q = 596 { A == 1, B == 2, C == 3, P == &A, Q == &C } 618 B = 4; 597 B = 4; 619 <write barrier> 598 <write barrier> 620 WRITE_ONCE(P, &B); 599 WRITE_ONCE(P, &B); 621 Q = READ_ONCE(P) 600 Q = READ_ONCE(P); 622 <implicit addres !! 601 <data dependency barrier> 623 D = *Q; 602 D = *Q; 624 603 625 This enforces the occurrence of one of the two 604 This enforces the occurrence of one of the two implications, and prevents the 626 third possibility from arising. 605 third possibility from arising. 627 606 628 607 629 [!] Note that this extremely counterintuitive 608 [!] Note that this extremely counterintuitive situation arises most easily on 630 machines with split caches, so that, for examp 609 machines with split caches, so that, for example, one cache bank processes 631 even-numbered cache lines and the other bank p 610 even-numbered cache lines and the other bank processes odd-numbered cache 632 lines. The pointer P might be stored in an od 611 lines. The pointer P might be stored in an odd-numbered cache line, and the 633 variable B might be stored in an even-numbered 612 variable B might be stored in an even-numbered cache line. Then, if the 634 even-numbered bank of the reading CPU's cache 613 even-numbered bank of the reading CPU's cache is extremely busy while the 635 odd-numbered bank is idle, one can see the new 614 odd-numbered bank is idle, one can see the new value of the pointer P (&B), 636 but the old value of the variable B (2). 615 but the old value of the variable B (2). 637 616 638 617 639 An address-dependency barrier is not required !! 618 A data-dependency barrier is not required to order dependent writes 640 because the CPUs that the Linux kernel support !! 619 because the CPUs that the Linux kernel supports don't do writes 641 are certain (1) that the write will actually h !! 620 until they are certain (1) that the write will actually happen, (2) 642 the write, and (3) of the value to be written. !! 621 of the location of the write, and (3) of the value to be written. 643 But please carefully read the "CONTROL DEPENDE 622 But please carefully read the "CONTROL DEPENDENCIES" section and the 644 Documentation/RCU/rcu_dereference.rst file: T !! 623 Documentation/RCU/rcu_dereference.rst file: The compiler can and does 645 dependencies in a great many highly creative w !! 624 break dependencies in a great many highly creative ways. 646 625 647 CPU 1 CPU 2 626 CPU 1 CPU 2 648 =============== =============== 627 =============== =============== 649 { A == 1, B == 2, C = 3, P == &A, Q == 628 { A == 1, B == 2, C = 3, P == &A, Q == &C } 650 B = 4; 629 B = 4; 651 <write barrier> 630 <write barrier> 652 WRITE_ONCE(P, &B); 631 WRITE_ONCE(P, &B); 653 Q = READ_ONCE_OL !! 632 Q = READ_ONCE(P); 654 WRITE_ONCE(*Q, 5 633 WRITE_ONCE(*Q, 5); 655 634 656 Therefore, no address-dependency barrier is re !! 635 Therefore, no data-dependency barrier is required to order the read into 657 Q with the store into *Q. In other words, thi 636 Q with the store into *Q. In other words, this outcome is prohibited, 658 even without an implicit address-dependency ba !! 637 even without a data-dependency barrier: 659 638 660 (Q == &B) && (B == 4) 639 (Q == &B) && (B == 4) 661 640 662 Please note that this pattern should be rare. 641 Please note that this pattern should be rare. After all, the whole point 663 of dependency ordering is to -prevent- writes 642 of dependency ordering is to -prevent- writes to the data structure, along 664 with the expensive cache misses associated wit 643 with the expensive cache misses associated with those writes. This pattern 665 can be used to record rare error conditions an 644 can be used to record rare error conditions and the like, and the CPUs' 666 naturally occurring ordering prevents such rec 645 naturally occurring ordering prevents such records from being lost. 667 646 668 647 669 Note well that the ordering provided by an add !! 648 Note well that the ordering provided by a data dependency is local to 670 the CPU containing it. See the section on "Mu 649 the CPU containing it. See the section on "Multicopy atomicity" for 671 more information. 650 more information. 672 651 673 652 674 The address-dependency barrier is very importa !! 653 The data dependency barrier is very important to the RCU system, 675 for example. See rcu_assign_pointer() and rcu 654 for example. See rcu_assign_pointer() and rcu_dereference() in 676 include/linux/rcupdate.h. This permits the cu 655 include/linux/rcupdate.h. This permits the current target of an RCU'd 677 pointer to be replaced with a new modified tar 656 pointer to be replaced with a new modified target, without the replacement 678 target appearing to be incompletely initialise 657 target appearing to be incompletely initialised. 679 658 680 See also the subsection on "Cache Coherency" f 659 See also the subsection on "Cache Coherency" for a more thorough example. 681 660 682 661 683 CONTROL DEPENDENCIES 662 CONTROL DEPENDENCIES 684 -------------------- 663 -------------------- 685 664 686 Control dependencies can be a bit tricky becau 665 Control dependencies can be a bit tricky because current compilers do 687 not understand them. The purpose of this sect 666 not understand them. The purpose of this section is to help you prevent 688 the compiler's ignorance from breaking your co 667 the compiler's ignorance from breaking your code. 689 668 690 A load-load control dependency requires a full 669 A load-load control dependency requires a full read memory barrier, not 691 simply an (implicit) address-dependency barrie !! 670 simply a data dependency barrier to make it work correctly. Consider the 692 Consider the following bit of code: !! 671 following bit of code: 693 672 694 q = READ_ONCE(a); 673 q = READ_ONCE(a); 695 <implicit address-dependency barrier> << 696 if (q) { 674 if (q) { 697 /* BUG: No address dependency! !! 675 <data dependency barrier> /* BUG: No data dependency!!! */ 698 p = READ_ONCE(b); 676 p = READ_ONCE(b); 699 } 677 } 700 678 701 This will not have the desired effect because !! 679 This will not have the desired effect because there is no actual data 702 dependency, but rather a control dependency th 680 dependency, but rather a control dependency that the CPU may short-circuit 703 by attempting to predict the outcome in advanc 681 by attempting to predict the outcome in advance, so that other CPUs see 704 the load from b as having happened before the !! 682 the load from b as having happened before the load from a. In such a 705 what's actually required is: !! 683 case what's actually required is: 706 684 707 q = READ_ONCE(a); 685 q = READ_ONCE(a); 708 if (q) { 686 if (q) { 709 <read barrier> 687 <read barrier> 710 p = READ_ONCE(b); 688 p = READ_ONCE(b); 711 } 689 } 712 690 713 However, stores are not speculated. This mean 691 However, stores are not speculated. This means that ordering -is- provided 714 for load-store control dependencies, as in the 692 for load-store control dependencies, as in the following example: 715 693 716 q = READ_ONCE(a); 694 q = READ_ONCE(a); 717 if (q) { 695 if (q) { 718 WRITE_ONCE(b, 1); 696 WRITE_ONCE(b, 1); 719 } 697 } 720 698 721 Control dependencies pair normally with other 699 Control dependencies pair normally with other types of barriers. 722 That said, please note that neither READ_ONCE( 700 That said, please note that neither READ_ONCE() nor WRITE_ONCE() 723 are optional! Without the READ_ONCE(), the com 701 are optional! Without the READ_ONCE(), the compiler might combine the 724 load from 'a' with other loads from 'a'. With 702 load from 'a' with other loads from 'a'. Without the WRITE_ONCE(), 725 the compiler might combine the store to 'b' wi 703 the compiler might combine the store to 'b' with other stores to 'b'. 726 Either can result in highly counterintuitive e 704 Either can result in highly counterintuitive effects on ordering. 727 705 728 Worse yet, if the compiler is able to prove (s 706 Worse yet, if the compiler is able to prove (say) that the value of 729 variable 'a' is always non-zero, it would be w 707 variable 'a' is always non-zero, it would be well within its rights 730 to optimize the original example by eliminatin 708 to optimize the original example by eliminating the "if" statement 731 as follows: 709 as follows: 732 710 733 q = a; 711 q = a; 734 b = 1; /* BUG: Compiler and CPU can b 712 b = 1; /* BUG: Compiler and CPU can both reorder!!! */ 735 713 736 So don't leave out the READ_ONCE(). 714 So don't leave out the READ_ONCE(). 737 715 738 It is tempting to try to enforce ordering on i 716 It is tempting to try to enforce ordering on identical stores on both 739 branches of the "if" statement as follows: 717 branches of the "if" statement as follows: 740 718 741 q = READ_ONCE(a); 719 q = READ_ONCE(a); 742 if (q) { 720 if (q) { 743 barrier(); 721 barrier(); 744 WRITE_ONCE(b, 1); 722 WRITE_ONCE(b, 1); 745 do_something(); 723 do_something(); 746 } else { 724 } else { 747 barrier(); 725 barrier(); 748 WRITE_ONCE(b, 1); 726 WRITE_ONCE(b, 1); 749 do_something_else(); 727 do_something_else(); 750 } 728 } 751 729 752 Unfortunately, current compilers will transfor 730 Unfortunately, current compilers will transform this as follows at high 753 optimization levels: 731 optimization levels: 754 732 755 q = READ_ONCE(a); 733 q = READ_ONCE(a); 756 barrier(); 734 barrier(); 757 WRITE_ONCE(b, 1); /* BUG: No ordering 735 WRITE_ONCE(b, 1); /* BUG: No ordering vs. load from a!!! */ 758 if (q) { 736 if (q) { 759 /* WRITE_ONCE(b, 1); -- moved 737 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */ 760 do_something(); 738 do_something(); 761 } else { 739 } else { 762 /* WRITE_ONCE(b, 1); -- moved 740 /* WRITE_ONCE(b, 1); -- moved up, BUG!!! */ 763 do_something_else(); 741 do_something_else(); 764 } 742 } 765 743 766 Now there is no conditional between the load f 744 Now there is no conditional between the load from 'a' and the store to 767 'b', which means that the CPU is within its ri 745 'b', which means that the CPU is within its rights to reorder them: 768 The conditional is absolutely required, and mu 746 The conditional is absolutely required, and must be present in the 769 assembly code even after all compiler optimiza 747 assembly code even after all compiler optimizations have been applied. 770 Therefore, if you need ordering in this exampl 748 Therefore, if you need ordering in this example, you need explicit 771 memory barriers, for example, smp_store_releas 749 memory barriers, for example, smp_store_release(): 772 750 773 q = READ_ONCE(a); 751 q = READ_ONCE(a); 774 if (q) { 752 if (q) { 775 smp_store_release(&b, 1); 753 smp_store_release(&b, 1); 776 do_something(); 754 do_something(); 777 } else { 755 } else { 778 smp_store_release(&b, 1); 756 smp_store_release(&b, 1); 779 do_something_else(); 757 do_something_else(); 780 } 758 } 781 759 782 In contrast, without explicit memory barriers, 760 In contrast, without explicit memory barriers, two-legged-if control 783 ordering is guaranteed only when the stores di 761 ordering is guaranteed only when the stores differ, for example: 784 762 785 q = READ_ONCE(a); 763 q = READ_ONCE(a); 786 if (q) { 764 if (q) { 787 WRITE_ONCE(b, 1); 765 WRITE_ONCE(b, 1); 788 do_something(); 766 do_something(); 789 } else { 767 } else { 790 WRITE_ONCE(b, 2); 768 WRITE_ONCE(b, 2); 791 do_something_else(); 769 do_something_else(); 792 } 770 } 793 771 794 The initial READ_ONCE() is still required to p 772 The initial READ_ONCE() is still required to prevent the compiler from 795 proving the value of 'a'. 773 proving the value of 'a'. 796 774 797 In addition, you need to be careful what you d 775 In addition, you need to be careful what you do with the local variable 'q', 798 otherwise the compiler might be able to guess 776 otherwise the compiler might be able to guess the value and again remove 799 the needed conditional. For example: 777 the needed conditional. For example: 800 778 801 q = READ_ONCE(a); 779 q = READ_ONCE(a); 802 if (q % MAX) { 780 if (q % MAX) { 803 WRITE_ONCE(b, 1); 781 WRITE_ONCE(b, 1); 804 do_something(); 782 do_something(); 805 } else { 783 } else { 806 WRITE_ONCE(b, 2); 784 WRITE_ONCE(b, 2); 807 do_something_else(); 785 do_something_else(); 808 } 786 } 809 787 810 If MAX is defined to be 1, then the compiler k 788 If MAX is defined to be 1, then the compiler knows that (q % MAX) is 811 equal to zero, in which case the compiler is w 789 equal to zero, in which case the compiler is within its rights to 812 transform the above code into the following: 790 transform the above code into the following: 813 791 814 q = READ_ONCE(a); 792 q = READ_ONCE(a); 815 WRITE_ONCE(b, 2); 793 WRITE_ONCE(b, 2); 816 do_something_else(); 794 do_something_else(); 817 795 818 Given this transformation, the CPU is not requ 796 Given this transformation, the CPU is not required to respect the ordering 819 between the load from variable 'a' and the sto 797 between the load from variable 'a' and the store to variable 'b'. It is 820 tempting to add a barrier(), but this does not 798 tempting to add a barrier(), but this does not help. The conditional 821 is gone, and the barrier won't bring it back. 799 is gone, and the barrier won't bring it back. Therefore, if you are 822 relying on this ordering, you should make sure 800 relying on this ordering, you should make sure that MAX is greater than 823 one, perhaps as follows: 801 one, perhaps as follows: 824 802 825 q = READ_ONCE(a); 803 q = READ_ONCE(a); 826 BUILD_BUG_ON(MAX <= 1); /* Order load 804 BUILD_BUG_ON(MAX <= 1); /* Order load from a with store to b. */ 827 if (q % MAX) { 805 if (q % MAX) { 828 WRITE_ONCE(b, 1); 806 WRITE_ONCE(b, 1); 829 do_something(); 807 do_something(); 830 } else { 808 } else { 831 WRITE_ONCE(b, 2); 809 WRITE_ONCE(b, 2); 832 do_something_else(); 810 do_something_else(); 833 } 811 } 834 812 835 Please note once again that the stores to 'b' 813 Please note once again that the stores to 'b' differ. If they were 836 identical, as noted earlier, the compiler coul 814 identical, as noted earlier, the compiler could pull this store outside 837 of the 'if' statement. 815 of the 'if' statement. 838 816 839 You must also be careful not to rely too much 817 You must also be careful not to rely too much on boolean short-circuit 840 evaluation. Consider this example: 818 evaluation. Consider this example: 841 819 842 q = READ_ONCE(a); 820 q = READ_ONCE(a); 843 if (q || 1 > 0) 821 if (q || 1 > 0) 844 WRITE_ONCE(b, 1); 822 WRITE_ONCE(b, 1); 845 823 846 Because the first condition cannot fault and t 824 Because the first condition cannot fault and the second condition is 847 always true, the compiler can transform this e 825 always true, the compiler can transform this example as following, 848 defeating control dependency: 826 defeating control dependency: 849 827 850 q = READ_ONCE(a); 828 q = READ_ONCE(a); 851 WRITE_ONCE(b, 1); 829 WRITE_ONCE(b, 1); 852 830 853 This example underscores the need to ensure th 831 This example underscores the need to ensure that the compiler cannot 854 out-guess your code. More generally, although 832 out-guess your code. More generally, although READ_ONCE() does force 855 the compiler to actually emit code for a given 833 the compiler to actually emit code for a given load, it does not force 856 the compiler to use the results. 834 the compiler to use the results. 857 835 858 In addition, control dependencies apply only t 836 In addition, control dependencies apply only to the then-clause and 859 else-clause of the if-statement in question. 837 else-clause of the if-statement in question. In particular, it does 860 not necessarily apply to code following the if 838 not necessarily apply to code following the if-statement: 861 839 862 q = READ_ONCE(a); 840 q = READ_ONCE(a); 863 if (q) { 841 if (q) { 864 WRITE_ONCE(b, 1); 842 WRITE_ONCE(b, 1); 865 } else { 843 } else { 866 WRITE_ONCE(b, 2); 844 WRITE_ONCE(b, 2); 867 } 845 } 868 WRITE_ONCE(c, 1); /* BUG: No ordering 846 WRITE_ONCE(c, 1); /* BUG: No ordering against the read from 'a'. */ 869 847 870 It is tempting to argue that there in fact is 848 It is tempting to argue that there in fact is ordering because the 871 compiler cannot reorder volatile accesses and 849 compiler cannot reorder volatile accesses and also cannot reorder 872 the writes to 'b' with the condition. Unfortu 850 the writes to 'b' with the condition. Unfortunately for this line 873 of reasoning, the compiler might compile the t 851 of reasoning, the compiler might compile the two writes to 'b' as 874 conditional-move instructions, as in this fanc 852 conditional-move instructions, as in this fanciful pseudo-assembly 875 language: 853 language: 876 854 877 ld r1,a 855 ld r1,a 878 cmp r1,$0 856 cmp r1,$0 879 cmov,ne r4,$1 857 cmov,ne r4,$1 880 cmov,eq r4,$2 858 cmov,eq r4,$2 881 st r4,b 859 st r4,b 882 st $1,c 860 st $1,c 883 861 884 A weakly ordered CPU would have no dependency 862 A weakly ordered CPU would have no dependency of any sort between the load 885 from 'a' and the store to 'c'. The control de 863 from 'a' and the store to 'c'. The control dependencies would extend 886 only to the pair of cmov instructions and the 864 only to the pair of cmov instructions and the store depending on them. 887 In short, control dependencies apply only to t 865 In short, control dependencies apply only to the stores in the then-clause 888 and else-clause of the if-statement in questio 866 and else-clause of the if-statement in question (including functions 889 invoked by those two clauses), not to code fol 867 invoked by those two clauses), not to code following that if-statement. 890 868 891 869 892 Note well that the ordering provided by a cont 870 Note well that the ordering provided by a control dependency is local 893 to the CPU containing it. See the section on 871 to the CPU containing it. See the section on "Multicopy atomicity" 894 for more information. 872 for more information. 895 873 896 874 897 In summary: 875 In summary: 898 876 899 (*) Control dependencies can order prior loa 877 (*) Control dependencies can order prior loads against later stores. 900 However, they do -not- guarantee any oth 878 However, they do -not- guarantee any other sort of ordering: 901 Not prior loads against later loads, nor 879 Not prior loads against later loads, nor prior stores against 902 later anything. If you need these other 880 later anything. If you need these other forms of ordering, 903 use smp_rmb(), smp_wmb(), or, in the cas 881 use smp_rmb(), smp_wmb(), or, in the case of prior stores and 904 later loads, smp_mb(). 882 later loads, smp_mb(). 905 883 906 (*) If both legs of the "if" statement begin 884 (*) If both legs of the "if" statement begin with identical stores to 907 the same variable, then those stores mus 885 the same variable, then those stores must be ordered, either by 908 preceding both of them with smp_mb() or 886 preceding both of them with smp_mb() or by using smp_store_release() 909 to carry out the stores. Please note th 887 to carry out the stores. Please note that it is -not- sufficient 910 to use barrier() at beginning of each le 888 to use barrier() at beginning of each leg of the "if" statement 911 because, as shown by the example above, 889 because, as shown by the example above, optimizing compilers can 912 destroy the control dependency while res 890 destroy the control dependency while respecting the letter of the 913 barrier() law. 891 barrier() law. 914 892 915 (*) Control dependencies require at least on 893 (*) Control dependencies require at least one run-time conditional 916 between the prior load and the subsequen 894 between the prior load and the subsequent store, and this 917 conditional must involve the prior load. 895 conditional must involve the prior load. If the compiler is able 918 to optimize the conditional away, it wil 896 to optimize the conditional away, it will have also optimized 919 away the ordering. Careful use of READ_ 897 away the ordering. Careful use of READ_ONCE() and WRITE_ONCE() 920 can help to preserve the needed conditio 898 can help to preserve the needed conditional. 921 899 922 (*) Control dependencies require that the co 900 (*) Control dependencies require that the compiler avoid reordering the 923 dependency into nonexistence. Careful u 901 dependency into nonexistence. Careful use of READ_ONCE() or 924 atomic{,64}_read() can help to preserve 902 atomic{,64}_read() can help to preserve your control dependency. 925 Please see the COMPILER BARRIER section 903 Please see the COMPILER BARRIER section for more information. 926 904 927 (*) Control dependencies apply only to the t 905 (*) Control dependencies apply only to the then-clause and else-clause 928 of the if-statement containing the contr 906 of the if-statement containing the control dependency, including 929 any functions that these two clauses cal 907 any functions that these two clauses call. Control dependencies 930 do -not- apply to code following the if- 908 do -not- apply to code following the if-statement containing the 931 control dependency. 909 control dependency. 932 910 933 (*) Control dependencies pair normally with 911 (*) Control dependencies pair normally with other types of barriers. 934 912 935 (*) Control dependencies do -not- provide mu 913 (*) Control dependencies do -not- provide multicopy atomicity. If you 936 need all the CPUs to see a given store a 914 need all the CPUs to see a given store at the same time, use smp_mb(). 937 915 938 (*) Compilers do not understand control depe 916 (*) Compilers do not understand control dependencies. It is therefore 939 your job to ensure that they do not brea 917 your job to ensure that they do not break your code. 940 918 941 919 942 SMP BARRIER PAIRING 920 SMP BARRIER PAIRING 943 ------------------- 921 ------------------- 944 922 945 When dealing with CPU-CPU interactions, certai 923 When dealing with CPU-CPU interactions, certain types of memory barrier should 946 always be paired. A lack of appropriate pairi 924 always be paired. A lack of appropriate pairing is almost certainly an error. 947 925 948 General barriers pair with each other, though 926 General barriers pair with each other, though they also pair with most 949 other types of barriers, albeit without multic 927 other types of barriers, albeit without multicopy atomicity. An acquire 950 barrier pairs with a release barrier, but both 928 barrier pairs with a release barrier, but both may also pair with other 951 barriers, including of course general barriers 929 barriers, including of course general barriers. A write barrier pairs 952 with an address-dependency barrier, a control !! 930 with a data dependency barrier, a control dependency, an acquire barrier, 953 a release barrier, a read barrier, or a genera 931 a release barrier, a read barrier, or a general barrier. Similarly a 954 read barrier, control dependency, or an addres !! 932 read barrier, control dependency, or a data dependency barrier pairs 955 with a write barrier, an acquire barrier, a re 933 with a write barrier, an acquire barrier, a release barrier, or a 956 general barrier: 934 general barrier: 957 935 958 CPU 1 CPU 2 936 CPU 1 CPU 2 959 =============== =============== 937 =============== =============== 960 WRITE_ONCE(a, 1); 938 WRITE_ONCE(a, 1); 961 <write barrier> 939 <write barrier> 962 WRITE_ONCE(b, 2); x = READ_ONCE(b) 940 WRITE_ONCE(b, 2); x = READ_ONCE(b); 963 <read barrier> 941 <read barrier> 964 y = READ_ONCE(a) 942 y = READ_ONCE(a); 965 943 966 Or: 944 Or: 967 945 968 CPU 1 CPU 2 946 CPU 1 CPU 2 969 =============== ================ 947 =============== =============================== 970 a = 1; 948 a = 1; 971 <write barrier> 949 <write barrier> 972 WRITE_ONCE(b, &a); x = READ_ONCE(b) 950 WRITE_ONCE(b, &a); x = READ_ONCE(b); 973 <implicit addres !! 951 <data dependency barrier> 974 y = *x; 952 y = *x; 975 953 976 Or even: 954 Or even: 977 955 978 CPU 1 CPU 2 956 CPU 1 CPU 2 979 =============== ================ 957 =============== =============================== 980 r1 = READ_ONCE(y); 958 r1 = READ_ONCE(y); 981 <general barrier> 959 <general barrier> 982 WRITE_ONCE(x, 1); if (r2 = READ_ON 960 WRITE_ONCE(x, 1); if (r2 = READ_ONCE(x)) { 983 <implicit con 961 <implicit control dependency> 984 WRITE_ONCE(y, 962 WRITE_ONCE(y, 1); 985 } 963 } 986 964 987 assert(r1 == 0 || r2 == 0); 965 assert(r1 == 0 || r2 == 0); 988 966 989 Basically, the read barrier always has to be t 967 Basically, the read barrier always has to be there, even though it can be of 990 the "weaker" type. 968 the "weaker" type. 991 969 992 [!] Note that the stores before the write barr 970 [!] Note that the stores before the write barrier would normally be expected to 993 match the loads after the read barrier or the !! 971 match the loads after the read barrier or the data dependency barrier, and vice 994 vice versa: !! 972 versa: 995 973 996 CPU 1 CP 974 CPU 1 CPU 2 997 =================== == 975 =================== =================== 998 WRITE_ONCE(a, 1); }---- --->{ v 976 WRITE_ONCE(a, 1); }---- --->{ v = READ_ONCE(c); 999 WRITE_ONCE(b, 2); } \ / { w 977 WRITE_ONCE(b, 2); } \ / { w = READ_ONCE(d); 1000 <write barrier> \ < 978 <write barrier> \ <read barrier> 1001 WRITE_ONCE(c, 3); } / \ { x 979 WRITE_ONCE(c, 3); } / \ { x = READ_ONCE(a); 1002 WRITE_ONCE(d, 4); }---- --->{ y 980 WRITE_ONCE(d, 4); }---- --->{ y = READ_ONCE(b); 1003 981 1004 982 1005 EXAMPLES OF MEMORY BARRIER SEQUENCES 983 EXAMPLES OF MEMORY BARRIER SEQUENCES 1006 ------------------------------------ 984 ------------------------------------ 1007 985 1008 Firstly, write barriers act as partial orderi 986 Firstly, write barriers act as partial orderings on store operations. 1009 Consider the following sequence of events: 987 Consider the following sequence of events: 1010 988 1011 CPU 1 989 CPU 1 1012 ======================= 990 ======================= 1013 STORE A = 1 991 STORE A = 1 1014 STORE B = 2 992 STORE B = 2 1015 STORE C = 3 993 STORE C = 3 1016 <write barrier> 994 <write barrier> 1017 STORE D = 4 995 STORE D = 4 1018 STORE E = 5 996 STORE E = 5 1019 997 1020 This sequence of events is committed to the m 998 This sequence of events is committed to the memory coherence system in an order 1021 that the rest of the system might perceive as 999 that the rest of the system might perceive as the unordered set of { STORE A, 1022 STORE B, STORE C } all occurring before the u 1000 STORE B, STORE C } all occurring before the unordered set of { STORE D, STORE E 1023 }: 1001 }: 1024 1002 1025 +-------+ : : 1003 +-------+ : : 1026 | | +------+ 1004 | | +------+ 1027 | |------>| C=3 | } /\ 1005 | |------>| C=3 | } /\ 1028 | | : +------+ }----- 1006 | | : +------+ }----- \ -----> Events perceptible to 1029 | | : | A=1 | } 1007 | | : | A=1 | } \/ the rest of the system 1030 | | : +------+ } 1008 | | : +------+ } 1031 | CPU 1 | : | B=2 | } 1009 | CPU 1 | : | B=2 | } 1032 | | +------+ } 1010 | | +------+ } 1033 | | wwwwwwwwwwwwwwww } <--- 1011 | | wwwwwwwwwwwwwwww } <--- At this point the write barrier 1034 | | +------+ } 1012 | | +------+ } requires all stores prior to the 1035 | | : | E=5 | } 1013 | | : | E=5 | } barrier to be committed before 1036 | | : +------+ } 1014 | | : +------+ } further stores may take place 1037 | |------>| D=4 | } 1015 | |------>| D=4 | } 1038 | | +------+ 1016 | | +------+ 1039 +-------+ : : 1017 +-------+ : : 1040 | 1018 | 1041 | Sequence in whic 1019 | Sequence in which stores are committed to the 1042 | memory system by 1020 | memory system by CPU 1 1043 V 1021 V 1044 1022 1045 1023 1046 Secondly, address-dependency barriers act as !! 1024 Secondly, data dependency barriers act as partial orderings on data-dependent 1047 dependent loads. Consider the following sequ !! 1025 loads. Consider the following sequence of events: 1048 1026 1049 CPU 1 CPU 2 1027 CPU 1 CPU 2 1050 ======================= ============= 1028 ======================= ======================= 1051 { B = 7; X = 9; Y = 8; C = &Y 1029 { B = 7; X = 9; Y = 8; C = &Y } 1052 STORE A = 1 1030 STORE A = 1 1053 STORE B = 2 1031 STORE B = 2 1054 <write barrier> 1032 <write barrier> 1055 STORE C = &B LOAD X 1033 STORE C = &B LOAD X 1056 STORE D = 4 LOAD C (gets 1034 STORE D = 4 LOAD C (gets &B) 1057 LOAD *C (read 1035 LOAD *C (reads B) 1058 1036 1059 Without intervention, CPU 2 may perceive the 1037 Without intervention, CPU 2 may perceive the events on CPU 1 in some 1060 effectively random order, despite the write b 1038 effectively random order, despite the write barrier issued by CPU 1: 1061 1039 1062 +-------+ : : 1040 +-------+ : : : : 1063 | | +------+ 1041 | | +------+ +-------+ | Sequence of update 1064 | |------>| B=2 |----- - 1042 | |------>| B=2 |----- --->| Y->8 | | of perception on 1065 | | : +------+ \ 1043 | | : +------+ \ +-------+ | CPU 2 1066 | CPU 1 | : | A=1 | \ - 1044 | CPU 1 | : | A=1 | \ --->| C->&Y | V 1067 | | +------+ | 1045 | | +------+ | +-------+ 1068 | | wwwwwwwwwwwwwwww | 1046 | | wwwwwwwwwwwwwwww | : : 1069 | | +------+ | 1047 | | +------+ | : : 1070 | | : | C=&B |--- | 1048 | | : | C=&B |--- | : : +-------+ 1071 | | : +------+ \ | 1049 | | : +------+ \ | +-------+ | | 1072 | |------>| D=4 | --------- 1050 | |------>| D=4 | ----------->| C->&B |------>| | 1073 | | +------+ | 1051 | | +------+ | +-------+ | | 1074 +-------+ : : | 1052 +-------+ : : | : : | | 1075 | 1053 | : : | | 1076 | 1054 | : : | CPU 2 | 1077 | 1055 | +-------+ | | 1078 Apparently incorrect ---> | 1056 Apparently incorrect ---> | | B->7 |------>| | 1079 perception of B (!) | 1057 perception of B (!) | +-------+ | | 1080 | 1058 | : : | | 1081 | 1059 | +-------+ | | 1082 The load of X holds ---> \ 1060 The load of X holds ---> \ | X->9 |------>| | 1083 up the maintenance \ 1061 up the maintenance \ +-------+ | | 1084 of coherence of B --- 1062 of coherence of B ----->| B->2 | +-------+ 1085 1063 +-------+ 1086 1064 : : 1087 1065 1088 1066 1089 In the above example, CPU 2 perceives that B 1067 In the above example, CPU 2 perceives that B is 7, despite the load of *C 1090 (which would be B) coming after the LOAD of C 1068 (which would be B) coming after the LOAD of C. 1091 1069 1092 If, however, an address-dependency barrier we !! 1070 If, however, a data dependency barrier were to be placed between the load of C 1093 of C and the load of *C (ie: B) on CPU 2: !! 1071 and the load of *C (ie: B) on CPU 2: 1094 1072 1095 CPU 1 CPU 2 1073 CPU 1 CPU 2 1096 ======================= ============= 1074 ======================= ======================= 1097 { B = 7; X = 9; Y = 8; C = &Y 1075 { B = 7; X = 9; Y = 8; C = &Y } 1098 STORE A = 1 1076 STORE A = 1 1099 STORE B = 2 1077 STORE B = 2 1100 <write barrier> 1078 <write barrier> 1101 STORE C = &B LOAD X 1079 STORE C = &B LOAD X 1102 STORE D = 4 LOAD C (gets 1080 STORE D = 4 LOAD C (gets &B) 1103 <address-depe !! 1081 <data dependency barrier> 1104 LOAD *C (read 1082 LOAD *C (reads B) 1105 1083 1106 then the following will occur: 1084 then the following will occur: 1107 1085 1108 +-------+ : : 1086 +-------+ : : : : 1109 | | +------+ 1087 | | +------+ +-------+ 1110 | |------>| B=2 |----- - 1088 | |------>| B=2 |----- --->| Y->8 | 1111 | | : +------+ \ 1089 | | : +------+ \ +-------+ 1112 | CPU 1 | : | A=1 | \ - 1090 | CPU 1 | : | A=1 | \ --->| C->&Y | 1113 | | +------+ | 1091 | | +------+ | +-------+ 1114 | | wwwwwwwwwwwwwwww | 1092 | | wwwwwwwwwwwwwwww | : : 1115 | | +------+ | 1093 | | +------+ | : : 1116 | | : | C=&B |--- | 1094 | | : | C=&B |--- | : : +-------+ 1117 | | : +------+ \ | 1095 | | : +------+ \ | +-------+ | | 1118 | |------>| D=4 | --------- 1096 | |------>| D=4 | ----------->| C->&B |------>| | 1119 | | +------+ | 1097 | | +------+ | +-------+ | | 1120 +-------+ : : | 1098 +-------+ : : | : : | | 1121 | 1099 | : : | | 1122 | 1100 | : : | CPU 2 | 1123 | 1101 | +-------+ | | 1124 | 1102 | | X->9 |------>| | 1125 | 1103 | +-------+ | | 1126 Makes sure all effects ---> \ a !! 1104 Makes sure all effects ---> \ ddddddddddddddddd | | 1127 prior to the store of C \ 1105 prior to the store of C \ +-------+ | | 1128 are perceptible to --- 1106 are perceptible to ----->| B->2 |------>| | 1129 subsequent loads 1107 subsequent loads +-------+ | | 1130 1108 : : +-------+ 1131 1109 1132 1110 1133 And thirdly, a read barrier acts as a partial 1111 And thirdly, a read barrier acts as a partial order on loads. Consider the 1134 following sequence of events: 1112 following sequence of events: 1135 1113 1136 CPU 1 CPU 2 1114 CPU 1 CPU 2 1137 ======================= ============= 1115 ======================= ======================= 1138 { A = 0, B = 9 } 1116 { A = 0, B = 9 } 1139 STORE A=1 1117 STORE A=1 1140 <write barrier> 1118 <write barrier> 1141 STORE B=2 1119 STORE B=2 1142 LOAD B 1120 LOAD B 1143 LOAD A 1121 LOAD A 1144 1122 1145 Without intervention, CPU 2 may then choose t 1123 Without intervention, CPU 2 may then choose to perceive the events on CPU 1 in 1146 some effectively random order, despite the wr 1124 some effectively random order, despite the write barrier issued by CPU 1: 1147 1125 1148 +-------+ : : 1126 +-------+ : : : : 1149 | | +------+ 1127 | | +------+ +-------+ 1150 | |------>| A=1 |------ - 1128 | |------>| A=1 |------ --->| A->0 | 1151 | | +------+ \ 1129 | | +------+ \ +-------+ 1152 | CPU 1 | wwwwwwwwwwwwwwww \ - 1130 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1153 | | +------+ | 1131 | | +------+ | +-------+ 1154 | |------>| B=2 |--- | 1132 | |------>| B=2 |--- | : : 1155 | | +------+ \ | 1133 | | +------+ \ | : : +-------+ 1156 +-------+ : : \ | 1134 +-------+ : : \ | +-------+ | | 1157 -------- 1135 ---------->| B->2 |------>| | 1158 | 1136 | +-------+ | CPU 2 | 1159 | 1137 | | A->0 |------>| | 1160 | 1138 | +-------+ | | 1161 | 1139 | : : +-------+ 1162 \ 1140 \ : : 1163 \ 1141 \ +-------+ 1164 -- 1142 ---->| A->1 | 1165 1143 +-------+ 1166 1144 : : 1167 1145 1168 1146 1169 If, however, a read barrier were to be placed 1147 If, however, a read barrier were to be placed between the load of B and the 1170 load of A on CPU 2: 1148 load of A on CPU 2: 1171 1149 1172 CPU 1 CPU 2 1150 CPU 1 CPU 2 1173 ======================= ============= 1151 ======================= ======================= 1174 { A = 0, B = 9 } 1152 { A = 0, B = 9 } 1175 STORE A=1 1153 STORE A=1 1176 <write barrier> 1154 <write barrier> 1177 STORE B=2 1155 STORE B=2 1178 LOAD B 1156 LOAD B 1179 <read barrier 1157 <read barrier> 1180 LOAD A 1158 LOAD A 1181 1159 1182 then the partial ordering imposed by CPU 1 wi 1160 then the partial ordering imposed by CPU 1 will be perceived correctly by CPU 1183 2: 1161 2: 1184 1162 1185 +-------+ : : 1163 +-------+ : : : : 1186 | | +------+ 1164 | | +------+ +-------+ 1187 | |------>| A=1 |------ - 1165 | |------>| A=1 |------ --->| A->0 | 1188 | | +------+ \ 1166 | | +------+ \ +-------+ 1189 | CPU 1 | wwwwwwwwwwwwwwww \ - 1167 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1190 | | +------+ | 1168 | | +------+ | +-------+ 1191 | |------>| B=2 |--- | 1169 | |------>| B=2 |--- | : : 1192 | | +------+ \ | 1170 | | +------+ \ | : : +-------+ 1193 +-------+ : : \ | 1171 +-------+ : : \ | +-------+ | | 1194 -------- 1172 ---------->| B->2 |------>| | 1195 | 1173 | +-------+ | CPU 2 | 1196 | 1174 | : : | | 1197 | 1175 | : : | | 1198 At this point the read ----> \ r 1176 At this point the read ----> \ rrrrrrrrrrrrrrrrr | | 1199 barrier causes all effects \ 1177 barrier causes all effects \ +-------+ | | 1200 prior to the storage of B -- 1178 prior to the storage of B ---->| A->1 |------>| | 1201 to be perceptible to CPU 2 1179 to be perceptible to CPU 2 +-------+ | | 1202 1180 : : +-------+ 1203 1181 1204 1182 1205 To illustrate this more completely, consider 1183 To illustrate this more completely, consider what could happen if the code 1206 contained a load of A either side of the read 1184 contained a load of A either side of the read barrier: 1207 1185 1208 CPU 1 CPU 2 1186 CPU 1 CPU 2 1209 ======================= ============= 1187 ======================= ======================= 1210 { A = 0, B = 9 } 1188 { A = 0, B = 9 } 1211 STORE A=1 1189 STORE A=1 1212 <write barrier> 1190 <write barrier> 1213 STORE B=2 1191 STORE B=2 1214 LOAD B 1192 LOAD B 1215 LOAD A [first 1193 LOAD A [first load of A] 1216 <read barrier 1194 <read barrier> 1217 LOAD A [secon 1195 LOAD A [second load of A] 1218 1196 1219 Even though the two loads of A both occur aft 1197 Even though the two loads of A both occur after the load of B, they may both 1220 come up with different values: 1198 come up with different values: 1221 1199 1222 +-------+ : : 1200 +-------+ : : : : 1223 | | +------+ 1201 | | +------+ +-------+ 1224 | |------>| A=1 |------ - 1202 | |------>| A=1 |------ --->| A->0 | 1225 | | +------+ \ 1203 | | +------+ \ +-------+ 1226 | CPU 1 | wwwwwwwwwwwwwwww \ - 1204 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1227 | | +------+ | 1205 | | +------+ | +-------+ 1228 | |------>| B=2 |--- | 1206 | |------>| B=2 |--- | : : 1229 | | +------+ \ | 1207 | | +------+ \ | : : +-------+ 1230 +-------+ : : \ | 1208 +-------+ : : \ | +-------+ | | 1231 -------- 1209 ---------->| B->2 |------>| | 1232 | 1210 | +-------+ | CPU 2 | 1233 | 1211 | : : | | 1234 | 1212 | : : | | 1235 | 1213 | +-------+ | | 1236 | 1214 | | A->0 |------>| 1st | 1237 | 1215 | +-------+ | | 1238 At this point the read ----> \ r 1216 At this point the read ----> \ rrrrrrrrrrrrrrrrr | | 1239 barrier causes all effects \ 1217 barrier causes all effects \ +-------+ | | 1240 prior to the storage of B -- 1218 prior to the storage of B ---->| A->1 |------>| 2nd | 1241 to be perceptible to CPU 2 1219 to be perceptible to CPU 2 +-------+ | | 1242 1220 : : +-------+ 1243 1221 1244 1222 1245 But it may be that the update to A from CPU 1 1223 But it may be that the update to A from CPU 1 becomes perceptible to CPU 2 1246 before the read barrier completes anyway: 1224 before the read barrier completes anyway: 1247 1225 1248 +-------+ : : 1226 +-------+ : : : : 1249 | | +------+ 1227 | | +------+ +-------+ 1250 | |------>| A=1 |------ - 1228 | |------>| A=1 |------ --->| A->0 | 1251 | | +------+ \ 1229 | | +------+ \ +-------+ 1252 | CPU 1 | wwwwwwwwwwwwwwww \ - 1230 | CPU 1 | wwwwwwwwwwwwwwww \ --->| B->9 | 1253 | | +------+ | 1231 | | +------+ | +-------+ 1254 | |------>| B=2 |--- | 1232 | |------>| B=2 |--- | : : 1255 | | +------+ \ | 1233 | | +------+ \ | : : +-------+ 1256 +-------+ : : \ | 1234 +-------+ : : \ | +-------+ | | 1257 -------- 1235 ---------->| B->2 |------>| | 1258 | 1236 | +-------+ | CPU 2 | 1259 | 1237 | : : | | 1260 \ 1238 \ : : | | 1261 \ 1239 \ +-------+ | | 1262 -- 1240 ---->| A->1 |------>| 1st | 1263 1241 +-------+ | | 1264 r 1242 rrrrrrrrrrrrrrrrr | | 1265 1243 +-------+ | | 1266 1244 | A->1 |------>| 2nd | 1267 1245 +-------+ | | 1268 1246 : : +-------+ 1269 1247 1270 1248 1271 The guarantee is that the second load will al 1249 The guarantee is that the second load will always come up with A == 1 if the 1272 load of B came up with B == 2. No such guara 1250 load of B came up with B == 2. No such guarantee exists for the first load of 1273 A; that may come up with either A == 0 or A = 1251 A; that may come up with either A == 0 or A == 1. 1274 1252 1275 1253 1276 READ MEMORY BARRIERS VS LOAD SPECULATION 1254 READ MEMORY BARRIERS VS LOAD SPECULATION 1277 ---------------------------------------- 1255 ---------------------------------------- 1278 1256 1279 Many CPUs speculate with loads: that is they 1257 Many CPUs speculate with loads: that is they see that they will need to load an 1280 item from memory, and they find a time where 1258 item from memory, and they find a time where they're not using the bus for any 1281 other loads, and so do the load in advance - 1259 other loads, and so do the load in advance - even though they haven't actually 1282 got to that point in the instruction executio 1260 got to that point in the instruction execution flow yet. This permits the 1283 actual load instruction to potentially comple 1261 actual load instruction to potentially complete immediately because the CPU 1284 already has the value to hand. 1262 already has the value to hand. 1285 1263 1286 It may turn out that the CPU didn't actually 1264 It may turn out that the CPU didn't actually need the value - perhaps because a 1287 branch circumvented the load - in which case 1265 branch circumvented the load - in which case it can discard the value or just 1288 cache it for later use. 1266 cache it for later use. 1289 1267 1290 Consider: 1268 Consider: 1291 1269 1292 CPU 1 CPU 2 1270 CPU 1 CPU 2 1293 ======================= ============= 1271 ======================= ======================= 1294 LOAD B 1272 LOAD B 1295 DIVIDE 1273 DIVIDE } Divide instructions generally 1296 DIVIDE 1274 DIVIDE } take a long time to perform 1297 LOAD A 1275 LOAD A 1298 1276 1299 Which might appear as this: 1277 Which might appear as this: 1300 1278 1301 1279 : : +-------+ 1302 1280 +-------+ | | 1303 - 1281 --->| B->2 |------>| | 1304 1282 +-------+ | CPU 2 | 1305 1283 : :DIVIDE | | 1306 1284 +-------+ | | 1307 The CPU being busy doing a ---> - 1285 The CPU being busy doing a ---> --->| A->0 |~~~~ | | 1308 division speculates on the 1286 division speculates on the +-------+ ~ | | 1309 LOAD of A 1287 LOAD of A : : ~ | | 1310 1288 : :DIVIDE | | 1311 1289 : : ~ | | 1312 Once the divisions are complete --> 1290 Once the divisions are complete --> : : ~-->| | 1313 the CPU can then perform the 1291 the CPU can then perform the : : | | 1314 LOAD with immediate effect 1292 LOAD with immediate effect : : +-------+ 1315 1293 1316 1294 1317 Placing a read barrier or an address-dependen !! 1295 Placing a read barrier or a data dependency barrier just before the second 1318 load: 1296 load: 1319 1297 1320 CPU 1 CPU 2 1298 CPU 1 CPU 2 1321 ======================= ============= 1299 ======================= ======================= 1322 LOAD B 1300 LOAD B 1323 DIVIDE 1301 DIVIDE 1324 DIVIDE 1302 DIVIDE 1325 <read barrier 1303 <read barrier> 1326 LOAD A 1304 LOAD A 1327 1305 1328 will force any value speculatively obtained t 1306 will force any value speculatively obtained to be reconsidered to an extent 1329 dependent on the type of barrier used. If th 1307 dependent on the type of barrier used. If there was no change made to the 1330 speculated memory location, then the speculat 1308 speculated memory location, then the speculated value will just be used: 1331 1309 1332 1310 : : +-------+ 1333 1311 +-------+ | | 1334 - 1312 --->| B->2 |------>| | 1335 1313 +-------+ | CPU 2 | 1336 1314 : :DIVIDE | | 1337 1315 +-------+ | | 1338 The CPU being busy doing a ---> - 1316 The CPU being busy doing a ---> --->| A->0 |~~~~ | | 1339 division speculates on the 1317 division speculates on the +-------+ ~ | | 1340 LOAD of A 1318 LOAD of A : : ~ | | 1341 1319 : :DIVIDE | | 1342 1320 : : ~ | | 1343 1321 : : ~ | | 1344 r 1322 rrrrrrrrrrrrrrrr~ | | 1345 1323 : : ~ | | 1346 1324 : : ~-->| | 1347 1325 : : | | 1348 1326 : : +-------+ 1349 1327 1350 1328 1351 but if there was an update or an invalidation 1329 but if there was an update or an invalidation from another CPU pending, then 1352 the speculation will be cancelled and the val 1330 the speculation will be cancelled and the value reloaded: 1353 1331 1354 1332 : : +-------+ 1355 1333 +-------+ | | 1356 - 1334 --->| B->2 |------>| | 1357 1335 +-------+ | CPU 2 | 1358 1336 : :DIVIDE | | 1359 1337 +-------+ | | 1360 The CPU being busy doing a ---> - 1338 The CPU being busy doing a ---> --->| A->0 |~~~~ | | 1361 division speculates on the 1339 division speculates on the +-------+ ~ | | 1362 LOAD of A 1340 LOAD of A : : ~ | | 1363 1341 : :DIVIDE | | 1364 1342 : : ~ | | 1365 1343 : : ~ | | 1366 r 1344 rrrrrrrrrrrrrrrrr | | 1367 1345 +-------+ | | 1368 The speculation is discarded ---> - 1346 The speculation is discarded ---> --->| A->1 |------>| | 1369 and an updated value is 1347 and an updated value is +-------+ | | 1370 retrieved 1348 retrieved : : +-------+ 1371 1349 1372 1350 1373 MULTICOPY ATOMICITY 1351 MULTICOPY ATOMICITY 1374 -------------------- 1352 -------------------- 1375 1353 1376 Multicopy atomicity is a deeply intuitive not 1354 Multicopy atomicity is a deeply intuitive notion about ordering that is 1377 not always provided by real computer systems, 1355 not always provided by real computer systems, namely that a given store 1378 becomes visible at the same time to all CPUs, 1356 becomes visible at the same time to all CPUs, or, alternatively, that all 1379 CPUs agree on the order in which all stores b 1357 CPUs agree on the order in which all stores become visible. However, 1380 support of full multicopy atomicity would rul 1358 support of full multicopy atomicity would rule out valuable hardware 1381 optimizations, so a weaker form called ``othe 1359 optimizations, so a weaker form called ``other multicopy atomicity'' 1382 instead guarantees only that a given store be 1360 instead guarantees only that a given store becomes visible at the same 1383 time to all -other- CPUs. The remainder of t 1361 time to all -other- CPUs. The remainder of this document discusses this 1384 weaker form, but for brevity will call it sim 1362 weaker form, but for brevity will call it simply ``multicopy atomicity''. 1385 1363 1386 The following example demonstrates multicopy 1364 The following example demonstrates multicopy atomicity: 1387 1365 1388 CPU 1 CPU 2 1366 CPU 1 CPU 2 CPU 3 1389 ======================= ============= 1367 ======================= ======================= ======================= 1390 { X = 0, Y = 0 } 1368 { X = 0, Y = 0 } 1391 STORE X=1 r1=LOAD X (re 1369 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1) 1392 <general barr 1370 <general barrier> <read barrier> 1393 STORE Y=r1 1371 STORE Y=r1 LOAD X 1394 1372 1395 Suppose that CPU 2's load from X returns 1, w 1373 Suppose that CPU 2's load from X returns 1, which it then stores to Y, 1396 and CPU 3's load from Y returns 1. This indi 1374 and CPU 3's load from Y returns 1. This indicates that CPU 1's store 1397 to X precedes CPU 2's load from X and that CP 1375 to X precedes CPU 2's load from X and that CPU 2's store to Y precedes 1398 CPU 3's load from Y. In addition, the memory 1376 CPU 3's load from Y. In addition, the memory barriers guarantee that 1399 CPU 2 executes its load before its store, and 1377 CPU 2 executes its load before its store, and CPU 3 loads from Y before 1400 it loads from X. The question is then "Can C 1378 it loads from X. The question is then "Can CPU 3's load from X return 0?" 1401 1379 1402 Because CPU 3's load from X in some sense com 1380 Because CPU 3's load from X in some sense comes after CPU 2's load, it 1403 is natural to expect that CPU 3's load from X 1381 is natural to expect that CPU 3's load from X must therefore return 1. 1404 This expectation follows from multicopy atomi 1382 This expectation follows from multicopy atomicity: if a load executing 1405 on CPU B follows a load from the same variabl 1383 on CPU B follows a load from the same variable executing on CPU A (and 1406 CPU A did not originally store the value whic 1384 CPU A did not originally store the value which it read), then on 1407 multicopy-atomic systems, CPU B's load must r 1385 multicopy-atomic systems, CPU B's load must return either the same value 1408 that CPU A's load did or some later value. H 1386 that CPU A's load did or some later value. However, the Linux kernel 1409 does not require systems to be multicopy atom 1387 does not require systems to be multicopy atomic. 1410 1388 1411 The use of a general memory barrier in the ex 1389 The use of a general memory barrier in the example above compensates 1412 for any lack of multicopy atomicity. In the 1390 for any lack of multicopy atomicity. In the example, if CPU 2's load 1413 from X returns 1 and CPU 3's load from Y retu 1391 from X returns 1 and CPU 3's load from Y returns 1, then CPU 3's load 1414 from X must indeed also return 1. 1392 from X must indeed also return 1. 1415 1393 1416 However, dependencies, read barriers, and wri 1394 However, dependencies, read barriers, and write barriers are not always 1417 able to compensate for non-multicopy atomicit 1395 able to compensate for non-multicopy atomicity. For example, suppose 1418 that CPU 2's general barrier is removed from 1396 that CPU 2's general barrier is removed from the above example, leaving 1419 only the data dependency shown below: 1397 only the data dependency shown below: 1420 1398 1421 CPU 1 CPU 2 1399 CPU 1 CPU 2 CPU 3 1422 ======================= ============= 1400 ======================= ======================= ======================= 1423 { X = 0, Y = 0 } 1401 { X = 0, Y = 0 } 1424 STORE X=1 r1=LOAD X (re 1402 STORE X=1 r1=LOAD X (reads 1) LOAD Y (reads 1) 1425 <data depende 1403 <data dependency> <read barrier> 1426 STORE Y=r1 1404 STORE Y=r1 LOAD X (reads 0) 1427 1405 1428 This substitution allows non-multicopy atomic 1406 This substitution allows non-multicopy atomicity to run rampant: in 1429 this example, it is perfectly legal for CPU 2 1407 this example, it is perfectly legal for CPU 2's load from X to return 1, 1430 CPU 3's load from Y to return 1, and its load 1408 CPU 3's load from Y to return 1, and its load from X to return 0. 1431 1409 1432 The key point is that although CPU 2's data d 1410 The key point is that although CPU 2's data dependency orders its load 1433 and store, it does not guarantee to order CPU 1411 and store, it does not guarantee to order CPU 1's store. Thus, if this 1434 example runs on a non-multicopy-atomic system 1412 example runs on a non-multicopy-atomic system where CPUs 1 and 2 share a 1435 store buffer or a level of cache, CPU 2 might 1413 store buffer or a level of cache, CPU 2 might have early access to CPU 1's 1436 writes. General barriers are therefore requi 1414 writes. General barriers are therefore required to ensure that all CPUs 1437 agree on the combined order of multiple acces 1415 agree on the combined order of multiple accesses. 1438 1416 1439 General barriers can compensate not only for 1417 General barriers can compensate not only for non-multicopy atomicity, 1440 but can also generate additional ordering tha 1418 but can also generate additional ordering that can ensure that -all- 1441 CPUs will perceive the same order of -all- op 1419 CPUs will perceive the same order of -all- operations. In contrast, a 1442 chain of release-acquire pairs do not provide 1420 chain of release-acquire pairs do not provide this additional ordering, 1443 which means that only those CPUs on the chain 1421 which means that only those CPUs on the chain are guaranteed to agree 1444 on the combined order of the accesses. For e 1422 on the combined order of the accesses. For example, switching to C code 1445 in deference to the ghost of Herman Hollerith 1423 in deference to the ghost of Herman Hollerith: 1446 1424 1447 int u, v, x, y, z; 1425 int u, v, x, y, z; 1448 1426 1449 void cpu0(void) 1427 void cpu0(void) 1450 { 1428 { 1451 r0 = smp_load_acquire(&x); 1429 r0 = smp_load_acquire(&x); 1452 WRITE_ONCE(u, 1); 1430 WRITE_ONCE(u, 1); 1453 smp_store_release(&y, 1); 1431 smp_store_release(&y, 1); 1454 } 1432 } 1455 1433 1456 void cpu1(void) 1434 void cpu1(void) 1457 { 1435 { 1458 r1 = smp_load_acquire(&y); 1436 r1 = smp_load_acquire(&y); 1459 r4 = READ_ONCE(v); 1437 r4 = READ_ONCE(v); 1460 r5 = READ_ONCE(u); 1438 r5 = READ_ONCE(u); 1461 smp_store_release(&z, 1); 1439 smp_store_release(&z, 1); 1462 } 1440 } 1463 1441 1464 void cpu2(void) 1442 void cpu2(void) 1465 { 1443 { 1466 r2 = smp_load_acquire(&z); 1444 r2 = smp_load_acquire(&z); 1467 smp_store_release(&x, 1); 1445 smp_store_release(&x, 1); 1468 } 1446 } 1469 1447 1470 void cpu3(void) 1448 void cpu3(void) 1471 { 1449 { 1472 WRITE_ONCE(v, 1); 1450 WRITE_ONCE(v, 1); 1473 smp_mb(); 1451 smp_mb(); 1474 r3 = READ_ONCE(u); 1452 r3 = READ_ONCE(u); 1475 } 1453 } 1476 1454 1477 Because cpu0(), cpu1(), and cpu2() participat 1455 Because cpu0(), cpu1(), and cpu2() participate in a chain of 1478 smp_store_release()/smp_load_acquire() pairs, 1456 smp_store_release()/smp_load_acquire() pairs, the following outcome 1479 is prohibited: 1457 is prohibited: 1480 1458 1481 r0 == 1 && r1 == 1 && r2 == 1 1459 r0 == 1 && r1 == 1 && r2 == 1 1482 1460 1483 Furthermore, because of the release-acquire r 1461 Furthermore, because of the release-acquire relationship between cpu0() 1484 and cpu1(), cpu1() must see cpu0()'s writes, 1462 and cpu1(), cpu1() must see cpu0()'s writes, so that the following 1485 outcome is prohibited: 1463 outcome is prohibited: 1486 1464 1487 r1 == 1 && r5 == 0 1465 r1 == 1 && r5 == 0 1488 1466 1489 However, the ordering provided by a release-a 1467 However, the ordering provided by a release-acquire chain is local 1490 to the CPUs participating in that chain and d 1468 to the CPUs participating in that chain and does not apply to cpu3(), 1491 at least aside from stores. Therefore, the f 1469 at least aside from stores. Therefore, the following outcome is possible: 1492 1470 1493 r0 == 0 && r1 == 1 && r2 == 1 && r3 = 1471 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 1494 1472 1495 As an aside, the following outcome is also po 1473 As an aside, the following outcome is also possible: 1496 1474 1497 r0 == 0 && r1 == 1 && r2 == 1 && r3 = 1475 r0 == 0 && r1 == 1 && r2 == 1 && r3 == 0 && r4 == 0 && r5 == 1 1498 1476 1499 Although cpu0(), cpu1(), and cpu2() will see 1477 Although cpu0(), cpu1(), and cpu2() will see their respective reads and 1500 writes in order, CPUs not involved in the rel 1478 writes in order, CPUs not involved in the release-acquire chain might 1501 well disagree on the order. This disagreemen 1479 well disagree on the order. This disagreement stems from the fact that 1502 the weak memory-barrier instructions used to 1480 the weak memory-barrier instructions used to implement smp_load_acquire() 1503 and smp_store_release() are not required to o 1481 and smp_store_release() are not required to order prior stores against 1504 subsequent loads in all cases. This means th 1482 subsequent loads in all cases. This means that cpu3() can see cpu0()'s 1505 store to u as happening -after- cpu1()'s load 1483 store to u as happening -after- cpu1()'s load from v, even though 1506 both cpu0() and cpu1() agree that these two o 1484 both cpu0() and cpu1() agree that these two operations occurred in the 1507 intended order. 1485 intended order. 1508 1486 1509 However, please keep in mind that smp_load_ac 1487 However, please keep in mind that smp_load_acquire() is not magic. 1510 In particular, it simply reads from its argum 1488 In particular, it simply reads from its argument with ordering. It does 1511 -not- ensure that any particular value will b 1489 -not- ensure that any particular value will be read. Therefore, the 1512 following outcome is possible: 1490 following outcome is possible: 1513 1491 1514 r0 == 0 && r1 == 0 && r2 == 0 && r5 = 1492 r0 == 0 && r1 == 0 && r2 == 0 && r5 == 0 1515 1493 1516 Note that this outcome can happen even on a m 1494 Note that this outcome can happen even on a mythical sequentially 1517 consistent system where nothing is ever reord 1495 consistent system where nothing is ever reordered. 1518 1496 1519 To reiterate, if your code requires full orde 1497 To reiterate, if your code requires full ordering of all operations, 1520 use general barriers throughout. 1498 use general barriers throughout. 1521 1499 1522 1500 1523 ======================== 1501 ======================== 1524 EXPLICIT KERNEL BARRIERS 1502 EXPLICIT KERNEL BARRIERS 1525 ======================== 1503 ======================== 1526 1504 1527 The Linux kernel has a variety of different b 1505 The Linux kernel has a variety of different barriers that act at different 1528 levels: 1506 levels: 1529 1507 1530 (*) Compiler barrier. 1508 (*) Compiler barrier. 1531 1509 1532 (*) CPU memory barriers. 1510 (*) CPU memory barriers. 1533 1511 1534 1512 1535 COMPILER BARRIER 1513 COMPILER BARRIER 1536 ---------------- 1514 ---------------- 1537 1515 1538 The Linux kernel has an explicit compiler bar 1516 The Linux kernel has an explicit compiler barrier function that prevents the 1539 compiler from moving the memory accesses eith 1517 compiler from moving the memory accesses either side of it to the other side: 1540 1518 1541 barrier(); 1519 barrier(); 1542 1520 1543 This is a general barrier -- there are no rea 1521 This is a general barrier -- there are no read-read or write-write 1544 variants of barrier(). However, READ_ONCE() 1522 variants of barrier(). However, READ_ONCE() and WRITE_ONCE() can be 1545 thought of as weak forms of barrier() that af 1523 thought of as weak forms of barrier() that affect only the specific 1546 accesses flagged by the READ_ONCE() or WRITE_ 1524 accesses flagged by the READ_ONCE() or WRITE_ONCE(). 1547 1525 1548 The barrier() function has the following effe 1526 The barrier() function has the following effects: 1549 1527 1550 (*) Prevents the compiler from reordering ac 1528 (*) Prevents the compiler from reordering accesses following the 1551 barrier() to precede any accesses preced 1529 barrier() to precede any accesses preceding the barrier(). 1552 One example use for this property is to 1530 One example use for this property is to ease communication between 1553 interrupt-handler code and the code that 1531 interrupt-handler code and the code that was interrupted. 1554 1532 1555 (*) Within a loop, forces the compiler to lo 1533 (*) Within a loop, forces the compiler to load the variables used 1556 in that loop's conditional on each pass 1534 in that loop's conditional on each pass through that loop. 1557 1535 1558 The READ_ONCE() and WRITE_ONCE() functions ca 1536 The READ_ONCE() and WRITE_ONCE() functions can prevent any number of 1559 optimizations that, while perfectly safe in s 1537 optimizations that, while perfectly safe in single-threaded code, can 1560 be fatal in concurrent code. Here are some e 1538 be fatal in concurrent code. Here are some examples of these sorts 1561 of optimizations: 1539 of optimizations: 1562 1540 1563 (*) The compiler is within its rights to reo 1541 (*) The compiler is within its rights to reorder loads and stores 1564 to the same variable, and in some cases, 1542 to the same variable, and in some cases, the CPU is within its 1565 rights to reorder loads to the same vari 1543 rights to reorder loads to the same variable. This means that 1566 the following code: 1544 the following code: 1567 1545 1568 a[0] = x; 1546 a[0] = x; 1569 a[1] = x; 1547 a[1] = x; 1570 1548 1571 Might result in an older value of x stor 1549 Might result in an older value of x stored in a[1] than in a[0]. 1572 Prevent both the compiler and the CPU fr 1550 Prevent both the compiler and the CPU from doing this as follows: 1573 1551 1574 a[0] = READ_ONCE(x); 1552 a[0] = READ_ONCE(x); 1575 a[1] = READ_ONCE(x); 1553 a[1] = READ_ONCE(x); 1576 1554 1577 In short, READ_ONCE() and WRITE_ONCE() p 1555 In short, READ_ONCE() and WRITE_ONCE() provide cache coherence for 1578 accesses from multiple CPUs to a single 1556 accesses from multiple CPUs to a single variable. 1579 1557 1580 (*) The compiler is within its rights to mer 1558 (*) The compiler is within its rights to merge successive loads from 1581 the same variable. Such merging can cau 1559 the same variable. Such merging can cause the compiler to "optimize" 1582 the following code: 1560 the following code: 1583 1561 1584 while (tmp = a) 1562 while (tmp = a) 1585 do_something_with(tmp); 1563 do_something_with(tmp); 1586 1564 1587 into the following code, which, although 1565 into the following code, which, although in some sense legitimate 1588 for single-threaded code, is almost cert 1566 for single-threaded code, is almost certainly not what the developer 1589 intended: 1567 intended: 1590 1568 1591 if (tmp = a) 1569 if (tmp = a) 1592 for (;;) 1570 for (;;) 1593 do_something_with(tmp 1571 do_something_with(tmp); 1594 1572 1595 Use READ_ONCE() to prevent the compiler 1573 Use READ_ONCE() to prevent the compiler from doing this to you: 1596 1574 1597 while (tmp = READ_ONCE(a)) 1575 while (tmp = READ_ONCE(a)) 1598 do_something_with(tmp); 1576 do_something_with(tmp); 1599 1577 1600 (*) The compiler is within its rights to rel 1578 (*) The compiler is within its rights to reload a variable, for example, 1601 in cases where high register pressure pr 1579 in cases where high register pressure prevents the compiler from 1602 keeping all data of interest in register 1580 keeping all data of interest in registers. The compiler might 1603 therefore optimize the variable 'tmp' ou 1581 therefore optimize the variable 'tmp' out of our previous example: 1604 1582 1605 while (tmp = a) 1583 while (tmp = a) 1606 do_something_with(tmp); 1584 do_something_with(tmp); 1607 1585 1608 This could result in the following code, 1586 This could result in the following code, which is perfectly safe in 1609 single-threaded code, but can be fatal i 1587 single-threaded code, but can be fatal in concurrent code: 1610 1588 1611 while (a) 1589 while (a) 1612 do_something_with(a); 1590 do_something_with(a); 1613 1591 1614 For example, the optimized version of th 1592 For example, the optimized version of this code could result in 1615 passing a zero to do_something_with() in 1593 passing a zero to do_something_with() in the case where the variable 1616 a was modified by some other CPU between 1594 a was modified by some other CPU between the "while" statement and 1617 the call to do_something_with(). 1595 the call to do_something_with(). 1618 1596 1619 Again, use READ_ONCE() to prevent the co 1597 Again, use READ_ONCE() to prevent the compiler from doing this: 1620 1598 1621 while (tmp = READ_ONCE(a)) 1599 while (tmp = READ_ONCE(a)) 1622 do_something_with(tmp); 1600 do_something_with(tmp); 1623 1601 1624 Note that if the compiler runs short of 1602 Note that if the compiler runs short of registers, it might save 1625 tmp onto the stack. The overhead of thi 1603 tmp onto the stack. The overhead of this saving and later restoring 1626 is why compilers reload variables. Doin 1604 is why compilers reload variables. Doing so is perfectly safe for 1627 single-threaded code, so you need to tel 1605 single-threaded code, so you need to tell the compiler about cases 1628 where it is not safe. 1606 where it is not safe. 1629 1607 1630 (*) The compiler is within its rights to omi 1608 (*) The compiler is within its rights to omit a load entirely if it knows 1631 what the value will be. For example, if 1609 what the value will be. For example, if the compiler can prove that 1632 the value of variable 'a' is always zero 1610 the value of variable 'a' is always zero, it can optimize this code: 1633 1611 1634 while (tmp = a) 1612 while (tmp = a) 1635 do_something_with(tmp); 1613 do_something_with(tmp); 1636 1614 1637 Into this: 1615 Into this: 1638 1616 1639 do { } while (0); 1617 do { } while (0); 1640 1618 1641 This transformation is a win for single- 1619 This transformation is a win for single-threaded code because it 1642 gets rid of a load and a branch. The pr 1620 gets rid of a load and a branch. The problem is that the compiler 1643 will carry out its proof assuming that t 1621 will carry out its proof assuming that the current CPU is the only 1644 one updating variable 'a'. If variable 1622 one updating variable 'a'. If variable 'a' is shared, then the 1645 compiler's proof will be erroneous. Use 1623 compiler's proof will be erroneous. Use READ_ONCE() to tell the 1646 compiler that it doesn't know as much as 1624 compiler that it doesn't know as much as it thinks it does: 1647 1625 1648 while (tmp = READ_ONCE(a)) 1626 while (tmp = READ_ONCE(a)) 1649 do_something_with(tmp); 1627 do_something_with(tmp); 1650 1628 1651 But please note that the compiler is als 1629 But please note that the compiler is also closely watching what you 1652 do with the value after the READ_ONCE(). 1630 do with the value after the READ_ONCE(). For example, suppose you 1653 do the following and MAX is a preprocess 1631 do the following and MAX is a preprocessor macro with the value 1: 1654 1632 1655 while ((tmp = READ_ONCE(a)) % MAX) 1633 while ((tmp = READ_ONCE(a)) % MAX) 1656 do_something_with(tmp); 1634 do_something_with(tmp); 1657 1635 1658 Then the compiler knows that the result 1636 Then the compiler knows that the result of the "%" operator applied 1659 to MAX will always be zero, again allowi 1637 to MAX will always be zero, again allowing the compiler to optimize 1660 the code into near-nonexistence. (It wi 1638 the code into near-nonexistence. (It will still load from the 1661 variable 'a'.) 1639 variable 'a'.) 1662 1640 1663 (*) Similarly, the compiler is within its ri 1641 (*) Similarly, the compiler is within its rights to omit a store entirely 1664 if it knows that the variable already ha 1642 if it knows that the variable already has the value being stored. 1665 Again, the compiler assumes that the cur 1643 Again, the compiler assumes that the current CPU is the only one 1666 storing into the variable, which can cau 1644 storing into the variable, which can cause the compiler to do the 1667 wrong thing for shared variables. For e 1645 wrong thing for shared variables. For example, suppose you have 1668 the following: 1646 the following: 1669 1647 1670 a = 0; 1648 a = 0; 1671 ... Code that does not store to varia 1649 ... Code that does not store to variable a ... 1672 a = 0; 1650 a = 0; 1673 1651 1674 The compiler sees that the value of vari 1652 The compiler sees that the value of variable 'a' is already zero, so 1675 it might well omit the second store. Th 1653 it might well omit the second store. This would come as a fatal 1676 surprise if some other CPU might have st 1654 surprise if some other CPU might have stored to variable 'a' in the 1677 meantime. 1655 meantime. 1678 1656 1679 Use WRITE_ONCE() to prevent the compiler 1657 Use WRITE_ONCE() to prevent the compiler from making this sort of 1680 wrong guess: 1658 wrong guess: 1681 1659 1682 WRITE_ONCE(a, 0); 1660 WRITE_ONCE(a, 0); 1683 ... Code that does not store to varia 1661 ... Code that does not store to variable a ... 1684 WRITE_ONCE(a, 0); 1662 WRITE_ONCE(a, 0); 1685 1663 1686 (*) The compiler is within its rights to reo 1664 (*) The compiler is within its rights to reorder memory accesses unless 1687 you tell it not to. For example, consid 1665 you tell it not to. For example, consider the following interaction 1688 between process-level code and an interr 1666 between process-level code and an interrupt handler: 1689 1667 1690 void process_level(void) 1668 void process_level(void) 1691 { 1669 { 1692 msg = get_message(); 1670 msg = get_message(); 1693 flag = true; 1671 flag = true; 1694 } 1672 } 1695 1673 1696 void interrupt_handler(void) 1674 void interrupt_handler(void) 1697 { 1675 { 1698 if (flag) 1676 if (flag) 1699 process_message(msg); 1677 process_message(msg); 1700 } 1678 } 1701 1679 1702 There is nothing to prevent the compiler 1680 There is nothing to prevent the compiler from transforming 1703 process_level() to the following, in fac 1681 process_level() to the following, in fact, this might well be a 1704 win for single-threaded code: 1682 win for single-threaded code: 1705 1683 1706 void process_level(void) 1684 void process_level(void) 1707 { 1685 { 1708 flag = true; 1686 flag = true; 1709 msg = get_message(); 1687 msg = get_message(); 1710 } 1688 } 1711 1689 1712 If the interrupt occurs between these tw 1690 If the interrupt occurs between these two statement, then 1713 interrupt_handler() might be passed a ga 1691 interrupt_handler() might be passed a garbled msg. Use WRITE_ONCE() 1714 to prevent this as follows: 1692 to prevent this as follows: 1715 1693 1716 void process_level(void) 1694 void process_level(void) 1717 { 1695 { 1718 WRITE_ONCE(msg, get_message() 1696 WRITE_ONCE(msg, get_message()); 1719 WRITE_ONCE(flag, true); 1697 WRITE_ONCE(flag, true); 1720 } 1698 } 1721 1699 1722 void interrupt_handler(void) 1700 void interrupt_handler(void) 1723 { 1701 { 1724 if (READ_ONCE(flag)) 1702 if (READ_ONCE(flag)) 1725 process_message(READ_ 1703 process_message(READ_ONCE(msg)); 1726 } 1704 } 1727 1705 1728 Note that the READ_ONCE() and WRITE_ONCE 1706 Note that the READ_ONCE() and WRITE_ONCE() wrappers in 1729 interrupt_handler() are needed if this i 1707 interrupt_handler() are needed if this interrupt handler can itself 1730 be interrupted by something that also ac 1708 be interrupted by something that also accesses 'flag' and 'msg', 1731 for example, a nested interrupt or an NM 1709 for example, a nested interrupt or an NMI. Otherwise, READ_ONCE() 1732 and WRITE_ONCE() are not needed in inter 1710 and WRITE_ONCE() are not needed in interrupt_handler() other than 1733 for documentation purposes. (Note also 1711 for documentation purposes. (Note also that nested interrupts 1734 do not typically occur in modern Linux k 1712 do not typically occur in modern Linux kernels, in fact, if an 1735 interrupt handler returns with interrupt 1713 interrupt handler returns with interrupts enabled, you will get a 1736 WARN_ONCE() splat.) 1714 WARN_ONCE() splat.) 1737 1715 1738 You should assume that the compiler can 1716 You should assume that the compiler can move READ_ONCE() and 1739 WRITE_ONCE() past code not containing RE 1717 WRITE_ONCE() past code not containing READ_ONCE(), WRITE_ONCE(), 1740 barrier(), or similar primitives. 1718 barrier(), or similar primitives. 1741 1719 1742 This effect could also be achieved using 1720 This effect could also be achieved using barrier(), but READ_ONCE() 1743 and WRITE_ONCE() are more selective: Wi 1721 and WRITE_ONCE() are more selective: With READ_ONCE() and 1744 WRITE_ONCE(), the compiler need only for 1722 WRITE_ONCE(), the compiler need only forget the contents of the 1745 indicated memory locations, while with b 1723 indicated memory locations, while with barrier() the compiler must 1746 discard the value of all memory location 1724 discard the value of all memory locations that it has currently 1747 cached in any machine registers. Of cou 1725 cached in any machine registers. Of course, the compiler must also 1748 respect the order in which the READ_ONCE 1726 respect the order in which the READ_ONCE()s and WRITE_ONCE()s occur, 1749 though the CPU of course need not do so. 1727 though the CPU of course need not do so. 1750 1728 1751 (*) The compiler is within its rights to inv 1729 (*) The compiler is within its rights to invent stores to a variable, 1752 as in the following example: 1730 as in the following example: 1753 1731 1754 if (a) 1732 if (a) 1755 b = a; 1733 b = a; 1756 else 1734 else 1757 b = 42; 1735 b = 42; 1758 1736 1759 The compiler might save a branch by opti 1737 The compiler might save a branch by optimizing this as follows: 1760 1738 1761 b = 42; 1739 b = 42; 1762 if (a) 1740 if (a) 1763 b = a; 1741 b = a; 1764 1742 1765 In single-threaded code, this is not onl 1743 In single-threaded code, this is not only safe, but also saves 1766 a branch. Unfortunately, in concurrent 1744 a branch. Unfortunately, in concurrent code, this optimization 1767 could cause some other CPU to see a spur 1745 could cause some other CPU to see a spurious value of 42 -- even 1768 if variable 'a' was never zero -- when l 1746 if variable 'a' was never zero -- when loading variable 'b'. 1769 Use WRITE_ONCE() to prevent this as foll 1747 Use WRITE_ONCE() to prevent this as follows: 1770 1748 1771 if (a) 1749 if (a) 1772 WRITE_ONCE(b, a); 1750 WRITE_ONCE(b, a); 1773 else 1751 else 1774 WRITE_ONCE(b, 42); 1752 WRITE_ONCE(b, 42); 1775 1753 1776 The compiler can also invent loads. The 1754 The compiler can also invent loads. These are usually less 1777 damaging, but they can result in cache-l 1755 damaging, but they can result in cache-line bouncing and thus in 1778 poor performance and scalability. Use R 1756 poor performance and scalability. Use READ_ONCE() to prevent 1779 invented loads. 1757 invented loads. 1780 1758 1781 (*) For aligned memory locations whose size 1759 (*) For aligned memory locations whose size allows them to be accessed 1782 with a single memory-reference instructi 1760 with a single memory-reference instruction, prevents "load tearing" 1783 and "store tearing," in which a single l 1761 and "store tearing," in which a single large access is replaced by 1784 multiple smaller accesses. For example, 1762 multiple smaller accesses. For example, given an architecture having 1785 16-bit store instructions with 7-bit imm 1763 16-bit store instructions with 7-bit immediate fields, the compiler 1786 might be tempted to use two 16-bit store 1764 might be tempted to use two 16-bit store-immediate instructions to 1787 implement the following 32-bit store: 1765 implement the following 32-bit store: 1788 1766 1789 p = 0x00010002; 1767 p = 0x00010002; 1790 1768 1791 Please note that GCC really does use thi 1769 Please note that GCC really does use this sort of optimization, 1792 which is not surprising given that it wo 1770 which is not surprising given that it would likely take more 1793 than two instructions to build the const 1771 than two instructions to build the constant and then store it. 1794 This optimization can therefore be a win 1772 This optimization can therefore be a win in single-threaded code. 1795 In fact, a recent bug (since fixed) caus 1773 In fact, a recent bug (since fixed) caused GCC to incorrectly use 1796 this optimization in a volatile store. 1774 this optimization in a volatile store. In the absence of such bugs, 1797 use of WRITE_ONCE() prevents store teari 1775 use of WRITE_ONCE() prevents store tearing in the following example: 1798 1776 1799 WRITE_ONCE(p, 0x00010002); 1777 WRITE_ONCE(p, 0x00010002); 1800 1778 1801 Use of packed structures can also result 1779 Use of packed structures can also result in load and store tearing, 1802 as in this example: 1780 as in this example: 1803 1781 1804 struct __attribute__((__packed__)) fo 1782 struct __attribute__((__packed__)) foo { 1805 short a; 1783 short a; 1806 int b; 1784 int b; 1807 short c; 1785 short c; 1808 }; 1786 }; 1809 struct foo foo1, foo2; 1787 struct foo foo1, foo2; 1810 ... 1788 ... 1811 1789 1812 foo2.a = foo1.a; 1790 foo2.a = foo1.a; 1813 foo2.b = foo1.b; 1791 foo2.b = foo1.b; 1814 foo2.c = foo1.c; 1792 foo2.c = foo1.c; 1815 1793 1816 Because there are no READ_ONCE() or WRIT 1794 Because there are no READ_ONCE() or WRITE_ONCE() wrappers and no 1817 volatile markings, the compiler would be 1795 volatile markings, the compiler would be well within its rights to 1818 implement these three assignment stateme 1796 implement these three assignment statements as a pair of 32-bit 1819 loads followed by a pair of 32-bit store 1797 loads followed by a pair of 32-bit stores. This would result in 1820 load tearing on 'foo1.b' and store teari 1798 load tearing on 'foo1.b' and store tearing on 'foo2.b'. READ_ONCE() 1821 and WRITE_ONCE() again prevent tearing i 1799 and WRITE_ONCE() again prevent tearing in this example: 1822 1800 1823 foo2.a = foo1.a; 1801 foo2.a = foo1.a; 1824 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b)) 1802 WRITE_ONCE(foo2.b, READ_ONCE(foo1.b)); 1825 foo2.c = foo1.c; 1803 foo2.c = foo1.c; 1826 1804 1827 All that aside, it is never necessary to use 1805 All that aside, it is never necessary to use READ_ONCE() and 1828 WRITE_ONCE() on a variable that has been mark 1806 WRITE_ONCE() on a variable that has been marked volatile. For example, 1829 because 'jiffies' is marked volatile, it is n 1807 because 'jiffies' is marked volatile, it is never necessary to 1830 say READ_ONCE(jiffies). The reason for this 1808 say READ_ONCE(jiffies). The reason for this is that READ_ONCE() and 1831 WRITE_ONCE() are implemented as volatile cast 1809 WRITE_ONCE() are implemented as volatile casts, which has no effect when 1832 its argument is already marked volatile. 1810 its argument is already marked volatile. 1833 1811 1834 Please note that these compiler barriers have 1812 Please note that these compiler barriers have no direct effect on the CPU, 1835 which may then reorder things however it wish 1813 which may then reorder things however it wishes. 1836 1814 1837 1815 1838 CPU MEMORY BARRIERS 1816 CPU MEMORY BARRIERS 1839 ------------------- 1817 ------------------- 1840 1818 1841 The Linux kernel has seven basic CPU memory b !! 1819 The Linux kernel has eight basic CPU memory barriers: 1842 1820 1843 TYPE MANDATORY !! 1821 TYPE MANDATORY SMP CONDITIONAL 1844 ======================= ============= !! 1822 =============== ======================= =========================== 1845 GENERAL mb() !! 1823 GENERAL mb() smp_mb() 1846 WRITE wmb() !! 1824 WRITE wmb() smp_wmb() 1847 READ rmb() !! 1825 READ rmb() smp_rmb() 1848 ADDRESS DEPENDENCY !! 1826 DATA DEPENDENCY READ_ONCE() 1849 1827 1850 1828 1851 All memory barriers except the address-depend !! 1829 All memory barriers except the data dependency barriers imply a compiler 1852 barrier. Address dependencies do not impose !! 1830 barrier. Data dependencies do not impose any additional compiler ordering. 1853 1831 1854 Aside: In the case of address dependencies, t !! 1832 Aside: In the case of data dependencies, the compiler would be expected 1855 to issue the loads in the correct order (eg. 1833 to issue the loads in the correct order (eg. `a[b]` would have to load 1856 the value of b before loading a[b]), however 1834 the value of b before loading a[b]), however there is no guarantee in 1857 the C specification that the compiler may not 1835 the C specification that the compiler may not speculate the value of b 1858 (eg. is equal to 1) and load a[b] before b (e 1836 (eg. is equal to 1) and load a[b] before b (eg. tmp = a[1]; if (b != 1) 1859 tmp = a[b]; ). There is also the problem of 1837 tmp = a[b]; ). There is also the problem of a compiler reloading b after 1860 having loaded a[b], thus having a newer copy 1838 having loaded a[b], thus having a newer copy of b than a[b]. A consensus 1861 has not yet been reached about these problems 1839 has not yet been reached about these problems, however the READ_ONCE() 1862 macro is a good place to start looking. 1840 macro is a good place to start looking. 1863 1841 1864 SMP memory barriers are reduced to compiler b 1842 SMP memory barriers are reduced to compiler barriers on uniprocessor compiled 1865 systems because it is assumed that a CPU will 1843 systems because it is assumed that a CPU will appear to be self-consistent, 1866 and will order overlapping accesses correctly 1844 and will order overlapping accesses correctly with respect to itself. 1867 However, see the subsection on "Virtual Machi 1845 However, see the subsection on "Virtual Machine Guests" below. 1868 1846 1869 [!] Note that SMP memory barriers _must_ be u 1847 [!] Note that SMP memory barriers _must_ be used to control the ordering of 1870 references to shared memory on SMP systems, t 1848 references to shared memory on SMP systems, though the use of locking instead 1871 is sufficient. 1849 is sufficient. 1872 1850 1873 Mandatory barriers should not be used to cont 1851 Mandatory barriers should not be used to control SMP effects, since mandatory 1874 barriers impose unnecessary overhead on both 1852 barriers impose unnecessary overhead on both SMP and UP systems. They may, 1875 however, be used to control MMIO effects on a 1853 however, be used to control MMIO effects on accesses through relaxed memory I/O 1876 windows. These barriers are required even on 1854 windows. These barriers are required even on non-SMP systems as they affect 1877 the order in which memory operations appear t 1855 the order in which memory operations appear to a device by prohibiting both the 1878 compiler and the CPU from reordering them. 1856 compiler and the CPU from reordering them. 1879 1857 1880 1858 1881 There are some more advanced barrier function 1859 There are some more advanced barrier functions: 1882 1860 1883 (*) smp_store_mb(var, value) 1861 (*) smp_store_mb(var, value) 1884 1862 1885 This assigns the value to the variable a 1863 This assigns the value to the variable and then inserts a full memory 1886 barrier after it. It isn't guaranteed t 1864 barrier after it. It isn't guaranteed to insert anything more than a 1887 compiler barrier in a UP compilation. 1865 compiler barrier in a UP compilation. 1888 1866 1889 1867 1890 (*) smp_mb__before_atomic(); 1868 (*) smp_mb__before_atomic(); 1891 (*) smp_mb__after_atomic(); 1869 (*) smp_mb__after_atomic(); 1892 1870 1893 These are for use with atomic RMW functi 1871 These are for use with atomic RMW functions that do not imply memory 1894 barriers, but where the code needs a mem 1872 barriers, but where the code needs a memory barrier. Examples for atomic 1895 RMW functions that do not imply a memory 1873 RMW functions that do not imply a memory barrier are e.g. add, 1896 subtract, (failed) conditional operation 1874 subtract, (failed) conditional operations, _relaxed functions, 1897 but not atomic_read or atomic_set. A com 1875 but not atomic_read or atomic_set. A common example where a memory 1898 barrier may be required is when atomic o 1876 barrier may be required is when atomic ops are used for reference 1899 counting. 1877 counting. 1900 1878 1901 These are also used for atomic RMW bitop 1879 These are also used for atomic RMW bitop functions that do not imply a 1902 memory barrier (such as set_bit and clea 1880 memory barrier (such as set_bit and clear_bit). 1903 1881 1904 As an example, consider a piece of code 1882 As an example, consider a piece of code that marks an object as being dead 1905 and then decrements the object's referen 1883 and then decrements the object's reference count: 1906 1884 1907 obj->dead = 1; 1885 obj->dead = 1; 1908 smp_mb__before_atomic(); 1886 smp_mb__before_atomic(); 1909 atomic_dec(&obj->ref_count); 1887 atomic_dec(&obj->ref_count); 1910 1888 1911 This makes sure that the death mark on t 1889 This makes sure that the death mark on the object is perceived to be set 1912 *before* the reference counter is decrem 1890 *before* the reference counter is decremented. 1913 1891 1914 See Documentation/atomic_{t,bitops}.txt 1892 See Documentation/atomic_{t,bitops}.txt for more information. 1915 1893 1916 1894 1917 (*) dma_wmb(); 1895 (*) dma_wmb(); 1918 (*) dma_rmb(); 1896 (*) dma_rmb(); 1919 (*) dma_mb(); << 1920 1897 1921 These are for use with consistent memory 1898 These are for use with consistent memory to guarantee the ordering 1922 of writes or reads of shared memory acce 1899 of writes or reads of shared memory accessible to both the CPU and a 1923 DMA capable device. See Documentation/co !! 1900 DMA capable device. 1924 information about consistent memory. << 1925 1901 1926 For example, consider a device driver th 1902 For example, consider a device driver that shares memory with a device 1927 and uses a descriptor status value to in 1903 and uses a descriptor status value to indicate if the descriptor belongs 1928 to the device or the CPU, and a doorbell 1904 to the device or the CPU, and a doorbell to notify it when new 1929 descriptors are available: 1905 descriptors are available: 1930 1906 1931 if (desc->status != DEVICE_OWN) { 1907 if (desc->status != DEVICE_OWN) { 1932 /* do not read data until we 1908 /* do not read data until we own descriptor */ 1933 dma_rmb(); 1909 dma_rmb(); 1934 1910 1935 /* read/modify data */ 1911 /* read/modify data */ 1936 read_data = desc->data; 1912 read_data = desc->data; 1937 desc->data = write_data; 1913 desc->data = write_data; 1938 1914 1939 /* flush modifications before 1915 /* flush modifications before status update */ 1940 dma_wmb(); 1916 dma_wmb(); 1941 1917 1942 /* assign ownership */ 1918 /* assign ownership */ 1943 desc->status = DEVICE_OWN; 1919 desc->status = DEVICE_OWN; 1944 1920 1945 /* Make descriptor status vis !! 1921 /* notify device of new descriptors */ 1946 * notify device of new descr << 1947 */ << 1948 writel(DESC_NOTIFY, doorbell) 1922 writel(DESC_NOTIFY, doorbell); 1949 } 1923 } 1950 1924 1951 The dma_rmb() allows us to guarantee tha !! 1925 The dma_rmb() allows us guarantee the device has released ownership 1952 before we read the data from the descrip 1926 before we read the data from the descriptor, and the dma_wmb() allows 1953 us to guarantee the data is written to t 1927 us to guarantee the data is written to the descriptor before the device 1954 can see it now has ownership. The dma_m !! 1928 can see it now has ownership. Note that, when using writel(), a prior 1955 a dma_wmb(). !! 1929 wmb() is not needed to guarantee that the cache coherent memory writes 1956 !! 1930 have completed before writing to the MMIO region. The cheaper 1957 Note that the dma_*() barriers do not pr !! 1931 writel_relaxed() does not provide this guarantee and must not be used 1958 accesses to MMIO regions. See the later !! 1932 here. 1959 subsection for more information about I/ !! 1933 >> 1934 See the subsection "Kernel I/O barrier effects" for more information on >> 1935 relaxed I/O accessors and the Documentation/core-api/dma-api.rst file for >> 1936 more information on consistent memory. 1960 1937 1961 (*) pmem_wmb(); 1938 (*) pmem_wmb(); 1962 1939 1963 This is for use with persistent memory t 1940 This is for use with persistent memory to ensure that stores for which 1964 modifications are written to persistent 1941 modifications are written to persistent storage reached a platform 1965 durability domain. 1942 durability domain. 1966 1943 1967 For example, after a non-temporal write 1944 For example, after a non-temporal write to pmem region, we use pmem_wmb() 1968 to ensure that stores have reached a pla 1945 to ensure that stores have reached a platform durability domain. This ensures 1969 that stores have updated persistent stor 1946 that stores have updated persistent storage before any data access or 1970 data transfer caused by subsequent instr 1947 data transfer caused by subsequent instructions is initiated. This is 1971 in addition to the ordering done by wmb( 1948 in addition to the ordering done by wmb(). 1972 1949 1973 For load from persistent memory, existin 1950 For load from persistent memory, existing read memory barriers are sufficient 1974 to ensure read ordering. 1951 to ensure read ordering. 1975 1952 1976 (*) io_stop_wc(); << 1977 << 1978 For memory accesses with write-combining << 1979 by ioremap_wc()), the CPU may wait for p << 1980 subsequent ones. io_stop_wc() can be use << 1981 write-combining memory accesses before t << 1982 such wait has performance implications. << 1983 << 1984 =============================== 1953 =============================== 1985 IMPLICIT KERNEL MEMORY BARRIERS 1954 IMPLICIT KERNEL MEMORY BARRIERS 1986 =============================== 1955 =============================== 1987 1956 1988 Some of the other functions in the linux kern 1957 Some of the other functions in the linux kernel imply memory barriers, amongst 1989 which are locking and scheduling functions. 1958 which are locking and scheduling functions. 1990 1959 1991 This specification is a _minimum_ guarantee; 1960 This specification is a _minimum_ guarantee; any particular architecture may 1992 provide more substantial guarantees, but thes 1961 provide more substantial guarantees, but these may not be relied upon outside 1993 of arch specific code. 1962 of arch specific code. 1994 1963 1995 1964 1996 LOCK ACQUISITION FUNCTIONS 1965 LOCK ACQUISITION FUNCTIONS 1997 -------------------------- 1966 -------------------------- 1998 1967 1999 The Linux kernel has a number of locking cons 1968 The Linux kernel has a number of locking constructs: 2000 1969 2001 (*) spin locks 1970 (*) spin locks 2002 (*) R/W spin locks 1971 (*) R/W spin locks 2003 (*) mutexes 1972 (*) mutexes 2004 (*) semaphores 1973 (*) semaphores 2005 (*) R/W semaphores 1974 (*) R/W semaphores 2006 1975 2007 In all cases there are variants on "ACQUIRE" 1976 In all cases there are variants on "ACQUIRE" operations and "RELEASE" operations 2008 for each construct. These operations all imp 1977 for each construct. These operations all imply certain barriers: 2009 1978 2010 (1) ACQUIRE operation implication: 1979 (1) ACQUIRE operation implication: 2011 1980 2012 Memory operations issued after the ACQUI 1981 Memory operations issued after the ACQUIRE will be completed after the 2013 ACQUIRE operation has completed. 1982 ACQUIRE operation has completed. 2014 1983 2015 Memory operations issued before the ACQU 1984 Memory operations issued before the ACQUIRE may be completed after 2016 the ACQUIRE operation has completed. 1985 the ACQUIRE operation has completed. 2017 1986 2018 (2) RELEASE operation implication: 1987 (2) RELEASE operation implication: 2019 1988 2020 Memory operations issued before the RELE 1989 Memory operations issued before the RELEASE will be completed before the 2021 RELEASE operation has completed. 1990 RELEASE operation has completed. 2022 1991 2023 Memory operations issued after the RELEA 1992 Memory operations issued after the RELEASE may be completed before the 2024 RELEASE operation has completed. 1993 RELEASE operation has completed. 2025 1994 2026 (3) ACQUIRE vs ACQUIRE implication: 1995 (3) ACQUIRE vs ACQUIRE implication: 2027 1996 2028 All ACQUIRE operations issued before ano 1997 All ACQUIRE operations issued before another ACQUIRE operation will be 2029 completed before that ACQUIRE operation. 1998 completed before that ACQUIRE operation. 2030 1999 2031 (4) ACQUIRE vs RELEASE implication: 2000 (4) ACQUIRE vs RELEASE implication: 2032 2001 2033 All ACQUIRE operations issued before a R 2002 All ACQUIRE operations issued before a RELEASE operation will be 2034 completed before the RELEASE operation. 2003 completed before the RELEASE operation. 2035 2004 2036 (5) Failed conditional ACQUIRE implication: 2005 (5) Failed conditional ACQUIRE implication: 2037 2006 2038 Certain locking variants of the ACQUIRE 2007 Certain locking variants of the ACQUIRE operation may fail, either due to 2039 being unable to get the lock immediately 2008 being unable to get the lock immediately, or due to receiving an unblocked 2040 signal while asleep waiting for the lock 2009 signal while asleep waiting for the lock to become available. Failed 2041 locks do not imply any sort of barrier. 2010 locks do not imply any sort of barrier. 2042 2011 2043 [!] Note: one of the consequences of lock ACQ 2012 [!] Note: one of the consequences of lock ACQUIREs and RELEASEs being only 2044 one-way barriers is that the effects of instr 2013 one-way barriers is that the effects of instructions outside of a critical 2045 section may seep into the inside of the criti 2014 section may seep into the inside of the critical section. 2046 2015 2047 An ACQUIRE followed by a RELEASE may not be a 2016 An ACQUIRE followed by a RELEASE may not be assumed to be full memory barrier 2048 because it is possible for an access precedin 2017 because it is possible for an access preceding the ACQUIRE to happen after the 2049 ACQUIRE, and an access following the RELEASE 2018 ACQUIRE, and an access following the RELEASE to happen before the RELEASE, and 2050 the two accesses can themselves then cross: 2019 the two accesses can themselves then cross: 2051 2020 2052 *A = a; 2021 *A = a; 2053 ACQUIRE M 2022 ACQUIRE M 2054 RELEASE M 2023 RELEASE M 2055 *B = b; 2024 *B = b; 2056 2025 2057 may occur as: 2026 may occur as: 2058 2027 2059 ACQUIRE M, STORE *B, STORE *A, RELEAS 2028 ACQUIRE M, STORE *B, STORE *A, RELEASE M 2060 2029 2061 When the ACQUIRE and RELEASE are a lock acqui 2030 When the ACQUIRE and RELEASE are a lock acquisition and release, 2062 respectively, this same reordering can occur 2031 respectively, this same reordering can occur if the lock's ACQUIRE and 2063 RELEASE are to the same lock variable, but on 2032 RELEASE are to the same lock variable, but only from the perspective of 2064 another CPU not holding that lock. In short, 2033 another CPU not holding that lock. In short, a ACQUIRE followed by an 2065 RELEASE may -not- be assumed to be a full mem 2034 RELEASE may -not- be assumed to be a full memory barrier. 2066 2035 2067 Similarly, the reverse case of a RELEASE foll 2036 Similarly, the reverse case of a RELEASE followed by an ACQUIRE does 2068 not imply a full memory barrier. Therefore, 2037 not imply a full memory barrier. Therefore, the CPU's execution of the 2069 critical sections corresponding to the RELEAS 2038 critical sections corresponding to the RELEASE and the ACQUIRE can cross, 2070 so that: 2039 so that: 2071 2040 2072 *A = a; 2041 *A = a; 2073 RELEASE M 2042 RELEASE M 2074 ACQUIRE N 2043 ACQUIRE N 2075 *B = b; 2044 *B = b; 2076 2045 2077 could occur as: 2046 could occur as: 2078 2047 2079 ACQUIRE N, STORE *B, STORE *A, RELEAS 2048 ACQUIRE N, STORE *B, STORE *A, RELEASE M 2080 2049 2081 It might appear that this reordering could in 2050 It might appear that this reordering could introduce a deadlock. 2082 However, this cannot happen because if such a 2051 However, this cannot happen because if such a deadlock threatened, 2083 the RELEASE would simply complete, thereby av 2052 the RELEASE would simply complete, thereby avoiding the deadlock. 2084 2053 2085 Why does this work? 2054 Why does this work? 2086 2055 2087 One key point is that we are only tal 2056 One key point is that we are only talking about the CPU doing 2088 the reordering, not the compiler. If 2057 the reordering, not the compiler. If the compiler (or, for 2089 that matter, the developer) switched 2058 that matter, the developer) switched the operations, deadlock 2090 -could- occur. 2059 -could- occur. 2091 2060 2092 But suppose the CPU reordered the ope 2061 But suppose the CPU reordered the operations. In this case, 2093 the unlock precedes the lock in the a 2062 the unlock precedes the lock in the assembly code. The CPU 2094 simply elected to try executing the l 2063 simply elected to try executing the later lock operation first. 2095 If there is a deadlock, this lock ope 2064 If there is a deadlock, this lock operation will simply spin (or 2096 try to sleep, but more on that later) 2065 try to sleep, but more on that later). The CPU will eventually 2097 execute the unlock operation (which p 2066 execute the unlock operation (which preceded the lock operation 2098 in the assembly code), which will unr 2067 in the assembly code), which will unravel the potential deadlock, 2099 allowing the lock operation to succee 2068 allowing the lock operation to succeed. 2100 2069 2101 But what if the lock is a sleeplock? 2070 But what if the lock is a sleeplock? In that case, the code will 2102 try to enter the scheduler, where it 2071 try to enter the scheduler, where it will eventually encounter 2103 a memory barrier, which will force th 2072 a memory barrier, which will force the earlier unlock operation 2104 to complete, again unraveling the dea 2073 to complete, again unraveling the deadlock. There might be 2105 a sleep-unlock race, but the locking 2074 a sleep-unlock race, but the locking primitive needs to resolve 2106 such races properly in any case. 2075 such races properly in any case. 2107 2076 2108 Locks and semaphores may not provide any guar 2077 Locks and semaphores may not provide any guarantee of ordering on UP compiled 2109 systems, and so cannot be counted on in such 2078 systems, and so cannot be counted on in such a situation to actually achieve 2110 anything at all - especially with respect to 2079 anything at all - especially with respect to I/O accesses - unless combined 2111 with interrupt disabling operations. 2080 with interrupt disabling operations. 2112 2081 2113 See also the section on "Inter-CPU acquiring 2082 See also the section on "Inter-CPU acquiring barrier effects". 2114 2083 2115 2084 2116 As an example, consider the following: 2085 As an example, consider the following: 2117 2086 2118 *A = a; 2087 *A = a; 2119 *B = b; 2088 *B = b; 2120 ACQUIRE 2089 ACQUIRE 2121 *C = c; 2090 *C = c; 2122 *D = d; 2091 *D = d; 2123 RELEASE 2092 RELEASE 2124 *E = e; 2093 *E = e; 2125 *F = f; 2094 *F = f; 2126 2095 2127 The following sequence of events is acceptabl 2096 The following sequence of events is acceptable: 2128 2097 2129 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RE 2098 ACQUIRE, {*F,*A}, *E, {*C,*D}, *B, RELEASE 2130 2099 2131 [+] Note that {*F,*A} indicates a com 2100 [+] Note that {*F,*A} indicates a combined access. 2132 2101 2133 But none of the following are: 2102 But none of the following are: 2134 2103 2135 {*F,*A}, *B, ACQUIRE, *C, *D, 2104 {*F,*A}, *B, ACQUIRE, *C, *D, RELEASE, *E 2136 *A, *B, *C, ACQUIRE, *D, 2105 *A, *B, *C, ACQUIRE, *D, RELEASE, *E, *F 2137 *A, *B, ACQUIRE, *C, 2106 *A, *B, ACQUIRE, *C, RELEASE, *D, *E, *F 2138 *B, ACQUIRE, *C, *D, 2107 *B, ACQUIRE, *C, *D, RELEASE, {*F,*A}, *E 2139 2108 2140 2109 2141 2110 2142 INTERRUPT DISABLING FUNCTIONS 2111 INTERRUPT DISABLING FUNCTIONS 2143 ----------------------------- 2112 ----------------------------- 2144 2113 2145 Functions that disable interrupts (ACQUIRE eq 2114 Functions that disable interrupts (ACQUIRE equivalent) and enable interrupts 2146 (RELEASE equivalent) will act as compiler bar 2115 (RELEASE equivalent) will act as compiler barriers only. So if memory or I/O 2147 barriers are required in such a situation, th 2116 barriers are required in such a situation, they must be provided from some 2148 other means. 2117 other means. 2149 2118 2150 2119 2151 SLEEP AND WAKE-UP FUNCTIONS 2120 SLEEP AND WAKE-UP FUNCTIONS 2152 --------------------------- 2121 --------------------------- 2153 2122 2154 Sleeping and waking on an event flagged in gl 2123 Sleeping and waking on an event flagged in global data can be viewed as an 2155 interaction between two pieces of data: the t 2124 interaction between two pieces of data: the task state of the task waiting for 2156 the event and the global data used to indicat 2125 the event and the global data used to indicate the event. To make sure that 2157 these appear to happen in the right order, th 2126 these appear to happen in the right order, the primitives to begin the process 2158 of going to sleep, and the primitives to init 2127 of going to sleep, and the primitives to initiate a wake up imply certain 2159 barriers. 2128 barriers. 2160 2129 2161 Firstly, the sleeper normally follows somethi 2130 Firstly, the sleeper normally follows something like this sequence of events: 2162 2131 2163 for (;;) { 2132 for (;;) { 2164 set_current_state(TASK_UNINTE 2133 set_current_state(TASK_UNINTERRUPTIBLE); 2165 if (event_indicated) 2134 if (event_indicated) 2166 break; 2135 break; 2167 schedule(); 2136 schedule(); 2168 } 2137 } 2169 2138 2170 A general memory barrier is interpolated auto 2139 A general memory barrier is interpolated automatically by set_current_state() 2171 after it has altered the task state: 2140 after it has altered the task state: 2172 2141 2173 CPU 1 2142 CPU 1 2174 =============================== 2143 =============================== 2175 set_current_state(); 2144 set_current_state(); 2176 smp_store_mb(); 2145 smp_store_mb(); 2177 STORE current->state 2146 STORE current->state 2178 <general barrier> 2147 <general barrier> 2179 LOAD event_indicated 2148 LOAD event_indicated 2180 2149 2181 set_current_state() may be wrapped by: 2150 set_current_state() may be wrapped by: 2182 2151 2183 prepare_to_wait(); 2152 prepare_to_wait(); 2184 prepare_to_wait_exclusive(); 2153 prepare_to_wait_exclusive(); 2185 2154 2186 which therefore also imply a general memory b 2155 which therefore also imply a general memory barrier after setting the state. 2187 The whole sequence above is available in vari 2156 The whole sequence above is available in various canned forms, all of which 2188 interpolate the memory barrier in the right p 2157 interpolate the memory barrier in the right place: 2189 2158 2190 wait_event(); 2159 wait_event(); 2191 wait_event_interruptible(); 2160 wait_event_interruptible(); 2192 wait_event_interruptible_exclusive(); 2161 wait_event_interruptible_exclusive(); 2193 wait_event_interruptible_timeout(); 2162 wait_event_interruptible_timeout(); 2194 wait_event_killable(); 2163 wait_event_killable(); 2195 wait_event_timeout(); 2164 wait_event_timeout(); 2196 wait_on_bit(); 2165 wait_on_bit(); 2197 wait_on_bit_lock(); 2166 wait_on_bit_lock(); 2198 2167 2199 2168 2200 Secondly, code that performs a wake up normal 2169 Secondly, code that performs a wake up normally follows something like this: 2201 2170 2202 event_indicated = 1; 2171 event_indicated = 1; 2203 wake_up(&event_wait_queue); 2172 wake_up(&event_wait_queue); 2204 2173 2205 or: 2174 or: 2206 2175 2207 event_indicated = 1; 2176 event_indicated = 1; 2208 wake_up_process(event_daemon); 2177 wake_up_process(event_daemon); 2209 2178 2210 A general memory barrier is executed by wake_ 2179 A general memory barrier is executed by wake_up() if it wakes something up. 2211 If it doesn't wake anything up then a memory 2180 If it doesn't wake anything up then a memory barrier may or may not be 2212 executed; you must not rely on it. The barri 2181 executed; you must not rely on it. The barrier occurs before the task state 2213 is accessed, in particular, it sits between t 2182 is accessed, in particular, it sits between the STORE to indicate the event 2214 and the STORE to set TASK_RUNNING: 2183 and the STORE to set TASK_RUNNING: 2215 2184 2216 CPU 1 (Sleeper) CPU 2 2185 CPU 1 (Sleeper) CPU 2 (Waker) 2217 =============================== ===== 2186 =============================== =============================== 2218 set_current_state(); STORE 2187 set_current_state(); STORE event_indicated 2219 smp_store_mb(); wake_ 2188 smp_store_mb(); wake_up(); 2220 STORE current->state ... 2189 STORE current->state ... 2221 <general barrier> <ge 2190 <general barrier> <general barrier> 2222 LOAD event_indicated if 2191 LOAD event_indicated if ((LOAD task->state) & TASK_NORMAL) 2223 S 2192 STORE task->state 2224 2193 2225 where "task" is the thread being woken up and 2194 where "task" is the thread being woken up and it equals CPU 1's "current". 2226 2195 2227 To repeat, a general memory barrier is guaran 2196 To repeat, a general memory barrier is guaranteed to be executed by wake_up() 2228 if something is actually awakened, but otherw 2197 if something is actually awakened, but otherwise there is no such guarantee. 2229 To see this, consider the following sequence 2198 To see this, consider the following sequence of events, where X and Y are both 2230 initially zero: 2199 initially zero: 2231 2200 2232 CPU 1 CPU 2 2201 CPU 1 CPU 2 2233 =============================== ===== 2202 =============================== =============================== 2234 X = 1; Y = 1 2203 X = 1; Y = 1; 2235 smp_mb(); wake_ 2204 smp_mb(); wake_up(); 2236 LOAD Y LOAD 2205 LOAD Y LOAD X 2237 2206 2238 If a wakeup does occur, one (at least) of the 2207 If a wakeup does occur, one (at least) of the two loads must see 1. If, on 2239 the other hand, a wakeup does not occur, both 2208 the other hand, a wakeup does not occur, both loads might see 0. 2240 2209 2241 wake_up_process() always executes a general m 2210 wake_up_process() always executes a general memory barrier. The barrier again 2242 occurs before the task state is accessed. In 2211 occurs before the task state is accessed. In particular, if the wake_up() in 2243 the previous snippet were replaced by a call 2212 the previous snippet were replaced by a call to wake_up_process() then one of 2244 the two loads would be guaranteed to see 1. 2213 the two loads would be guaranteed to see 1. 2245 2214 2246 The available waker functions include: 2215 The available waker functions include: 2247 2216 2248 complete(); 2217 complete(); 2249 wake_up(); 2218 wake_up(); 2250 wake_up_all(); 2219 wake_up_all(); 2251 wake_up_bit(); 2220 wake_up_bit(); 2252 wake_up_interruptible(); 2221 wake_up_interruptible(); 2253 wake_up_interruptible_all(); 2222 wake_up_interruptible_all(); 2254 wake_up_interruptible_nr(); 2223 wake_up_interruptible_nr(); 2255 wake_up_interruptible_poll(); 2224 wake_up_interruptible_poll(); 2256 wake_up_interruptible_sync(); 2225 wake_up_interruptible_sync(); 2257 wake_up_interruptible_sync_poll(); 2226 wake_up_interruptible_sync_poll(); 2258 wake_up_locked(); 2227 wake_up_locked(); 2259 wake_up_locked_poll(); 2228 wake_up_locked_poll(); 2260 wake_up_nr(); 2229 wake_up_nr(); 2261 wake_up_poll(); 2230 wake_up_poll(); 2262 wake_up_process(); 2231 wake_up_process(); 2263 2232 2264 In terms of memory ordering, these functions 2233 In terms of memory ordering, these functions all provide the same guarantees of 2265 a wake_up() (or stronger). 2234 a wake_up() (or stronger). 2266 2235 2267 [!] Note that the memory barriers implied by 2236 [!] Note that the memory barriers implied by the sleeper and the waker do _not_ 2268 order multiple stores before the wake-up with 2237 order multiple stores before the wake-up with respect to loads of those stored 2269 values after the sleeper has called set_curre 2238 values after the sleeper has called set_current_state(). For instance, if the 2270 sleeper does: 2239 sleeper does: 2271 2240 2272 set_current_state(TASK_INTERRUPTIBLE) 2241 set_current_state(TASK_INTERRUPTIBLE); 2273 if (event_indicated) 2242 if (event_indicated) 2274 break; 2243 break; 2275 __set_current_state(TASK_RUNNING); 2244 __set_current_state(TASK_RUNNING); 2276 do_something(my_data); 2245 do_something(my_data); 2277 2246 2278 and the waker does: 2247 and the waker does: 2279 2248 2280 my_data = value; 2249 my_data = value; 2281 event_indicated = 1; 2250 event_indicated = 1; 2282 wake_up(&event_wait_queue); 2251 wake_up(&event_wait_queue); 2283 2252 2284 there's no guarantee that the change to event 2253 there's no guarantee that the change to event_indicated will be perceived by 2285 the sleeper as coming after the change to my_ 2254 the sleeper as coming after the change to my_data. In such a circumstance, the 2286 code on both sides must interpolate its own m 2255 code on both sides must interpolate its own memory barriers between the 2287 separate data accesses. Thus the above sleep 2256 separate data accesses. Thus the above sleeper ought to do: 2288 2257 2289 set_current_state(TASK_INTERRUPTIBLE) 2258 set_current_state(TASK_INTERRUPTIBLE); 2290 if (event_indicated) { 2259 if (event_indicated) { 2291 smp_rmb(); 2260 smp_rmb(); 2292 do_something(my_data); 2261 do_something(my_data); 2293 } 2262 } 2294 2263 2295 and the waker should do: 2264 and the waker should do: 2296 2265 2297 my_data = value; 2266 my_data = value; 2298 smp_wmb(); 2267 smp_wmb(); 2299 event_indicated = 1; 2268 event_indicated = 1; 2300 wake_up(&event_wait_queue); 2269 wake_up(&event_wait_queue); 2301 2270 2302 2271 2303 MISCELLANEOUS FUNCTIONS 2272 MISCELLANEOUS FUNCTIONS 2304 ----------------------- 2273 ----------------------- 2305 2274 2306 Other functions that imply barriers: 2275 Other functions that imply barriers: 2307 2276 2308 (*) schedule() and similar imply full memory 2277 (*) schedule() and similar imply full memory barriers. 2309 2278 2310 2279 2311 =================================== 2280 =================================== 2312 INTER-CPU ACQUIRING BARRIER EFFECTS 2281 INTER-CPU ACQUIRING BARRIER EFFECTS 2313 =================================== 2282 =================================== 2314 2283 2315 On SMP systems locking primitives give a more 2284 On SMP systems locking primitives give a more substantial form of barrier: one 2316 that does affect memory access ordering on ot 2285 that does affect memory access ordering on other CPUs, within the context of 2317 conflict on any particular lock. 2286 conflict on any particular lock. 2318 2287 2319 2288 2320 ACQUIRES VS MEMORY ACCESSES 2289 ACQUIRES VS MEMORY ACCESSES 2321 --------------------------- 2290 --------------------------- 2322 2291 2323 Consider the following: the system has a pair 2292 Consider the following: the system has a pair of spinlocks (M) and (Q), and 2324 three CPUs; then should the following sequenc 2293 three CPUs; then should the following sequence of events occur: 2325 2294 2326 CPU 1 CPU 2 2295 CPU 1 CPU 2 2327 =============================== ===== 2296 =============================== =============================== 2328 WRITE_ONCE(*A, a); WRITE 2297 WRITE_ONCE(*A, a); WRITE_ONCE(*E, e); 2329 ACQUIRE M ACQUI 2298 ACQUIRE M ACQUIRE Q 2330 WRITE_ONCE(*B, b); WRITE 2299 WRITE_ONCE(*B, b); WRITE_ONCE(*F, f); 2331 WRITE_ONCE(*C, c); WRITE 2300 WRITE_ONCE(*C, c); WRITE_ONCE(*G, g); 2332 RELEASE M RELEA 2301 RELEASE M RELEASE Q 2333 WRITE_ONCE(*D, d); WRITE 2302 WRITE_ONCE(*D, d); WRITE_ONCE(*H, h); 2334 2303 2335 Then there is no guarantee as to what order C 2304 Then there is no guarantee as to what order CPU 3 will see the accesses to *A 2336 through *H occur in, other than the constrain 2305 through *H occur in, other than the constraints imposed by the separate locks 2337 on the separate CPUs. It might, for example, 2306 on the separate CPUs. It might, for example, see: 2338 2307 2339 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, 2308 *E, ACQUIRE M, ACQUIRE Q, *G, *C, *F, *A, *B, RELEASE Q, *D, *H, RELEASE M 2340 2309 2341 But it won't see any of: 2310 But it won't see any of: 2342 2311 2343 *B, *C or *D preceding ACQUIRE M 2312 *B, *C or *D preceding ACQUIRE M 2344 *A, *B or *C following RELEASE M 2313 *A, *B or *C following RELEASE M 2345 *F, *G or *H preceding ACQUIRE Q 2314 *F, *G or *H preceding ACQUIRE Q 2346 *E, *F or *G following RELEASE Q 2315 *E, *F or *G following RELEASE Q 2347 2316 2348 2317 2349 ================================= 2318 ================================= 2350 WHERE ARE MEMORY BARRIERS NEEDED? 2319 WHERE ARE MEMORY BARRIERS NEEDED? 2351 ================================= 2320 ================================= 2352 2321 2353 Under normal operation, memory operation reor 2322 Under normal operation, memory operation reordering is generally not going to 2354 be a problem as a single-threaded linear piec 2323 be a problem as a single-threaded linear piece of code will still appear to 2355 work correctly, even if it's in an SMP kernel 2324 work correctly, even if it's in an SMP kernel. There are, however, four 2356 circumstances in which reordering definitely 2325 circumstances in which reordering definitely _could_ be a problem: 2357 2326 2358 (*) Interprocessor interaction. 2327 (*) Interprocessor interaction. 2359 2328 2360 (*) Atomic operations. 2329 (*) Atomic operations. 2361 2330 2362 (*) Accessing devices. 2331 (*) Accessing devices. 2363 2332 2364 (*) Interrupts. 2333 (*) Interrupts. 2365 2334 2366 2335 2367 INTERPROCESSOR INTERACTION 2336 INTERPROCESSOR INTERACTION 2368 -------------------------- 2337 -------------------------- 2369 2338 2370 When there's a system with more than one proc 2339 When there's a system with more than one processor, more than one CPU in the 2371 system may be working on the same data set at 2340 system may be working on the same data set at the same time. This can cause 2372 synchronisation problems, and the usual way o 2341 synchronisation problems, and the usual way of dealing with them is to use 2373 locks. Locks, however, are quite expensive, 2342 locks. Locks, however, are quite expensive, and so it may be preferable to 2374 operate without the use of a lock if at all p 2343 operate without the use of a lock if at all possible. In such a case 2375 operations that affect both CPUs may have to 2344 operations that affect both CPUs may have to be carefully ordered to prevent 2376 a malfunction. 2345 a malfunction. 2377 2346 2378 Consider, for example, the R/W semaphore slow 2347 Consider, for example, the R/W semaphore slow path. Here a waiting process is 2379 queued on the semaphore, by virtue of it havi 2348 queued on the semaphore, by virtue of it having a piece of its stack linked to 2380 the semaphore's list of waiting processes: 2349 the semaphore's list of waiting processes: 2381 2350 2382 struct rw_semaphore { 2351 struct rw_semaphore { 2383 ... 2352 ... 2384 spinlock_t lock; 2353 spinlock_t lock; 2385 struct list_head waiters; 2354 struct list_head waiters; 2386 }; 2355 }; 2387 2356 2388 struct rwsem_waiter { 2357 struct rwsem_waiter { 2389 struct list_head list; 2358 struct list_head list; 2390 struct task_struct *task; 2359 struct task_struct *task; 2391 }; 2360 }; 2392 2361 2393 To wake up a particular waiter, the up_read() 2362 To wake up a particular waiter, the up_read() or up_write() functions have to: 2394 2363 2395 (1) read the next pointer from this waiter's 2364 (1) read the next pointer from this waiter's record to know as to where the 2396 next waiter record is; 2365 next waiter record is; 2397 2366 2398 (2) read the pointer to the waiter's task st 2367 (2) read the pointer to the waiter's task structure; 2399 2368 2400 (3) clear the task pointer to tell the waite 2369 (3) clear the task pointer to tell the waiter it has been given the semaphore; 2401 2370 2402 (4) call wake_up_process() on the task; and 2371 (4) call wake_up_process() on the task; and 2403 2372 2404 (5) release the reference held on the waiter 2373 (5) release the reference held on the waiter's task struct. 2405 2374 2406 In other words, it has to perform this sequen 2375 In other words, it has to perform this sequence of events: 2407 2376 2408 LOAD waiter->list.next; 2377 LOAD waiter->list.next; 2409 LOAD waiter->task; 2378 LOAD waiter->task; 2410 STORE waiter->task; 2379 STORE waiter->task; 2411 CALL wakeup 2380 CALL wakeup 2412 RELEASE task 2381 RELEASE task 2413 2382 2414 and if any of these steps occur out of order, 2383 and if any of these steps occur out of order, then the whole thing may 2415 malfunction. 2384 malfunction. 2416 2385 2417 Once it has queued itself and dropped the sem 2386 Once it has queued itself and dropped the semaphore lock, the waiter does not 2418 get the lock again; it instead just waits for 2387 get the lock again; it instead just waits for its task pointer to be cleared 2419 before proceeding. Since the record is on th 2388 before proceeding. Since the record is on the waiter's stack, this means that 2420 if the task pointer is cleared _before_ the n 2389 if the task pointer is cleared _before_ the next pointer in the list is read, 2421 another CPU might start processing the waiter 2390 another CPU might start processing the waiter and might clobber the waiter's 2422 stack before the up*() function has a chance 2391 stack before the up*() function has a chance to read the next pointer. 2423 2392 2424 Consider then what might happen to the above 2393 Consider then what might happen to the above sequence of events: 2425 2394 2426 CPU 1 CPU 2 2395 CPU 1 CPU 2 2427 =============================== ===== 2396 =============================== =============================== 2428 down_ 2397 down_xxx() 2429 Queue 2398 Queue waiter 2430 Sleep 2399 Sleep 2431 up_yyy() 2400 up_yyy() 2432 LOAD waiter->task; 2401 LOAD waiter->task; 2433 STORE waiter->task; 2402 STORE waiter->task; 2434 Woken 2403 Woken up by other event 2435 <preempt> 2404 <preempt> 2436 Resum 2405 Resume processing 2437 down_ 2406 down_xxx() returns 2438 call 2407 call foo() 2439 foo() 2408 foo() clobbers *waiter 2440 </preempt> 2409 </preempt> 2441 LOAD waiter->list.next; 2410 LOAD waiter->list.next; 2442 --- OOPS --- 2411 --- OOPS --- 2443 2412 2444 This could be dealt with using the semaphore 2413 This could be dealt with using the semaphore lock, but then the down_xxx() 2445 function has to needlessly get the spinlock a 2414 function has to needlessly get the spinlock again after being woken up. 2446 2415 2447 The way to deal with this is to insert a gene 2416 The way to deal with this is to insert a general SMP memory barrier: 2448 2417 2449 LOAD waiter->list.next; 2418 LOAD waiter->list.next; 2450 LOAD waiter->task; 2419 LOAD waiter->task; 2451 smp_mb(); 2420 smp_mb(); 2452 STORE waiter->task; 2421 STORE waiter->task; 2453 CALL wakeup 2422 CALL wakeup 2454 RELEASE task 2423 RELEASE task 2455 2424 2456 In this case, the barrier makes a guarantee t 2425 In this case, the barrier makes a guarantee that all memory accesses before the 2457 barrier will appear to happen before all the 2426 barrier will appear to happen before all the memory accesses after the barrier 2458 with respect to the other CPUs on the system. 2427 with respect to the other CPUs on the system. It does _not_ guarantee that all 2459 the memory accesses before the barrier will b 2428 the memory accesses before the barrier will be complete by the time the barrier 2460 instruction itself is complete. 2429 instruction itself is complete. 2461 2430 2462 On a UP system - where this wouldn't be a pro 2431 On a UP system - where this wouldn't be a problem - the smp_mb() is just a 2463 compiler barrier, thus making sure the compil 2432 compiler barrier, thus making sure the compiler emits the instructions in the 2464 right order without actually intervening in t 2433 right order without actually intervening in the CPU. Since there's only one 2465 CPU, that CPU's dependency ordering logic wil 2434 CPU, that CPU's dependency ordering logic will take care of everything else. 2466 2435 2467 2436 2468 ATOMIC OPERATIONS 2437 ATOMIC OPERATIONS 2469 ----------------- 2438 ----------------- 2470 2439 2471 While they are technically interprocessor int 2440 While they are technically interprocessor interaction considerations, atomic 2472 operations are noted specially as some of the 2441 operations are noted specially as some of them imply full memory barriers and 2473 some don't, but they're very heavily relied o 2442 some don't, but they're very heavily relied on as a group throughout the 2474 kernel. 2443 kernel. 2475 2444 2476 See Documentation/atomic_t.txt for more infor 2445 See Documentation/atomic_t.txt for more information. 2477 2446 2478 2447 2479 ACCESSING DEVICES 2448 ACCESSING DEVICES 2480 ----------------- 2449 ----------------- 2481 2450 2482 Many devices can be memory mapped, and so app 2451 Many devices can be memory mapped, and so appear to the CPU as if they're just 2483 a set of memory locations. To control such a 2452 a set of memory locations. To control such a device, the driver usually has to 2484 make the right memory accesses in exactly the 2453 make the right memory accesses in exactly the right order. 2485 2454 2486 However, having a clever CPU or a clever comp 2455 However, having a clever CPU or a clever compiler creates a potential problem 2487 in that the carefully sequenced accesses in t 2456 in that the carefully sequenced accesses in the driver code won't reach the 2488 device in the requisite order if the CPU or t 2457 device in the requisite order if the CPU or the compiler thinks it is more 2489 efficient to reorder, combine or merge access 2458 efficient to reorder, combine or merge accesses - something that would cause 2490 the device to malfunction. 2459 the device to malfunction. 2491 2460 2492 Inside of the Linux kernel, I/O should be don 2461 Inside of the Linux kernel, I/O should be done through the appropriate accessor 2493 routines - such as inb() or writel() - which 2462 routines - such as inb() or writel() - which know how to make such accesses 2494 appropriately sequential. While this, for th 2463 appropriately sequential. While this, for the most part, renders the explicit 2495 use of memory barriers unnecessary, if the ac 2464 use of memory barriers unnecessary, if the accessor functions are used to refer 2496 to an I/O memory window with relaxed memory a 2465 to an I/O memory window with relaxed memory access properties, then _mandatory_ 2497 memory barriers are required to enforce order 2466 memory barriers are required to enforce ordering. 2498 2467 2499 See Documentation/driver-api/device-io.rst fo 2468 See Documentation/driver-api/device-io.rst for more information. 2500 2469 2501 2470 2502 INTERRUPTS 2471 INTERRUPTS 2503 ---------- 2472 ---------- 2504 2473 2505 A driver may be interrupted by its own interr 2474 A driver may be interrupted by its own interrupt service routine, and thus the 2506 two parts of the driver may interfere with ea 2475 two parts of the driver may interfere with each other's attempts to control or 2507 access the device. 2476 access the device. 2508 2477 2509 This may be alleviated - at least in part - b 2478 This may be alleviated - at least in part - by disabling local interrupts (a 2510 form of locking), such that the critical oper 2479 form of locking), such that the critical operations are all contained within 2511 the interrupt-disabled section in the driver. 2480 the interrupt-disabled section in the driver. While the driver's interrupt 2512 routine is executing, the driver's core may n 2481 routine is executing, the driver's core may not run on the same CPU, and its 2513 interrupt is not permitted to happen again un 2482 interrupt is not permitted to happen again until the current interrupt has been 2514 handled, thus the interrupt handler does not 2483 handled, thus the interrupt handler does not need to lock against that. 2515 2484 2516 However, consider a driver that was talking t 2485 However, consider a driver that was talking to an ethernet card that sports an 2517 address register and a data register. If tha 2486 address register and a data register. If that driver's core talks to the card 2518 under interrupt-disablement and then the driv 2487 under interrupt-disablement and then the driver's interrupt handler is invoked: 2519 2488 2520 LOCAL IRQ DISABLE 2489 LOCAL IRQ DISABLE 2521 writew(ADDR, 3); 2490 writew(ADDR, 3); 2522 writew(DATA, y); 2491 writew(DATA, y); 2523 LOCAL IRQ ENABLE 2492 LOCAL IRQ ENABLE 2524 <interrupt> 2493 <interrupt> 2525 writew(ADDR, 4); 2494 writew(ADDR, 4); 2526 q = readw(DATA); 2495 q = readw(DATA); 2527 </interrupt> 2496 </interrupt> 2528 2497 2529 The store to the data register might happen a 2498 The store to the data register might happen after the second store to the 2530 address register if ordering rules are suffic 2499 address register if ordering rules are sufficiently relaxed: 2531 2500 2532 STORE *ADDR = 3, STORE *ADDR = 4, STO 2501 STORE *ADDR = 3, STORE *ADDR = 4, STORE *DATA = y, q = LOAD *DATA 2533 2502 2534 2503 2535 If ordering rules are relaxed, it must be ass 2504 If ordering rules are relaxed, it must be assumed that accesses done inside an 2536 interrupt disabled section may leak outside o 2505 interrupt disabled section may leak outside of it and may interleave with 2537 accesses performed in an interrupt - and vice 2506 accesses performed in an interrupt - and vice versa - unless implicit or 2538 explicit barriers are used. 2507 explicit barriers are used. 2539 2508 2540 Normally this won't be a problem because the 2509 Normally this won't be a problem because the I/O accesses done inside such 2541 sections will include synchronous load operat 2510 sections will include synchronous load operations on strictly ordered I/O 2542 registers that form implicit I/O barriers. 2511 registers that form implicit I/O barriers. 2543 2512 2544 2513 2545 A similar situation may occur between an inte 2514 A similar situation may occur between an interrupt routine and two routines 2546 running on separate CPUs that communicate wit 2515 running on separate CPUs that communicate with each other. If such a case is 2547 likely, then interrupt-disabling locks should 2516 likely, then interrupt-disabling locks should be used to guarantee ordering. 2548 2517 2549 2518 2550 ========================== 2519 ========================== 2551 KERNEL I/O BARRIER EFFECTS 2520 KERNEL I/O BARRIER EFFECTS 2552 ========================== 2521 ========================== 2553 2522 2554 Interfacing with peripherals via I/O accesses 2523 Interfacing with peripherals via I/O accesses is deeply architecture and device 2555 specific. Therefore, drivers which are inhere 2524 specific. Therefore, drivers which are inherently non-portable may rely on 2556 specific behaviours of their target systems i 2525 specific behaviours of their target systems in order to achieve synchronization 2557 in the most lightweight manner possible. For 2526 in the most lightweight manner possible. For drivers intending to be portable 2558 between multiple architectures and bus implem 2527 between multiple architectures and bus implementations, the kernel offers a 2559 series of accessor functions that provide var 2528 series of accessor functions that provide various degrees of ordering 2560 guarantees: 2529 guarantees: 2561 2530 2562 (*) readX(), writeX(): 2531 (*) readX(), writeX(): 2563 2532 2564 The readX() and writeX() MMIO accesso 2533 The readX() and writeX() MMIO accessors take a pointer to the 2565 peripheral being accessed as an __iom 2534 peripheral being accessed as an __iomem * parameter. For pointers 2566 mapped with the default I/O attribute 2535 mapped with the default I/O attributes (e.g. those returned by 2567 ioremap()), the ordering guarantees a 2536 ioremap()), the ordering guarantees are as follows: 2568 2537 2569 1. All readX() and writeX() accesses 2538 1. All readX() and writeX() accesses to the same peripheral are ordered 2570 with respect to each other. This e 2539 with respect to each other. This ensures that MMIO register accesses 2571 by the same CPU thread to a partic 2540 by the same CPU thread to a particular device will arrive in program 2572 order. 2541 order. 2573 2542 2574 2. A writeX() issued by a CPU thread 2543 2. A writeX() issued by a CPU thread holding a spinlock is ordered 2575 before a writeX() to the same peri 2544 before a writeX() to the same peripheral from another CPU thread 2576 issued after a later acquisition o 2545 issued after a later acquisition of the same spinlock. This ensures 2577 that MMIO register writes to a par 2546 that MMIO register writes to a particular device issued while holding 2578 a spinlock will arrive in an order 2547 a spinlock will arrive in an order consistent with acquisitions of 2579 the lock. 2548 the lock. 2580 2549 2581 3. A writeX() by a CPU thread to the 2550 3. A writeX() by a CPU thread to the peripheral will first wait for the 2582 completion of all prior writes to 2551 completion of all prior writes to memory either issued by, or 2583 propagated to, the same thread. Th 2552 propagated to, the same thread. This ensures that writes by the CPU 2584 to an outbound DMA buffer allocate 2553 to an outbound DMA buffer allocated by dma_alloc_coherent() will be 2585 visible to a DMA engine when the C 2554 visible to a DMA engine when the CPU writes to its MMIO control 2586 register to trigger the transfer. 2555 register to trigger the transfer. 2587 2556 2588 4. A readX() by a CPU thread from the 2557 4. A readX() by a CPU thread from the peripheral will complete before 2589 any subsequent reads from memory b 2558 any subsequent reads from memory by the same thread can begin. This 2590 ensures that reads by the CPU from 2559 ensures that reads by the CPU from an incoming DMA buffer allocated 2591 by dma_alloc_coherent() will not s 2560 by dma_alloc_coherent() will not see stale data after reading from 2592 the DMA engine's MMIO status regis 2561 the DMA engine's MMIO status register to establish that the DMA 2593 transfer has completed. 2562 transfer has completed. 2594 2563 2595 5. A readX() by a CPU thread from the 2564 5. A readX() by a CPU thread from the peripheral will complete before 2596 any subsequent delay() loop can be 2565 any subsequent delay() loop can begin execution on the same thread. 2597 This ensures that two MMIO registe 2566 This ensures that two MMIO register writes by the CPU to a peripheral 2598 will arrive at least 1us apart if 2567 will arrive at least 1us apart if the first write is immediately read 2599 back with readX() and udelay(1) is 2568 back with readX() and udelay(1) is called prior to the second 2600 writeX(): 2569 writeX(): 2601 2570 2602 writel(42, DEVICE_REGISTER_0) 2571 writel(42, DEVICE_REGISTER_0); // Arrives at the device... 2603 readl(DEVICE_REGISTER_0); 2572 readl(DEVICE_REGISTER_0); 2604 udelay(1); 2573 udelay(1); 2605 writel(42, DEVICE_REGISTER_1) 2574 writel(42, DEVICE_REGISTER_1); // ...at least 1us before this. 2606 2575 2607 The ordering properties of __iomem po 2576 The ordering properties of __iomem pointers obtained with non-default 2608 attributes (e.g. those returned by io 2577 attributes (e.g. those returned by ioremap_wc()) are specific to the 2609 underlying architecture and therefore 2578 underlying architecture and therefore the guarantees listed above cannot 2610 generally be relied upon for accesses 2579 generally be relied upon for accesses to these types of mappings. 2611 2580 2612 (*) readX_relaxed(), writeX_relaxed(): 2581 (*) readX_relaxed(), writeX_relaxed(): 2613 2582 2614 These are similar to readX() and writ 2583 These are similar to readX() and writeX(), but provide weaker memory 2615 ordering guarantees. Specifically, th 2584 ordering guarantees. Specifically, they do not guarantee ordering with 2616 respect to locking, normal memory acc 2585 respect to locking, normal memory accesses or delay() loops (i.e. 2617 bullets 2-5 above) but they are still 2586 bullets 2-5 above) but they are still guaranteed to be ordered with 2618 respect to other accesses from the sa 2587 respect to other accesses from the same CPU thread to the same 2619 peripheral when operating on __iomem 2588 peripheral when operating on __iomem pointers mapped with the default 2620 I/O attributes. 2589 I/O attributes. 2621 2590 2622 (*) readsX(), writesX(): 2591 (*) readsX(), writesX(): 2623 2592 2624 The readsX() and writesX() MMIO acces 2593 The readsX() and writesX() MMIO accessors are designed for accessing 2625 register-based, memory-mapped FIFOs r 2594 register-based, memory-mapped FIFOs residing on peripherals that are not 2626 capable of performing DMA. Consequent 2595 capable of performing DMA. Consequently, they provide only the ordering 2627 guarantees of readX_relaxed() and wri 2596 guarantees of readX_relaxed() and writeX_relaxed(), as documented above. 2628 2597 2629 (*) inX(), outX(): 2598 (*) inX(), outX(): 2630 2599 2631 The inX() and outX() accessors are in 2600 The inX() and outX() accessors are intended to access legacy port-mapped 2632 I/O peripherals, which may require sp 2601 I/O peripherals, which may require special instructions on some 2633 architectures (notably x86). The port 2602 architectures (notably x86). The port number of the peripheral being 2634 accessed is passed as an argument. 2603 accessed is passed as an argument. 2635 2604 2636 Since many CPU architectures ultimate 2605 Since many CPU architectures ultimately access these peripherals via an 2637 internal virtual memory mapping, the 2606 internal virtual memory mapping, the portable ordering guarantees 2638 provided by inX() and outX() are the 2607 provided by inX() and outX() are the same as those provided by readX() 2639 and writeX() respectively when access 2608 and writeX() respectively when accessing a mapping with the default I/O 2640 attributes. 2609 attributes. 2641 2610 2642 Device drivers may expect outX() to e 2611 Device drivers may expect outX() to emit a non-posted write transaction 2643 that waits for a completion response 2612 that waits for a completion response from the I/O peripheral before 2644 returning. This is not guaranteed by 2613 returning. This is not guaranteed by all architectures and is therefore 2645 not part of the portable ordering sem 2614 not part of the portable ordering semantics. 2646 2615 2647 (*) insX(), outsX(): 2616 (*) insX(), outsX(): 2648 2617 2649 As above, the insX() and outsX() acce 2618 As above, the insX() and outsX() accessors provide the same ordering 2650 guarantees as readsX() and writesX() 2619 guarantees as readsX() and writesX() respectively when accessing a 2651 mapping with the default I/O attribut 2620 mapping with the default I/O attributes. 2652 2621 2653 (*) ioreadX(), iowriteX(): 2622 (*) ioreadX(), iowriteX(): 2654 2623 2655 These will perform appropriately for 2624 These will perform appropriately for the type of access they're actually 2656 doing, be it inX()/outX() or readX()/ 2625 doing, be it inX()/outX() or readX()/writeX(). 2657 2626 2658 With the exception of the string accessors (i 2627 With the exception of the string accessors (insX(), outsX(), readsX() and 2659 writesX()), all of the above assume that the 2628 writesX()), all of the above assume that the underlying peripheral is 2660 little-endian and will therefore perform byte 2629 little-endian and will therefore perform byte-swapping operations on big-endian 2661 architectures. 2630 architectures. 2662 2631 2663 2632 2664 ======================================== 2633 ======================================== 2665 ASSUMED MINIMUM EXECUTION ORDERING MODEL 2634 ASSUMED MINIMUM EXECUTION ORDERING MODEL 2666 ======================================== 2635 ======================================== 2667 2636 2668 It has to be assumed that the conceptual CPU 2637 It has to be assumed that the conceptual CPU is weakly-ordered but that it will 2669 maintain the appearance of program causality 2638 maintain the appearance of program causality with respect to itself. Some CPUs 2670 (such as i386 or x86_64) are more constrained 2639 (such as i386 or x86_64) are more constrained than others (such as powerpc or 2671 frv), and so the most relaxed case (namely DE 2640 frv), and so the most relaxed case (namely DEC Alpha) must be assumed outside 2672 of arch-specific code. 2641 of arch-specific code. 2673 2642 2674 This means that it must be considered that th 2643 This means that it must be considered that the CPU will execute its instruction 2675 stream in any order it feels like - or even i 2644 stream in any order it feels like - or even in parallel - provided that if an 2676 instruction in the stream depends on an earli 2645 instruction in the stream depends on an earlier instruction, then that 2677 earlier instruction must be sufficiently comp 2646 earlier instruction must be sufficiently complete[*] before the later 2678 instruction may proceed; in other words: prov 2647 instruction may proceed; in other words: provided that the appearance of 2679 causality is maintained. 2648 causality is maintained. 2680 2649 2681 [*] Some instructions have more than one eff 2650 [*] Some instructions have more than one effect - such as changing the 2682 condition codes, changing registers or c 2651 condition codes, changing registers or changing memory - and different 2683 instructions may depend on different eff 2652 instructions may depend on different effects. 2684 2653 2685 A CPU may also discard any instruction sequen 2654 A CPU may also discard any instruction sequence that winds up having no 2686 ultimate effect. For example, if two adjacen 2655 ultimate effect. For example, if two adjacent instructions both load an 2687 immediate value into the same register, the f 2656 immediate value into the same register, the first may be discarded. 2688 2657 2689 2658 2690 Similarly, it has to be assumed that compiler 2659 Similarly, it has to be assumed that compiler might reorder the instruction 2691 stream in any way it sees fit, again provided 2660 stream in any way it sees fit, again provided the appearance of causality is 2692 maintained. 2661 maintained. 2693 2662 2694 2663 2695 ============================ 2664 ============================ 2696 THE EFFECTS OF THE CPU CACHE 2665 THE EFFECTS OF THE CPU CACHE 2697 ============================ 2666 ============================ 2698 2667 2699 The way cached memory operations are perceive 2668 The way cached memory operations are perceived across the system is affected to 2700 a certain extent by the caches that lie betwe 2669 a certain extent by the caches that lie between CPUs and memory, and by the 2701 memory coherence system that maintains the co 2670 memory coherence system that maintains the consistency of state in the system. 2702 2671 2703 As far as the way a CPU interacts with anothe 2672 As far as the way a CPU interacts with another part of the system through the 2704 caches goes, the memory system has to include 2673 caches goes, the memory system has to include the CPU's caches, and memory 2705 barriers for the most part act at the interfa 2674 barriers for the most part act at the interface between the CPU and its cache 2706 (memory barriers logically act on the dotted 2675 (memory barriers logically act on the dotted line in the following diagram): 2707 2676 2708 <--- CPU ---> : <-- 2677 <--- CPU ---> : <----------- Memory -----------> 2709 : 2678 : 2710 +--------+ +--------+ : +------ 2679 +--------+ +--------+ : +--------+ +-----------+ 2711 | | | | : | 2680 | | | | : | | | | +--------+ 2712 | CPU | | Memory | : | CPU 2681 | CPU | | Memory | : | CPU | | | | | 2713 | Core |--->| Access |----->| Cache 2682 | Core |--->| Access |----->| Cache |<-->| | | | 2714 | | | Queue | : | 2683 | | | Queue | : | | | |--->| Memory | 2715 | | | | : | 2684 | | | | : | | | | | | 2716 +--------+ +--------+ : +------ 2685 +--------+ +--------+ : +--------+ | | | | 2717 : 2686 : | Cache | +--------+ 2718 : 2687 : | Coherency | 2719 : 2688 : | Mechanism | +--------+ 2720 +--------+ +--------+ : +------ 2689 +--------+ +--------+ : +--------+ | | | | 2721 | | | | : | 2690 | | | | : | | | | | | 2722 | CPU | | Memory | : | CPU 2691 | CPU | | Memory | : | CPU | | |--->| Device | 2723 | Core |--->| Access |----->| Cache 2692 | Core |--->| Access |----->| Cache |<-->| | | | 2724 | | | Queue | : | 2693 | | | Queue | : | | | | | | 2725 | | | | : | 2694 | | | | : | | | | +--------+ 2726 +--------+ +--------+ : +------ 2695 +--------+ +--------+ : +--------+ +-----------+ 2727 : 2696 : 2728 : 2697 : 2729 2698 2730 Although any particular load or store may not 2699 Although any particular load or store may not actually appear outside of the 2731 CPU that issued it since it may have been sat 2700 CPU that issued it since it may have been satisfied within the CPU's own cache, 2732 it will still appear as if the full memory ac 2701 it will still appear as if the full memory access had taken place as far as the 2733 other CPUs are concerned since the cache cohe 2702 other CPUs are concerned since the cache coherency mechanisms will migrate the 2734 cacheline over to the accessing CPU and propa 2703 cacheline over to the accessing CPU and propagate the effects upon conflict. 2735 2704 2736 The CPU core may execute instructions in any 2705 The CPU core may execute instructions in any order it deems fit, provided the 2737 expected program causality appears to be main 2706 expected program causality appears to be maintained. Some of the instructions 2738 generate load and store operations which then 2707 generate load and store operations which then go into the queue of memory 2739 accesses to be performed. The core may place 2708 accesses to be performed. The core may place these in the queue in any order 2740 it wishes, and continue execution until it is 2709 it wishes, and continue execution until it is forced to wait for an instruction 2741 to complete. 2710 to complete. 2742 2711 2743 What memory barriers are concerned with is co 2712 What memory barriers are concerned with is controlling the order in which 2744 accesses cross from the CPU side of things to 2713 accesses cross from the CPU side of things to the memory side of things, and 2745 the order in which the effects are perceived 2714 the order in which the effects are perceived to happen by the other observers 2746 in the system. 2715 in the system. 2747 2716 2748 [!] Memory barriers are _not_ needed within a 2717 [!] Memory barriers are _not_ needed within a given CPU, as CPUs always see 2749 their own loads and stores as if they had hap 2718 their own loads and stores as if they had happened in program order. 2750 2719 2751 [!] MMIO or other device accesses may bypass 2720 [!] MMIO or other device accesses may bypass the cache system. This depends on 2752 the properties of the memory window through w 2721 the properties of the memory window through which devices are accessed and/or 2753 the use of any special device communication i 2722 the use of any special device communication instructions the CPU may have. 2754 2723 2755 2724 2756 CACHE COHERENCY VS DMA 2725 CACHE COHERENCY VS DMA 2757 ---------------------- 2726 ---------------------- 2758 2727 2759 Not all systems maintain cache coherency with 2728 Not all systems maintain cache coherency with respect to devices doing DMA. In 2760 such cases, a device attempting DMA may obtai 2729 such cases, a device attempting DMA may obtain stale data from RAM because 2761 dirty cache lines may be resident in the cach 2730 dirty cache lines may be resident in the caches of various CPUs, and may not 2762 have been written back to RAM yet. To deal w 2731 have been written back to RAM yet. To deal with this, the appropriate part of 2763 the kernel must flush the overlapping bits of 2732 the kernel must flush the overlapping bits of cache on each CPU (and maybe 2764 invalidate them as well). 2733 invalidate them as well). 2765 2734 2766 In addition, the data DMA'd to RAM by a devic 2735 In addition, the data DMA'd to RAM by a device may be overwritten by dirty 2767 cache lines being written back to RAM from a 2736 cache lines being written back to RAM from a CPU's cache after the device has 2768 installed its own data, or cache lines presen 2737 installed its own data, or cache lines present in the CPU's cache may simply 2769 obscure the fact that RAM has been updated, u 2738 obscure the fact that RAM has been updated, until at such time as the cacheline 2770 is discarded from the CPU's cache and reloade 2739 is discarded from the CPU's cache and reloaded. To deal with this, the 2771 appropriate part of the kernel must invalidat 2740 appropriate part of the kernel must invalidate the overlapping bits of the 2772 cache on each CPU. 2741 cache on each CPU. 2773 2742 2774 See Documentation/core-api/cachetlb.rst for m !! 2743 See Documentation/core-api/cachetlb.rst for more information on cache management. 2775 management. << 2776 2744 2777 2745 2778 CACHE COHERENCY VS MMIO 2746 CACHE COHERENCY VS MMIO 2779 ----------------------- 2747 ----------------------- 2780 2748 2781 Memory mapped I/O usually takes place through 2749 Memory mapped I/O usually takes place through memory locations that are part of 2782 a window in the CPU's memory space that has d 2750 a window in the CPU's memory space that has different properties assigned than 2783 the usual RAM directed window. 2751 the usual RAM directed window. 2784 2752 2785 Amongst these properties is usually the fact 2753 Amongst these properties is usually the fact that such accesses bypass the 2786 caching entirely and go directly to the devic 2754 caching entirely and go directly to the device buses. This means MMIO accesses 2787 may, in effect, overtake accesses to cached m 2755 may, in effect, overtake accesses to cached memory that were emitted earlier. 2788 A memory barrier isn't sufficient in such a c 2756 A memory barrier isn't sufficient in such a case, but rather the cache must be 2789 flushed between the cached memory write and t 2757 flushed between the cached memory write and the MMIO access if the two are in 2790 any way dependent. 2758 any way dependent. 2791 2759 2792 2760 2793 ========================= 2761 ========================= 2794 THE THINGS CPUS GET UP TO 2762 THE THINGS CPUS GET UP TO 2795 ========================= 2763 ========================= 2796 2764 2797 A programmer might take it for granted that t 2765 A programmer might take it for granted that the CPU will perform memory 2798 operations in exactly the order specified, so 2766 operations in exactly the order specified, so that if the CPU is, for example, 2799 given the following piece of code to execute: 2767 given the following piece of code to execute: 2800 2768 2801 a = READ_ONCE(*A); 2769 a = READ_ONCE(*A); 2802 WRITE_ONCE(*B, b); 2770 WRITE_ONCE(*B, b); 2803 c = READ_ONCE(*C); 2771 c = READ_ONCE(*C); 2804 d = READ_ONCE(*D); 2772 d = READ_ONCE(*D); 2805 WRITE_ONCE(*E, e); 2773 WRITE_ONCE(*E, e); 2806 2774 2807 they would then expect that the CPU will comp 2775 they would then expect that the CPU will complete the memory operation for each 2808 instruction before moving on to the next one, 2776 instruction before moving on to the next one, leading to a definite sequence of 2809 operations as seen by external observers in t 2777 operations as seen by external observers in the system: 2810 2778 2811 LOAD *A, STORE *B, LOAD *C, LOAD *D, 2779 LOAD *A, STORE *B, LOAD *C, LOAD *D, STORE *E. 2812 2780 2813 2781 2814 Reality is, of course, much messier. With ma 2782 Reality is, of course, much messier. With many CPUs and compilers, the above 2815 assumption doesn't hold because: 2783 assumption doesn't hold because: 2816 2784 2817 (*) loads are more likely to need to be comp 2785 (*) loads are more likely to need to be completed immediately to permit 2818 execution progress, whereas stores can o 2786 execution progress, whereas stores can often be deferred without a 2819 problem; 2787 problem; 2820 2788 2821 (*) loads may be done speculatively, and the 2789 (*) loads may be done speculatively, and the result discarded should it prove 2822 to have been unnecessary; 2790 to have been unnecessary; 2823 2791 2824 (*) loads may be done speculatively, leading 2792 (*) loads may be done speculatively, leading to the result having been fetched 2825 at the wrong time in the expected sequen 2793 at the wrong time in the expected sequence of events; 2826 2794 2827 (*) the order of the memory accesses may be 2795 (*) the order of the memory accesses may be rearranged to promote better use 2828 of the CPU buses and caches; 2796 of the CPU buses and caches; 2829 2797 2830 (*) loads and stores may be combined to impr 2798 (*) loads and stores may be combined to improve performance when talking to 2831 memory or I/O hardware that can do batch 2799 memory or I/O hardware that can do batched accesses of adjacent locations, 2832 thus cutting down on transaction setup c 2800 thus cutting down on transaction setup costs (memory and PCI devices may 2833 both be able to do this); and 2801 both be able to do this); and 2834 2802 2835 (*) the CPU's data cache may affect the orde 2803 (*) the CPU's data cache may affect the ordering, and while cache-coherency 2836 mechanisms may alleviate this - once the 2804 mechanisms may alleviate this - once the store has actually hit the cache 2837 - there's no guarantee that the coherenc 2805 - there's no guarantee that the coherency management will be propagated in 2838 order to other CPUs. 2806 order to other CPUs. 2839 2807 2840 So what another CPU, say, might actually obse 2808 So what another CPU, say, might actually observe from the above piece of code 2841 is: 2809 is: 2842 2810 2843 LOAD *A, ..., LOAD {*C,*D}, STORE *E, 2811 LOAD *A, ..., LOAD {*C,*D}, STORE *E, STORE *B 2844 2812 2845 (Where "LOAD {*C,*D}" is a combined l 2813 (Where "LOAD {*C,*D}" is a combined load) 2846 2814 2847 2815 2848 However, it is guaranteed that a CPU will be 2816 However, it is guaranteed that a CPU will be self-consistent: it will see its 2849 _own_ accesses appear to be correctly ordered 2817 _own_ accesses appear to be correctly ordered, without the need for a memory 2850 barrier. For instance with the following cod 2818 barrier. For instance with the following code: 2851 2819 2852 U = READ_ONCE(*A); 2820 U = READ_ONCE(*A); 2853 WRITE_ONCE(*A, V); 2821 WRITE_ONCE(*A, V); 2854 WRITE_ONCE(*A, W); 2822 WRITE_ONCE(*A, W); 2855 X = READ_ONCE(*A); 2823 X = READ_ONCE(*A); 2856 WRITE_ONCE(*A, Y); 2824 WRITE_ONCE(*A, Y); 2857 Z = READ_ONCE(*A); 2825 Z = READ_ONCE(*A); 2858 2826 2859 and assuming no intervention by an external i 2827 and assuming no intervention by an external influence, it can be assumed that 2860 the final result will appear to be: 2828 the final result will appear to be: 2861 2829 2862 U == the original value of *A 2830 U == the original value of *A 2863 X == W 2831 X == W 2864 Z == Y 2832 Z == Y 2865 *A == Y 2833 *A == Y 2866 2834 2867 The code above may cause the CPU to generate 2835 The code above may cause the CPU to generate the full sequence of memory 2868 accesses: 2836 accesses: 2869 2837 2870 U=LOAD *A, STORE *A=V, STORE *A=W, X= 2838 U=LOAD *A, STORE *A=V, STORE *A=W, X=LOAD *A, STORE *A=Y, Z=LOAD *A 2871 2839 2872 in that order, but, without intervention, the 2840 in that order, but, without intervention, the sequence may have almost any 2873 combination of elements combined or discarded 2841 combination of elements combined or discarded, provided the program's view 2874 of the world remains consistent. Note that R 2842 of the world remains consistent. Note that READ_ONCE() and WRITE_ONCE() 2875 are -not- optional in the above example, as t 2843 are -not- optional in the above example, as there are architectures 2876 where a given CPU might reorder successive lo 2844 where a given CPU might reorder successive loads to the same location. 2877 On such architectures, READ_ONCE() and WRITE_ 2845 On such architectures, READ_ONCE() and WRITE_ONCE() do whatever is 2878 necessary to prevent this, for example, on It 2846 necessary to prevent this, for example, on Itanium the volatile casts 2879 used by READ_ONCE() and WRITE_ONCE() cause GC 2847 used by READ_ONCE() and WRITE_ONCE() cause GCC to emit the special ld.acq 2880 and st.rel instructions (respectively) that p 2848 and st.rel instructions (respectively) that prevent such reordering. 2881 2849 2882 The compiler may also combine, discard or def 2850 The compiler may also combine, discard or defer elements of the sequence before 2883 the CPU even sees them. 2851 the CPU even sees them. 2884 2852 2885 For instance: 2853 For instance: 2886 2854 2887 *A = V; 2855 *A = V; 2888 *A = W; 2856 *A = W; 2889 2857 2890 may be reduced to: 2858 may be reduced to: 2891 2859 2892 *A = W; 2860 *A = W; 2893 2861 2894 since, without either a write barrier or an W 2862 since, without either a write barrier or an WRITE_ONCE(), it can be 2895 assumed that the effect of the storage of V t 2863 assumed that the effect of the storage of V to *A is lost. Similarly: 2896 2864 2897 *A = Y; 2865 *A = Y; 2898 Z = *A; 2866 Z = *A; 2899 2867 2900 may, without a memory barrier or an READ_ONCE 2868 may, without a memory barrier or an READ_ONCE() and WRITE_ONCE(), be 2901 reduced to: 2869 reduced to: 2902 2870 2903 *A = Y; 2871 *A = Y; 2904 Z = Y; 2872 Z = Y; 2905 2873 2906 and the LOAD operation never appear outside o 2874 and the LOAD operation never appear outside of the CPU. 2907 2875 2908 2876 2909 AND THEN THERE'S THE ALPHA 2877 AND THEN THERE'S THE ALPHA 2910 -------------------------- 2878 -------------------------- 2911 2879 2912 The DEC Alpha CPU is one of the most relaxed 2880 The DEC Alpha CPU is one of the most relaxed CPUs there is. Not only that, 2913 some versions of the Alpha CPU have a split d 2881 some versions of the Alpha CPU have a split data cache, permitting them to have 2914 two semantically-related cache lines updated 2882 two semantically-related cache lines updated at separate times. This is where 2915 the address-dependency barrier really becomes !! 2883 the data dependency barrier really becomes necessary as this synchronises both 2916 both caches with the memory coherence system, !! 2884 caches with the memory coherence system, thus making it seem like pointer 2917 changes vs new data occur in the right order. 2885 changes vs new data occur in the right order. 2918 2886 2919 The Alpha defines the Linux kernel's memory m 2887 The Alpha defines the Linux kernel's memory model, although as of v4.15 2920 the Linux kernel's addition of smp_mb() to RE 2888 the Linux kernel's addition of smp_mb() to READ_ONCE() on Alpha greatly 2921 reduced its impact on the memory model. 2889 reduced its impact on the memory model. 2922 2890 2923 2891 2924 VIRTUAL MACHINE GUESTS 2892 VIRTUAL MACHINE GUESTS 2925 ---------------------- 2893 ---------------------- 2926 2894 2927 Guests running within virtual machines might 2895 Guests running within virtual machines might be affected by SMP effects even if 2928 the guest itself is compiled without SMP supp 2896 the guest itself is compiled without SMP support. This is an artifact of 2929 interfacing with an SMP host while running an 2897 interfacing with an SMP host while running an UP kernel. Using mandatory 2930 barriers for this use-case would be possible 2898 barriers for this use-case would be possible but is often suboptimal. 2931 2899 2932 To handle this case optimally, low-level virt 2900 To handle this case optimally, low-level virt_mb() etc macros are available. 2933 These have the same effect as smp_mb() etc wh 2901 These have the same effect as smp_mb() etc when SMP is enabled, but generate 2934 identical code for SMP and non-SMP systems. 2902 identical code for SMP and non-SMP systems. For example, virtual machine guests 2935 should use virt_mb() rather than smp_mb() whe 2903 should use virt_mb() rather than smp_mb() when synchronizing against a 2936 (possibly SMP) host. 2904 (possibly SMP) host. 2937 2905 2938 These are equivalent to smp_mb() etc counterp 2906 These are equivalent to smp_mb() etc counterparts in all other respects, 2939 in particular, they do not control MMIO effec 2907 in particular, they do not control MMIO effects: to control 2940 MMIO effects, use mandatory barriers. 2908 MMIO effects, use mandatory barriers. 2941 2909 2942 2910 2943 ============ 2911 ============ 2944 EXAMPLE USES 2912 EXAMPLE USES 2945 ============ 2913 ============ 2946 2914 2947 CIRCULAR BUFFERS 2915 CIRCULAR BUFFERS 2948 ---------------- 2916 ---------------- 2949 2917 2950 Memory barriers can be used to implement circ 2918 Memory barriers can be used to implement circular buffering without the need 2951 of a lock to serialise the producer with the 2919 of a lock to serialise the producer with the consumer. See: 2952 2920 2953 Documentation/core-api/circular-buffe 2921 Documentation/core-api/circular-buffers.rst 2954 2922 2955 for details. 2923 for details. 2956 2924 2957 2925 2958 ========== 2926 ========== 2959 REFERENCES 2927 REFERENCES 2960 ========== 2928 ========== 2961 2929 2962 Alpha AXP Architecture Reference Manual, Seco 2930 Alpha AXP Architecture Reference Manual, Second Edition (Sites & Witek, 2963 Digital Press) 2931 Digital Press) 2964 Chapter 5.2: Physical Address Space C 2932 Chapter 5.2: Physical Address Space Characteristics 2965 Chapter 5.4: Caches and Write Buffers 2933 Chapter 5.4: Caches and Write Buffers 2966 Chapter 5.5: Data Sharing 2934 Chapter 5.5: Data Sharing 2967 Chapter 5.6: Read/Write Ordering 2935 Chapter 5.6: Read/Write Ordering 2968 2936 2969 AMD64 Architecture Programmer's Manual Volume 2937 AMD64 Architecture Programmer's Manual Volume 2: System Programming 2970 Chapter 7.1: Memory-Access Ordering 2938 Chapter 7.1: Memory-Access Ordering 2971 Chapter 7.4: Buffering and Combining 2939 Chapter 7.4: Buffering and Combining Memory Writes 2972 2940 2973 ARM Architecture Reference Manual (ARMv8, for 2941 ARM Architecture Reference Manual (ARMv8, for ARMv8-A architecture profile) 2974 Chapter B2: The AArch64 Application L 2942 Chapter B2: The AArch64 Application Level Memory Model 2975 2943 2976 IA-32 Intel Architecture Software Developer's 2944 IA-32 Intel Architecture Software Developer's Manual, Volume 3: 2977 System Programming Guide 2945 System Programming Guide 2978 Chapter 7.1: Locked Atomic Operations 2946 Chapter 7.1: Locked Atomic Operations 2979 Chapter 7.2: Memory Ordering 2947 Chapter 7.2: Memory Ordering 2980 Chapter 7.4: Serializing Instructions 2948 Chapter 7.4: Serializing Instructions 2981 2949 2982 The SPARC Architecture Manual, Version 9 2950 The SPARC Architecture Manual, Version 9 2983 Chapter 8: Memory Models 2951 Chapter 8: Memory Models 2984 Appendix D: Formal Specification of t 2952 Appendix D: Formal Specification of the Memory Models 2985 Appendix J: Programming with the Memo 2953 Appendix J: Programming with the Memory Models 2986 2954 2987 Storage in the PowerPC (Stone and Fitzgerald) 2955 Storage in the PowerPC (Stone and Fitzgerald) 2988 2956 2989 UltraSPARC Programmer Reference Manual 2957 UltraSPARC Programmer Reference Manual 2990 Chapter 5: Memory Accesses and Cachea 2958 Chapter 5: Memory Accesses and Cacheability 2991 Chapter 15: Sparc-V9 Memory Models 2959 Chapter 15: Sparc-V9 Memory Models 2992 2960 2993 UltraSPARC III Cu User's Manual 2961 UltraSPARC III Cu User's Manual 2994 Chapter 9: Memory Models 2962 Chapter 9: Memory Models 2995 2963 2996 UltraSPARC IIIi Processor User's Manual 2964 UltraSPARC IIIi Processor User's Manual 2997 Chapter 8: Memory Models 2965 Chapter 8: Memory Models 2998 2966 2999 UltraSPARC Architecture 2005 2967 UltraSPARC Architecture 2005 3000 Chapter 9: Memory 2968 Chapter 9: Memory 3001 Appendix D: Formal Specifications of 2969 Appendix D: Formal Specifications of the Memory Models 3002 2970 3003 UltraSPARC T1 Supplement to the UltraSPARC Ar 2971 UltraSPARC T1 Supplement to the UltraSPARC Architecture 2005 3004 Chapter 8: Memory Models 2972 Chapter 8: Memory Models 3005 Appendix F: Caches and Cache Coherenc 2973 Appendix F: Caches and Cache Coherency 3006 2974 3007 Solaris Internals, Core Kernel Architecture, 2975 Solaris Internals, Core Kernel Architecture, p63-68: 3008 Chapter 3.3: Hardware Considerations 2976 Chapter 3.3: Hardware Considerations for Locks and 3009 Synchronization 2977 Synchronization 3010 2978 3011 Unix Systems for Modern Architectures, Symmet 2979 Unix Systems for Modern Architectures, Symmetric Multiprocessing and Caching 3012 for Kernel Programmers: 2980 for Kernel Programmers: 3013 Chapter 13: Other Memory Models 2981 Chapter 13: Other Memory Models 3014 2982 3015 Intel Itanium Architecture Software Developer 2983 Intel Itanium Architecture Software Developer's Manual: Volume 1: 3016 Section 2.6: Speculation 2984 Section 2.6: Speculation 3017 Section 4.4: Memory Access 2985 Section 4.4: Memory Access
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.