1 /* Macros for atomic functionality for tile. 2 Copyright (C) 2011-2020 Free Software Foundation, Inc. 3 Contributed by Walter Lee (walt@tilera.com) 4 5 This file is free software; you can redistribute it and/or modify it 6 under the terms of the GNU General Public License as published by the 7 Free Software Foundation; either version 3, or (at your option) any 8 later version. 9 10 This file is distributed in the hope that it will be useful, but 11 WITHOUT ANY WARRANTY; without even the implied warranty of 12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 13 General Public License for more details. 14 15 Under Section 7 of GPL version 3, you are granted additional 16 permissions described in the GCC Runtime Library Exception, version 17 3.1, as published by the Free Software Foundation. 18 19 You should have received a copy of the GNU General Public License and 20 a copy of the GCC Runtime Library Exception along with this program; 21 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 22 <http://www.gnu.org/licenses/>. */ 23 24 25 /* Provides macros for common atomic functionality. */ 26 27 #ifndef _ATOMIC_H_ 28 #define _ATOMIC_H_ 29 30 #ifdef __tilegx__ 31 /* Atomic instruction macros 32 33 The macros provided by atomic.h simplify access to the TILE-Gx 34 architecture's atomic instructions. The architecture provides a 35 variety of atomic instructions, including "exchange", "compare and 36 exchange", "fetch and ADD", "fetch and AND", "fetch and OR", and 37 "fetch and ADD if greater than or equal to zero". 38 39 No barrier or fence semantics are implied by any of the atomic 40 instructions for manipulating memory; you must specify the barriers 41 that you wish explicitly, using the provided macros. 42 43 Any integral 32- or 64-bit value can be used as the argument 44 to these macros, such as "int", "long long", "unsigned long", etc. 45 The pointers must be aligned to 4 or 8 bytes for 32- or 64-bit data. 46 The "exchange" and "compare and exchange" macros may also take 47 pointer values. We use the pseudo-type "VAL" in the documentation 48 to indicate the use of an appropriate type. */ 49 #else 50 /* Atomic instruction macros 51 52 The macros provided by atomic.h simplify access to the Tile 53 architecture's atomic instructions. Since the architecture 54 supports test-and-set as its only in-silicon atomic operation, many 55 of the operations provided by this header are implemented as 56 fast-path calls to Linux emulation routines. 57 58 Using the kernel for atomic operations allows userspace to take 59 advantage of the kernel's existing atomic-integer support (managed 60 by a distributed array of locks). The kernel provides proper 61 ordering among simultaneous atomic operations on different cores, 62 and guarantees a process cannot be context-switched part way 63 through an atomic operation. By virtue of sharing the kernel 64 atomic implementation, the userspace atomic operations 65 are compatible with the atomic methods provided by the kernel's 66 futex() syscall API. Note that these operations never cause Linux 67 kernel scheduling, and are in fact invisible to the kernel; they 68 simply act as regular function calls but with an elevated privilege 69 level. Note that the kernel's distributed lock array is hashed by 70 using only VA bits from the atomic value's address (to avoid the 71 performance hit of page table locking and multiple page-table 72 lookups to get the PA) and only the VA bits that are below page 73 granularity (to properly lock simultaneous accesses to the same 74 page mapped at different VAs). As a result, simultaneous atomic 75 operations on values whose addresses are at the same offset on a 76 page will contend in the kernel for the same lock array element. 77 78 No barrier or fence semantics are implied by any of the atomic 79 instructions for manipulating memory; you must specify the barriers 80 that you wish explicitly, using the provided macros. 81 82 Any integral 32- or 64-bit value can be used as the argument 83 to these macros, such as "int", "long long", "unsigned long", etc. 84 The pointers must be aligned to 4 or 8 bytes for 32- or 64-bit data. 85 The "exchange" and "compare and exchange" macros may also take 86 pointer values. We use the pseudo-type "VAL" in the documentation 87 to indicate the use of an appropriate type. 88 89 The 32-bit routines are implemented using a single kernel fast 90 syscall, as is the 64-bit compare-and-exchange. The other 64-bit 91 routines are implemented by looping over the 64-bit 92 compare-and-exchange routine, so may be potentially less efficient. */ 93 #endif 94 95 #ifdef __tilegx__ 96 #define SPR_CMPEXCH_VALUE 0x2780 97 #else 98 #define __NR_FAST_cmpxchg -1 99 #define __NR_FAST_atomic_update -2 100 #define __NR_FAST_cmpxchg64 -3 101 #endif 102 103 104 /* 32-bit integer compare-and-exchange. */ 105 static __inline __attribute__ ((always_inline)) arch_atomic_val_compare_and_exchange_4(volatile int * mem,int oldval,int newval)106 int arch_atomic_val_compare_and_exchange_4 (volatile int *mem, 107 int oldval, int newval) 108 { 109 #ifdef __tilegx__ 110 __insn_mtspr (SPR_CMPEXCH_VALUE, oldval); 111 return __insn_cmpexch4 (mem, newval); 112 #else 113 int result; 114 __asm__ __volatile__ ("swint1":"=R00" (result), 115 "=m" (*mem):"R10" (__NR_FAST_cmpxchg), "R00" (mem), 116 "R01" (oldval), "R02" (newval), "m" (*mem):"r20", 117 "r21", "r22", "r23", "r24", "r25", "r26", "r27", 118 "r28", "r29", "memory"); 119 return result; 120 #endif 121 } 122 123 /* 64-bit integer compare-and-exchange. */ 124 static __inline __attribute__ ((always_inline)) arch_atomic_val_compare_and_exchange_8(volatile long long * mem,long long oldval,long long newval)125 long long arch_atomic_val_compare_and_exchange_8 (volatile long long 126 *mem, long long oldval, 127 long long newval) 128 { 129 #ifdef __tilegx__ 130 __insn_mtspr (SPR_CMPEXCH_VALUE, oldval); 131 return __insn_cmpexch (mem, newval); 132 #else 133 unsigned int result_lo, result_hi; 134 unsigned int oldval_lo = oldval & 0xffffffffu, oldval_hi = oldval >> 32; 135 unsigned int newval_lo = newval & 0xffffffffu, newval_hi = newval >> 32; 136 __asm__ __volatile__ ("swint1":"=R00" (result_lo), "=R01" (result_hi), 137 "=m" (*mem):"R10" (__NR_FAST_cmpxchg64), "R00" (mem), 138 "R02" (oldval_lo), "R03" (oldval_hi), 139 "R04" (newval_lo), "R05" (newval_hi), 140 "m" (*mem):"r20", "r21", "r22", "r23", "r24", "r25", 141 "r26", "r27", "r28", "r29", "memory"); 142 return ((long long) result_hi) << 32 | result_lo; 143 #endif 144 } 145 146 /* This non-existent symbol is called for sizes other than "4" and "8", 147 indicating a bug in the caller. */ 148 extern int __arch_atomic_error_bad_argument_size (void) 149 __attribute__ ((warning ("sizeof atomic argument not 4 or 8"))); 150 151 152 #define arch_atomic_val_compare_and_exchange(mem, o, n) \ 153 __extension__ ({ \ 154 (__typeof(*(mem)))(__typeof(*(mem)-*(mem))) \ 155 ((sizeof(*(mem)) == 8) ? \ 156 arch_atomic_val_compare_and_exchange_8( \ 157 (volatile long long*)(mem), (__typeof((o)-(o)))(o), \ 158 (__typeof((n)-(n)))(n)) : \ 159 (sizeof(*(mem)) == 4) ? \ 160 arch_atomic_val_compare_and_exchange_4( \ 161 (volatile int*)(mem), (__typeof((o)-(o)))(o), \ 162 (__typeof((n)-(n)))(n)) : \ 163 __arch_atomic_error_bad_argument_size()); \ 164 }) 165 166 #define arch_atomic_bool_compare_and_exchange(mem, o, n) \ 167 __extension__ ({ \ 168 __typeof(o) __o = (o); \ 169 __builtin_expect( \ 170 __o == arch_atomic_val_compare_and_exchange((mem), __o, (n)), 1); \ 171 }) 172 173 174 /* Loop with compare_and_exchange until we guess the correct value. 175 Normally "expr" will be an expression using __old and __value. */ 176 #define __arch_atomic_update_cmpxchg(mem, value, expr) \ 177 __extension__ ({ \ 178 __typeof(value) __value = (value); \ 179 __typeof(*(mem)) *__mem = (mem), __old = *__mem, __guess; \ 180 do { \ 181 __guess = __old; \ 182 __old = arch_atomic_val_compare_and_exchange(__mem, __old, (expr)); \ 183 } while (__builtin_expect(__old != __guess, 0)); \ 184 __old; \ 185 }) 186 187 #ifdef __tilegx__ 188 189 /* Generic atomic op with 8- or 4-byte variant. 190 The _mask, _addend, and _expr arguments are ignored on tilegx. */ 191 #define __arch_atomic_update(mem, value, op, _mask, _addend, _expr) \ 192 __extension__ ({ \ 193 ((__typeof(*(mem))) \ 194 ((sizeof(*(mem)) == 8) ? (__typeof(*(mem)-*(mem)))__insn_##op( \ 195 (volatile void *)(mem), \ 196 (long long)(__typeof((value)-(value)))(value)) : \ 197 (sizeof(*(mem)) == 4) ? (int)__insn_##op##4( \ 198 (volatile void *)(mem), \ 199 (int)(__typeof((value)-(value)))(value)) : \ 200 __arch_atomic_error_bad_argument_size())); \ 201 }) 202 203 #else 204 205 /* This uses TILEPro's fast syscall support to atomically compute: 206 207 int old = *ptr; 208 *ptr = (old & mask) + addend; 209 return old; 210 211 This primitive can be used for atomic exchange, add, or, and. 212 Only 32-bit support is provided. */ 213 static __inline __attribute__ ((always_inline)) 214 int __arch_atomic_update_4(volatile int * mem,int mask,int addend)215 __arch_atomic_update_4 (volatile int *mem, int mask, int addend) 216 { 217 int result; 218 __asm__ __volatile__ ("swint1":"=R00" (result), 219 "=m" (*mem):"R10" (__NR_FAST_atomic_update), 220 "R00" (mem), "R01" (mask), "R02" (addend), 221 "m" (*mem):"r20", "r21", "r22", "r23", "r24", "r25", 222 "r26", "r27", "r28", "r29", "memory"); 223 return result; 224 } 225 226 /* Generic atomic op with 8- or 4-byte variant. 227 The _op argument is ignored on tilepro. */ 228 #define __arch_atomic_update(mem, value, _op, mask, addend, expr) \ 229 __extension__ ({ \ 230 (__typeof(*(mem)))(__typeof(*(mem)-*(mem))) \ 231 ((sizeof(*(mem)) == 8) ? \ 232 __arch_atomic_update_cmpxchg((mem), (value), (expr)) : \ 233 (sizeof(*(mem)) == 4) ? \ 234 __arch_atomic_update_4((volatile int*)(mem), \ 235 (__typeof((mask)-(mask)))(mask), \ 236 (__typeof((addend)-(addend)))(addend)) : \ 237 __arch_atomic_error_bad_argument_size()); \ 238 }) 239 240 #endif /* __tilegx__ */ 241 242 243 #define arch_atomic_exchange(mem, newvalue) \ 244 __arch_atomic_update(mem, newvalue, exch, 0, newvalue, __value) 245 246 #define arch_atomic_add(mem, value) \ 247 __arch_atomic_update(mem, value, fetchadd, -1, value, __old + __value) 248 249 #define arch_atomic_sub(mem, value) arch_atomic_add((mem), -(value)) 250 251 #define arch_atomic_increment(mem) arch_atomic_add((mem), 1) 252 253 #define arch_atomic_decrement(mem) arch_atomic_add((mem), -1) 254 255 #define arch_atomic_and(mem, mask) \ 256 __arch_atomic_update(mem, mask, fetchand, mask, 0, __old & __value) 257 258 #define arch_atomic_or(mem, mask) \ 259 __arch_atomic_update(mem, mask, fetchor, ~mask, mask, __old | __value) 260 261 #define arch_atomic_xor(mem, mask) \ 262 __arch_atomic_update_cmpxchg(mem, mask, __old ^ __value) 263 264 #define arch_atomic_nand(mem, mask) \ 265 __arch_atomic_update_cmpxchg(mem, mask, ~(__old & __value)) 266 267 #define arch_atomic_bit_set(mem, bit) \ 268 __extension__ ({ \ 269 __typeof(*(mem)) __mask = (__typeof(*(mem)))1 << (bit); \ 270 __mask & arch_atomic_or((mem), __mask); \ 271 }) 272 273 #define arch_atomic_bit_clear(mem, bit) \ 274 __extension__ ({ \ 275 __typeof(*(mem)) __mask = (__typeof(*(mem)))1 << (bit); \ 276 __mask & arch_atomic_and((mem), ~__mask); \ 277 }) 278 279 #ifdef __tilegx__ 280 /* Atomically store a new value to memory. 281 Note that you can freely use types of any size here, unlike the 282 other atomic routines, which require 32- or 64-bit types. 283 This accessor is provided for compatibility with TILEPro, which 284 required an explicit atomic operation for stores that needed 285 to be atomic with respect to other atomic methods in this header. */ 286 #define arch_atomic_write(mem, value) ((void) (*(mem) = (value))) 287 #else 288 #define arch_atomic_write(mem, value) \ 289 do { \ 290 __typeof(mem) __aw_mem = (mem); \ 291 __typeof(value) __aw_val = (value); \ 292 unsigned int *__aw_mem32, __aw_intval, __aw_val32, __aw_off, __aw_mask; \ 293 __aw_intval = (__typeof((value) - (value)))__aw_val; \ 294 switch (sizeof(*__aw_mem)) { \ 295 case 8: \ 296 __arch_atomic_update_cmpxchg(__aw_mem, __aw_val, __value); \ 297 break; \ 298 case 4: \ 299 __arch_atomic_update_4((int *)__aw_mem, 0, __aw_intval); \ 300 break; \ 301 case 2: \ 302 __aw_off = 8 * ((long)__aw_mem & 0x2); \ 303 __aw_mask = 0xffffU << __aw_off; \ 304 __aw_mem32 = (unsigned int *)((long)__aw_mem & ~0x2); \ 305 __aw_val32 = (__aw_intval << __aw_off) & __aw_mask; \ 306 __arch_atomic_update_cmpxchg(__aw_mem32, __aw_val32, \ 307 (__old & ~__aw_mask) | __value); \ 308 break; \ 309 case 1: \ 310 __aw_off = 8 * ((long)__aw_mem & 0x3); \ 311 __aw_mask = 0xffU << __aw_off; \ 312 __aw_mem32 = (unsigned int *)((long)__aw_mem & ~0x3); \ 313 __aw_val32 = (__aw_intval << __aw_off) & __aw_mask; \ 314 __arch_atomic_update_cmpxchg(__aw_mem32, __aw_val32, \ 315 (__old & ~__aw_mask) | __value); \ 316 break; \ 317 } \ 318 } while (0) 319 #endif 320 321 /* Compiler barrier. 322 323 This macro prevents loads or stores from being moved by the compiler 324 across the macro. Any loaded value that was loaded before this 325 macro must then be reloaded by the compiler. */ 326 #define arch_atomic_compiler_barrier() __asm__ __volatile__("" ::: "memory") 327 328 /* Full memory barrier. 329 330 This macro has the semantics of arch_atomic_compiler_barrer(), but also 331 ensures that previous stores are visible to other cores, and that 332 all previous loaded values have been placed into their target 333 register on this core. */ 334 #define arch_atomic_full_barrier() __insn_mf() 335 336 /* Read memory barrier. 337 338 Ensure that all reads by this processor that occurred prior to the 339 read memory barrier have completed, and that no reads that occur 340 after the read memory barrier on this processor are initiated 341 before the barrier. 342 343 On current TILE chips a read barrier is implemented as a full barrier, 344 but this may not be true in later versions of the architecture. 345 346 See also arch_atomic_acquire_barrier() for the appropriate idiom to use 347 to ensure no reads are lifted above an atomic lock instruction. */ 348 #define arch_atomic_read_barrier() arch_atomic_full_barrier() 349 350 /* Write memory barrier. 351 352 Ensure that all writes by this processor that occurred prior to the 353 write memory barrier have completed, and that no writes that occur 354 after the write memory barrier on this processor are initiated 355 before the barrier. 356 357 On current TILE chips a write barrier is implemented as a full barrier, 358 but this may not be true in later versions of the architecture. 359 360 See also arch_atomic_release_barrier() for the appropriate idiom to use 361 to ensure all writes are complete prior to an atomic unlock instruction. */ 362 #define arch_atomic_write_barrier() arch_atomic_full_barrier() 363 364 /* Lock acquisition barrier. 365 366 Ensure that no load operations that follow this macro in the 367 program can issue prior to the barrier. Without such a barrier, 368 the compiler can reorder them to issue earlier, or the hardware can 369 issue them speculatively. The latter is not currently done in the 370 Tile microarchitecture, but using this operation improves 371 portability to future implementations. 372 373 This operation is intended to be used as part of the "acquire" 374 path for locking, that is, when entering a critical section. 375 This should be done after the atomic operation that actually 376 acquires the lock, and in conjunction with a "control dependency" 377 that checks the atomic operation result to see if the lock was 378 in fact acquired. See the arch_atomic_read_barrier() macro 379 for a heavier-weight barrier to use in certain unusual constructs, 380 or arch_atomic_acquire_barrier_value() if no control dependency exists. */ 381 #define arch_atomic_acquire_barrier() arch_atomic_compiler_barrier() 382 383 /* Lock release barrier. 384 385 Ensure that no store operations that precede this macro in the 386 program complete subsequent to the barrier. Without such a 387 barrier, the compiler can reorder stores to issue later, or stores 388 can be still outstanding in the memory network. 389 390 This operation is intended to be used as part of the "release" path 391 for locking, that is, when leaving a critical section. This should 392 be done before the operation (such as a store of zero) that 393 actually releases the lock. */ 394 #define arch_atomic_release_barrier() arch_atomic_write_barrier() 395 396 /* Barrier until the read of a particular value is complete. 397 398 This is occasionally useful when constructing certain locking 399 scenarios. For example, you might write a routine that issues an 400 atomic instruction to enter a critical section, then reads one or 401 more values within the critical section without checking to see if 402 the critical section was in fact acquired, and only later checks 403 the atomic instruction result to see if the lock was acquired. If 404 so the routine could properly release the lock and know that the 405 values that were read were valid. 406 407 In this scenario, it is required to wait for the result of the 408 atomic instruction, even if the value itself is not checked. This 409 guarantees that if the atomic instruction succeeded in taking the lock, 410 the lock was held before any reads in the critical section issued. */ 411 #define arch_atomic_acquire_barrier_value(val) \ 412 __asm__ __volatile__("move %0, %0" :: "r"(val)) 413 414 /* Access the given variable in memory exactly once. 415 416 In some contexts, an algorithm may need to force access to memory, 417 since otherwise the compiler may think it can optimize away a 418 memory load or store; for example, in a loop when polling memory to 419 see if another cpu has updated it yet. Generally this is only 420 required for certain very carefully hand-tuned algorithms; using it 421 unnecessarily may result in performance losses. 422 423 A related use of this macro is to ensure that the compiler does not 424 rematerialize the value of "x" by reloading it from memory 425 unexpectedly; the "volatile" marking will prevent the compiler from 426 being able to rematerialize. This is helpful if an algorithm needs 427 to read a variable without locking, but needs it to have the same 428 value if it ends up being used several times within the algorithm. 429 430 Note that multiple uses of this macro are guaranteed to be ordered, 431 i.e. the compiler will not reorder stores or loads that are wrapped 432 in arch_atomic_access_once(). */ 433 #define arch_atomic_access_once(x) (*(volatile __typeof(x) *)&(x)) 434 435 436 437 #endif /* !_ATOMIC_H_ */ 438