1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright (C) 2010-2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> 4 * 5 * membarrier system call 6 */ 7 8 /* 9 * For documentation purposes, here are some membarrier ordering 10 * scenarios to keep in mind: 11 * 12 * A) Userspace thread execution after IPI vs membarrier's memory 13 * barrier before sending the IPI 14 * 15 * Userspace variables: 16 * 17 * int x = 0, y = 0; 18 * 19 * The memory barrier at the start of membarrier() on CPU0 is necessary in 20 * order to enforce the guarantee that any writes occurring on CPU0 before 21 * the membarrier() is executed will be visible to any code executing on 22 * CPU1 after the IPI-induced memory barrier: 23 * 24 * CPU0 CPU1 25 * 26 * x = 1 27 * membarrier(): 28 * a: smp_mb() 29 * b: send IPI IPI-induced mb 30 * c: smp_mb() 31 * r2 = y 32 * y = 1 33 * barrier() 34 * r1 = x 35 * 36 * BUG_ON(r1 == 0 && r2 == 0) 37 * 38 * The write to y and load from x by CPU1 are unordered by the hardware, 39 * so it's possible to have "r1 = x" reordered before "y = 1" at any 40 * point after (b). If the memory barrier at (a) is omitted, then "x = 1" 41 * can be reordered after (a) (although not after (c)), so we get r1 == 0 42 * and r2 == 0. This violates the guarantee that membarrier() is 43 * supposed by provide. 44 * 45 * The timing of the memory barrier at (a) has to ensure that it executes 46 * before the IPI-induced memory barrier on CPU1. 47 * 48 * B) Userspace thread execution before IPI vs membarrier's memory 49 * barrier after completing the IPI 50 * 51 * Userspace variables: 52 * 53 * int x = 0, y = 0; 54 * 55 * The memory barrier at the end of membarrier() on CPU0 is necessary in 56 * order to enforce the guarantee that any writes occurring on CPU1 before 57 * the membarrier() is executed will be visible to any code executing on 58 * CPU0 after the membarrier(): 59 * 60 * CPU0 CPU1 61 * 62 * x = 1 63 * barrier() 64 * y = 1 65 * r2 = y 66 * membarrier(): 67 * a: smp_mb() 68 * b: send IPI IPI-induced mb 69 * c: smp_mb() 70 * r1 = x 71 * BUG_ON(r1 == 0 && r2 == 1) 72 * 73 * The writes to x and y are unordered by the hardware, so it's possible to 74 * have "r2 = 1" even though the write to x doesn't execute until (b). If 75 * the memory barrier at (c) is omitted then "r1 = x" can be reordered 76 * before (b) (although not before (a)), so we get "r1 = 0". This violates 77 * the guarantee that membarrier() is supposed to provide. 78 * 79 * The timing of the memory barrier at (c) has to ensure that it executes 80 * after the IPI-induced memory barrier on CPU1. 81 * 82 * C) Scheduling userspace thread -> kthread -> userspace thread vs membarrier 83 * 84 * CPU0 CPU1 85 * 86 * membarrier(): 87 * a: smp_mb() 88 * d: switch to kthread (includes mb) 89 * b: read rq->curr->mm == NULL 90 * e: switch to user (includes mb) 91 * c: smp_mb() 92 * 93 * Using the scenario from (A), we can show that (a) needs to be paired 94 * with (e). Using the scenario from (B), we can show that (c) needs to 95 * be paired with (d). 96 * 97 * D) exit_mm vs membarrier 98 * 99 * Two thread groups are created, A and B. Thread group B is created by 100 * issuing clone from group A with flag CLONE_VM set, but not CLONE_THREAD. 101 * Let's assume we have a single thread within each thread group (Thread A 102 * and Thread B). Thread A runs on CPU0, Thread B runs on CPU1. 103 * 104 * CPU0 CPU1 105 * 106 * membarrier(): 107 * a: smp_mb() 108 * exit_mm(): 109 * d: smp_mb() 110 * e: current->mm = NULL 111 * b: read rq->curr->mm == NULL 112 * c: smp_mb() 113 * 114 * Using scenario (B), we can show that (c) needs to be paired with (d). 115 * 116 * E) kthread_{use,unuse}_mm vs membarrier 117 * 118 * CPU0 CPU1 119 * 120 * membarrier(): 121 * a: smp_mb() 122 * kthread_unuse_mm() 123 * d: smp_mb() 124 * e: current->mm = NULL 125 * b: read rq->curr->mm == NULL 126 * kthread_use_mm() 127 * f: current->mm = mm 128 * g: smp_mb() 129 * c: smp_mb() 130 * 131 * Using the scenario from (A), we can show that (a) needs to be paired 132 * with (g). Using the scenario from (B), we can show that (c) needs to 133 * be paired with (d). 134 */ 135 136 /* 137 * Bitmask made from a "or" of all commands within enum membarrier_cmd, 138 * except MEMBARRIER_CMD_QUERY. 139 */ 140 #ifdef CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE 141 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK \ 142 (MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE \ 143 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE) 144 #else 145 #define MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK 0 146 #endif 147 148 #ifdef CONFIG_RSEQ 149 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK \ 150 (MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ \ 151 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ) 152 #else 153 #define MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK 0 154 #endif 155 156 #define MEMBARRIER_CMD_BITMASK \ 157 (MEMBARRIER_CMD_GLOBAL | MEMBARRIER_CMD_GLOBAL_EXPEDITED \ 158 | MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED \ 159 | MEMBARRIER_CMD_PRIVATE_EXPEDITED \ 160 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED \ 161 | MEMBARRIER_PRIVATE_EXPEDITED_SYNC_CORE_BITMASK \ 162 | MEMBARRIER_PRIVATE_EXPEDITED_RSEQ_BITMASK \ 163 | MEMBARRIER_CMD_GET_REGISTRATIONS) 164 165 static DEFINE_MUTEX(membarrier_ipi_mutex); 166 #define SERIALIZE_IPI() guard(mutex)(&membarrier_ipi_mutex) 167 168 static void ipi_mb(void *info) 169 { 170 smp_mb(); /* IPIs should be serializing but paranoid. */ 171 } 172 173 static void ipi_sync_core(void *info) 174 { 175 /* 176 * The smp_mb() in membarrier after all the IPIs is supposed to 177 * ensure that memory on remote CPUs that occur before the IPI 178 * become visible to membarrier()'s caller -- see scenario B in 179 * the big comment at the top of this file. 180 * 181 * A sync_core() would provide this guarantee, but 182 * sync_core_before_usermode() might end up being deferred until 183 * after membarrier()'s smp_mb(). 184 */ 185 smp_mb(); /* IPIs should be serializing but paranoid. */ 186 187 sync_core_before_usermode(); 188 } 189 190 static void ipi_rseq(void *info) 191 { 192 /* 193 * Ensure that all stores done by the calling thread are visible 194 * to the current task before the current task resumes. We could 195 * probably optimize this away on most architectures, but by the 196 * time we've already sent an IPI, the cost of the extra smp_mb() 197 * is negligible. 198 */ 199 smp_mb(); 200 rseq_preempt(current); 201 } 202 203 static void ipi_sync_rq_state(void *info) 204 { 205 struct mm_struct *mm = (struct mm_struct *) info; 206 207 if (current->mm != mm) 208 return; 209 this_cpu_write(runqueues.membarrier_state, 210 atomic_read(&mm->membarrier_state)); 211 /* 212 * Issue a memory barrier after setting 213 * MEMBARRIER_STATE_GLOBAL_EXPEDITED in the current runqueue to 214 * guarantee that no memory access following registration is reordered 215 * before registration. 216 */ 217 smp_mb(); 218 } 219 220 void membarrier_exec_mmap(struct mm_struct *mm) 221 { 222 /* 223 * Issue a memory barrier before clearing membarrier_state to 224 * guarantee that no memory access prior to exec is reordered after 225 * clearing this state. 226 */ 227 smp_mb(); 228 atomic_set(&mm->membarrier_state, 0); 229 /* 230 * Keep the runqueue membarrier_state in sync with this mm 231 * membarrier_state. 232 */ 233 this_cpu_write(runqueues.membarrier_state, 0); 234 } 235 236 void membarrier_update_current_mm(struct mm_struct *next_mm) 237 { 238 struct rq *rq = this_rq(); 239 int membarrier_state = 0; 240 241 if (next_mm) 242 membarrier_state = atomic_read(&next_mm->membarrier_state); 243 if (READ_ONCE(rq->membarrier_state) == membarrier_state) 244 return; 245 WRITE_ONCE(rq->membarrier_state, membarrier_state); 246 } 247 248 static int membarrier_global_expedited(void) 249 { 250 int cpu; 251 cpumask_var_t tmpmask; 252 253 if (num_online_cpus() == 1) 254 return 0; 255 256 /* 257 * Matches memory barriers around rq->curr modification in 258 * scheduler. 259 */ 260 smp_mb(); /* system call entry is not a mb. */ 261 262 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) 263 return -ENOMEM; 264 265 SERIALIZE_IPI(); 266 cpus_read_lock(); 267 rcu_read_lock(); 268 for_each_online_cpu(cpu) { 269 struct task_struct *p; 270 271 /* 272 * Skipping the current CPU is OK even through we can be 273 * migrated at any point. The current CPU, at the point 274 * where we read raw_smp_processor_id(), is ensured to 275 * be in program order with respect to the caller 276 * thread. Therefore, we can skip this CPU from the 277 * iteration. 278 */ 279 if (cpu == raw_smp_processor_id()) 280 continue; 281 282 if (!(READ_ONCE(cpu_rq(cpu)->membarrier_state) & 283 MEMBARRIER_STATE_GLOBAL_EXPEDITED)) 284 continue; 285 286 /* 287 * Skip the CPU if it runs a kernel thread which is not using 288 * a task mm. 289 */ 290 p = rcu_dereference(cpu_rq(cpu)->curr); 291 if (!p->mm) 292 continue; 293 294 __cpumask_set_cpu(cpu, tmpmask); 295 } 296 rcu_read_unlock(); 297 298 preempt_disable(); 299 smp_call_function_many(tmpmask, ipi_mb, NULL, 1); 300 preempt_enable(); 301 302 free_cpumask_var(tmpmask); 303 cpus_read_unlock(); 304 305 /* 306 * Memory barrier on the caller thread _after_ we finished 307 * waiting for the last IPI. Matches memory barriers around 308 * rq->curr modification in scheduler. 309 */ 310 smp_mb(); /* exit from system call is not a mb */ 311 return 0; 312 } 313 314 static int membarrier_private_expedited(int flags, int cpu_id) 315 { 316 cpumask_var_t tmpmask; 317 struct mm_struct *mm = current->mm; 318 smp_call_func_t ipi_func = ipi_mb; 319 320 if (flags == MEMBARRIER_FLAG_SYNC_CORE) { 321 if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE)) 322 return -EINVAL; 323 if (!(atomic_read(&mm->membarrier_state) & 324 MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY)) 325 return -EPERM; 326 ipi_func = ipi_sync_core; 327 } else if (flags == MEMBARRIER_FLAG_RSEQ) { 328 if (!IS_ENABLED(CONFIG_RSEQ)) 329 return -EINVAL; 330 if (!(atomic_read(&mm->membarrier_state) & 331 MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY)) 332 return -EPERM; 333 ipi_func = ipi_rseq; 334 } else { 335 WARN_ON_ONCE(flags); 336 if (!(atomic_read(&mm->membarrier_state) & 337 MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY)) 338 return -EPERM; 339 } 340 341 if (flags != MEMBARRIER_FLAG_SYNC_CORE && 342 (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1)) 343 return 0; 344 345 /* 346 * Matches memory barriers around rq->curr modification in 347 * scheduler. 348 */ 349 smp_mb(); /* system call entry is not a mb. */ 350 351 if (cpu_id < 0 && !zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) 352 return -ENOMEM; 353 354 SERIALIZE_IPI(); 355 cpus_read_lock(); 356 357 if (cpu_id >= 0) { 358 struct task_struct *p; 359 360 if (cpu_id >= nr_cpu_ids || !cpu_online(cpu_id)) 361 goto out; 362 rcu_read_lock(); 363 p = rcu_dereference(cpu_rq(cpu_id)->curr); 364 if (!p || p->mm != mm) { 365 rcu_read_unlock(); 366 goto out; 367 } 368 rcu_read_unlock(); 369 } else { 370 int cpu; 371 372 rcu_read_lock(); 373 for_each_online_cpu(cpu) { 374 struct task_struct *p; 375 376 p = rcu_dereference(cpu_rq(cpu)->curr); 377 if (p && p->mm == mm) 378 __cpumask_set_cpu(cpu, tmpmask); 379 } 380 rcu_read_unlock(); 381 } 382 383 if (cpu_id >= 0) { 384 /* 385 * smp_call_function_single() will call ipi_func() if cpu_id 386 * is the calling CPU. 387 */ 388 smp_call_function_single(cpu_id, ipi_func, NULL, 1); 389 } else { 390 /* 391 * For regular membarrier, we can save a few cycles by 392 * skipping the current cpu -- we're about to do smp_mb() 393 * below, and if we migrate to a different cpu, this cpu 394 * and the new cpu will execute a full barrier in the 395 * scheduler. 396 * 397 * For SYNC_CORE, we do need a barrier on the current cpu -- 398 * otherwise, if we are migrated and replaced by a different 399 * task in the same mm just before, during, or after 400 * membarrier, we will end up with some thread in the mm 401 * running without a core sync. 402 * 403 * For RSEQ, don't rseq_preempt() the caller. User code 404 * is not supposed to issue syscalls at all from inside an 405 * rseq critical section. 406 */ 407 if (flags != MEMBARRIER_FLAG_SYNC_CORE) { 408 preempt_disable(); 409 smp_call_function_many(tmpmask, ipi_func, NULL, true); 410 preempt_enable(); 411 } else { 412 on_each_cpu_mask(tmpmask, ipi_func, NULL, true); 413 } 414 } 415 416 out: 417 if (cpu_id < 0) 418 free_cpumask_var(tmpmask); 419 cpus_read_unlock(); 420 421 /* 422 * Memory barrier on the caller thread _after_ we finished 423 * waiting for the last IPI. Matches memory barriers around 424 * rq->curr modification in scheduler. 425 */ 426 smp_mb(); /* exit from system call is not a mb */ 427 428 return 0; 429 } 430 431 static int sync_runqueues_membarrier_state(struct mm_struct *mm) 432 { 433 int membarrier_state = atomic_read(&mm->membarrier_state); 434 cpumask_var_t tmpmask; 435 int cpu; 436 437 if (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1) { 438 this_cpu_write(runqueues.membarrier_state, membarrier_state); 439 440 /* 441 * For single mm user, we can simply issue a memory barrier 442 * after setting MEMBARRIER_STATE_GLOBAL_EXPEDITED in the 443 * mm and in the current runqueue to guarantee that no memory 444 * access following registration is reordered before 445 * registration. 446 */ 447 smp_mb(); 448 return 0; 449 } 450 451 if (!zalloc_cpumask_var(&tmpmask, GFP_KERNEL)) 452 return -ENOMEM; 453 454 /* 455 * For mm with multiple users, we need to ensure all future 456 * scheduler executions will observe @mm's new membarrier 457 * state. 458 */ 459 synchronize_rcu(); 460 461 /* 462 * For each cpu runqueue, if the task's mm match @mm, ensure that all 463 * @mm's membarrier state set bits are also set in the runqueue's 464 * membarrier state. This ensures that a runqueue scheduling 465 * between threads which are users of @mm has its membarrier state 466 * updated. 467 */ 468 SERIALIZE_IPI(); 469 cpus_read_lock(); 470 rcu_read_lock(); 471 for_each_online_cpu(cpu) { 472 struct rq *rq = cpu_rq(cpu); 473 struct task_struct *p; 474 475 p = rcu_dereference(rq->curr); 476 if (p && p->mm == mm) 477 __cpumask_set_cpu(cpu, tmpmask); 478 } 479 rcu_read_unlock(); 480 481 on_each_cpu_mask(tmpmask, ipi_sync_rq_state, mm, true); 482 483 free_cpumask_var(tmpmask); 484 cpus_read_unlock(); 485 486 return 0; 487 } 488 489 static int membarrier_register_global_expedited(void) 490 { 491 struct task_struct *p = current; 492 struct mm_struct *mm = p->mm; 493 int ret; 494 495 if (atomic_read(&mm->membarrier_state) & 496 MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY) 497 return 0; 498 atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED, &mm->membarrier_state); 499 ret = sync_runqueues_membarrier_state(mm); 500 if (ret) 501 return ret; 502 atomic_or(MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY, 503 &mm->membarrier_state); 504 505 return 0; 506 } 507 508 static int membarrier_register_private_expedited(int flags) 509 { 510 struct task_struct *p = current; 511 struct mm_struct *mm = p->mm; 512 int ready_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY, 513 set_state = MEMBARRIER_STATE_PRIVATE_EXPEDITED, 514 ret; 515 516 if (flags == MEMBARRIER_FLAG_SYNC_CORE) { 517 if (!IS_ENABLED(CONFIG_ARCH_HAS_MEMBARRIER_SYNC_CORE)) 518 return -EINVAL; 519 ready_state = 520 MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY; 521 } else if (flags == MEMBARRIER_FLAG_RSEQ) { 522 if (!IS_ENABLED(CONFIG_RSEQ)) 523 return -EINVAL; 524 ready_state = 525 MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY; 526 } else { 527 WARN_ON_ONCE(flags); 528 } 529 530 /* 531 * We need to consider threads belonging to different thread 532 * groups, which use the same mm. (CLONE_VM but not 533 * CLONE_THREAD). 534 */ 535 if ((atomic_read(&mm->membarrier_state) & ready_state) == ready_state) 536 return 0; 537 if (flags & MEMBARRIER_FLAG_SYNC_CORE) 538 set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE; 539 if (flags & MEMBARRIER_FLAG_RSEQ) 540 set_state |= MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ; 541 atomic_or(set_state, &mm->membarrier_state); 542 ret = sync_runqueues_membarrier_state(mm); 543 if (ret) 544 return ret; 545 atomic_or(ready_state, &mm->membarrier_state); 546 547 return 0; 548 } 549 550 static int membarrier_get_registrations(void) 551 { 552 struct task_struct *p = current; 553 struct mm_struct *mm = p->mm; 554 int registrations_mask = 0, membarrier_state, i; 555 static const int states[] = { 556 MEMBARRIER_STATE_GLOBAL_EXPEDITED | 557 MEMBARRIER_STATE_GLOBAL_EXPEDITED_READY, 558 MEMBARRIER_STATE_PRIVATE_EXPEDITED | 559 MEMBARRIER_STATE_PRIVATE_EXPEDITED_READY, 560 MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE | 561 MEMBARRIER_STATE_PRIVATE_EXPEDITED_SYNC_CORE_READY, 562 MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ | 563 MEMBARRIER_STATE_PRIVATE_EXPEDITED_RSEQ_READY 564 }; 565 static const int registration_cmds[] = { 566 MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED, 567 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 568 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE, 569 MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ 570 }; 571 BUILD_BUG_ON(ARRAY_SIZE(states) != ARRAY_SIZE(registration_cmds)); 572 573 membarrier_state = atomic_read(&mm->membarrier_state); 574 for (i = 0; i < ARRAY_SIZE(states); ++i) { 575 if (membarrier_state & states[i]) { 576 registrations_mask |= registration_cmds[i]; 577 membarrier_state &= ~states[i]; 578 } 579 } 580 WARN_ON_ONCE(membarrier_state != 0); 581 return registrations_mask; 582 } 583 584 /** 585 * sys_membarrier - issue memory barriers on a set of threads 586 * @cmd: Takes command values defined in enum membarrier_cmd. 587 * @flags: Currently needs to be 0 for all commands other than 588 * MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: in the latter 589 * case it can be MEMBARRIER_CMD_FLAG_CPU, indicating that @cpu_id 590 * contains the CPU on which to interrupt (= restart) 591 * the RSEQ critical section. 592 * @cpu_id: if @flags == MEMBARRIER_CMD_FLAG_CPU, indicates the cpu on which 593 * RSEQ CS should be interrupted (@cmd must be 594 * MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ). 595 * 596 * If this system call is not implemented, -ENOSYS is returned. If the 597 * command specified does not exist, not available on the running 598 * kernel, or if the command argument is invalid, this system call 599 * returns -EINVAL. For a given command, with flags argument set to 0, 600 * if this system call returns -ENOSYS or -EINVAL, it is guaranteed to 601 * always return the same value until reboot. In addition, it can return 602 * -ENOMEM if there is not enough memory available to perform the system 603 * call. 604 * 605 * All memory accesses performed in program order from each targeted thread 606 * is guaranteed to be ordered with respect to sys_membarrier(). If we use 607 * the semantic "barrier()" to represent a compiler barrier forcing memory 608 * accesses to be performed in program order across the barrier, and 609 * smp_mb() to represent explicit memory barriers forcing full memory 610 * ordering across the barrier, we have the following ordering table for 611 * each pair of barrier(), sys_membarrier() and smp_mb(): 612 * 613 * The pair ordering is detailed as (O: ordered, X: not ordered): 614 * 615 * barrier() smp_mb() sys_membarrier() 616 * barrier() X X O 617 * smp_mb() X O O 618 * sys_membarrier() O O O 619 */ 620 SYSCALL_DEFINE3(membarrier, int, cmd, unsigned int, flags, int, cpu_id) 621 { 622 switch (cmd) { 623 case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: 624 if (unlikely(flags && flags != MEMBARRIER_CMD_FLAG_CPU)) 625 return -EINVAL; 626 break; 627 default: 628 if (unlikely(flags)) 629 return -EINVAL; 630 } 631 632 if (!(flags & MEMBARRIER_CMD_FLAG_CPU)) 633 cpu_id = -1; 634 635 switch (cmd) { 636 case MEMBARRIER_CMD_QUERY: 637 { 638 int cmd_mask = MEMBARRIER_CMD_BITMASK; 639 640 if (tick_nohz_full_enabled()) 641 cmd_mask &= ~MEMBARRIER_CMD_GLOBAL; 642 return cmd_mask; 643 } 644 case MEMBARRIER_CMD_GLOBAL: 645 /* MEMBARRIER_CMD_GLOBAL is not compatible with nohz_full. */ 646 if (tick_nohz_full_enabled()) 647 return -EINVAL; 648 if (num_online_cpus() > 1) 649 synchronize_rcu(); 650 return 0; 651 case MEMBARRIER_CMD_GLOBAL_EXPEDITED: 652 return membarrier_global_expedited(); 653 case MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED: 654 return membarrier_register_global_expedited(); 655 case MEMBARRIER_CMD_PRIVATE_EXPEDITED: 656 return membarrier_private_expedited(0, cpu_id); 657 case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED: 658 return membarrier_register_private_expedited(0); 659 case MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE: 660 return membarrier_private_expedited(MEMBARRIER_FLAG_SYNC_CORE, cpu_id); 661 case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_SYNC_CORE: 662 return membarrier_register_private_expedited(MEMBARRIER_FLAG_SYNC_CORE); 663 case MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: 664 return membarrier_private_expedited(MEMBARRIER_FLAG_RSEQ, cpu_id); 665 case MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ: 666 return membarrier_register_private_expedited(MEMBARRIER_FLAG_RSEQ); 667 case MEMBARRIER_CMD_GET_REGISTRATIONS: 668 return membarrier_get_registrations(); 669 default: 670 return -EINVAL; 671 } 672 } 673