git.sesse.net Git - bcachefs-tools-debian/blob - linux/six.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include <linux/export.h>
   4 #include <linux/log2.h>
   5 #include <linux/percpu.h>
   6 #include <linux/preempt.h>
   7 #include <linux/rcupdate.h>
   8 #include <linux/sched.h>
   9 #include <linux/sched/clock.h>
  10 #include <linux/sched/rt.h>
  11 #include <linux/six.h>
  12 #include <linux/slab.h>
  13
  14 #include <trace/events/lock.h>
  15
  16 #ifdef DEBUG
  17 #define EBUG_ON(cond)           BUG_ON(cond)
  18 #else
  19 #define EBUG_ON(cond)           do {} while (0)
  20 #endif
  21
  22 #define six_acquire(l, t, r, ip)        lock_acquire(l, 0, t, r, 1, NULL, ip)
  23 #define six_release(l, ip)              lock_release(l, ip)
  24
  25 static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type);
  26
  27 struct six_lock_vals {
  28         /* Value we add to the lock in order to take the lock: */
  29         u64                     lock_val;
  30
  31         /* If the lock has this value (used as a mask), taking the lock fails: */
  32         u64                     lock_fail;
  33
  34         /* Value we add to the lock in order to release the lock: */
  35         u64                     unlock_val;
  36
  37         /* Mask that indicates lock is held for this type: */
  38         u64                     held_mask;
  39
  40         /* Waitlist we wakeup when releasing the lock: */
  41         enum six_lock_type      unlock_wakeup;
  42 };
  43
  44 #define __SIX_LOCK_HELD_read    __SIX_VAL(read_lock, ~0)
  45 #define __SIX_LOCK_HELD_intent  __SIX_VAL(intent_lock, ~0)
  46 #define __SIX_LOCK_HELD_write   __SIX_VAL(seq, 1)
  47
  48 #define LOCK_VALS {                                                     \
  49         [SIX_LOCK_read] = {                                             \
  50                 .lock_val       = __SIX_VAL(read_lock, 1),              \
  51                 .lock_fail      = __SIX_LOCK_HELD_write + __SIX_VAL(write_locking, 1),\
  52                 .unlock_val     = -__SIX_VAL(read_lock, 1),             \
  53                 .held_mask      = __SIX_LOCK_HELD_read,                 \
  54                 .unlock_wakeup  = SIX_LOCK_write,                       \
  55         },                                                              \
  56         [SIX_LOCK_intent] = {                                           \
  57                 .lock_val       = __SIX_VAL(intent_lock, 1),            \
  58                 .lock_fail      = __SIX_LOCK_HELD_intent,               \
  59                 .unlock_val     = -__SIX_VAL(intent_lock, 1),           \
  60                 .held_mask      = __SIX_LOCK_HELD_intent,               \
  61                 .unlock_wakeup  = SIX_LOCK_intent,                      \
  62         },                                                              \
  63         [SIX_LOCK_write] = {                                            \
  64                 .lock_val       = __SIX_VAL(seq, 1),                    \
  65                 .lock_fail      = __SIX_LOCK_HELD_read,                 \
  66                 .unlock_val     = __SIX_VAL(seq, 1),                    \
  67                 .held_mask      = __SIX_LOCK_HELD_write,                \
  68                 .unlock_wakeup  = SIX_LOCK_read,                        \
  69         },                                                              \
  70 }
  71
  72 static inline void six_set_owner(struct six_lock *lock, enum six_lock_type type,
  73                                  union six_lock_state old,
  74                                  struct task_struct *owner)
  75 {
  76         if (type != SIX_LOCK_intent)
  77                 return;
  78
  79         if (!old.intent_lock) {
  80                 EBUG_ON(lock->owner);
  81                 lock->owner = owner;
  82         } else {
  83                 EBUG_ON(lock->owner != current);
  84         }
  85 }
  86
  87 static inline unsigned pcpu_read_count(struct six_lock *lock)
  88 {
  89         unsigned read_count = 0;
  90         int cpu;
  91
  92         for_each_possible_cpu(cpu)
  93                 read_count += *per_cpu_ptr(lock->readers, cpu);
  94         return read_count;
  95 }
  96
  97 /* This is probably up there with the more evil things I've done */
  98 #define waitlist_bitnr(id) ilog2((((union six_lock_state) { .waiters = 1 << (id) }).l))
  99
 100 static int __do_six_trylock_type(struct six_lock *lock,
 101                                  enum six_lock_type type,
 102                                  struct task_struct *task,
 103                                  bool try)
 104 {
 105         const struct six_lock_vals l[] = LOCK_VALS;
 106         union six_lock_state old, new;
 107         int ret;
 108         u64 v;
 109
 110         EBUG_ON(type == SIX_LOCK_write && lock->owner != task);
 111         EBUG_ON(type == SIX_LOCK_write && (lock->state.seq & 1));
 112         EBUG_ON(type == SIX_LOCK_write && (try != !(lock->state.write_locking)));
 113
 114         /*
 115          * Percpu reader mode:
 116          *
 117          * The basic idea behind this algorithm is that you can implement a lock
 118          * between two threads without any atomics, just memory barriers:
 119          *
 120          * For two threads you'll need two variables, one variable for "thread a
 121          * has the lock" and another for "thread b has the lock".
 122          *
 123          * To take the lock, a thread sets its variable indicating that it holds
 124          * the lock, then issues a full memory barrier, then reads from the
 125          * other thread's variable to check if the other thread thinks it has
 126          * the lock. If we raced, we backoff and retry/sleep.
 127          */
 128
 129         if (type == SIX_LOCK_read && lock->readers) {
 130                 preempt_disable();
 131                 this_cpu_inc(*lock->readers); /* signal that we own lock */
 132
 133                 smp_mb();
 134
 135                 old.v = READ_ONCE(lock->state.v);
 136                 ret = !(old.v & l[type].lock_fail);
 137
 138                 this_cpu_sub(*lock->readers, !ret);
 139                 preempt_enable();
 140
 141                 /*
 142                  * If we failed because a writer was trying to take the
 143                  * lock, issue a wakeup because we might have caused a
 144                  * spurious trylock failure:
 145                  */
 146                 if (old.write_locking)
 147                         ret = -1 - SIX_LOCK_write;
 148         } else if (type == SIX_LOCK_write && lock->readers) {
 149                 if (try) {
 150                         atomic64_add(__SIX_VAL(write_locking, 1),
 151                                      &lock->state.counter);
 152                         smp_mb__after_atomic();
 153                 } else if (!(lock->state.waiters & (1 << SIX_LOCK_write))) {
 154                         atomic64_add(__SIX_VAL(waiters, 1 << SIX_LOCK_write),
 155                                      &lock->state.counter);
 156                         /*
 157                          * pairs with barrier after unlock and before checking
 158                          * for readers in unlock path
 159                          */
 160                         smp_mb__after_atomic();
 161                 }
 162
 163                 ret = !pcpu_read_count(lock);
 164
 165                 /*
 166                  * On success, we increment lock->seq; also we clear
 167                  * write_locking unless we failed from the lock path:
 168                  */
 169                 v = 0;
 170                 if (ret)
 171                         v += __SIX_VAL(seq, 1);
 172                 if (ret || try)
 173                         v -= __SIX_VAL(write_locking, 1);
 174
 175                 if (try && !ret) {
 176                         old.v = atomic64_add_return(v, &lock->state.counter);
 177                         if (old.waiters & (1 << SIX_LOCK_read))
 178                                 ret = -1 - SIX_LOCK_read;
 179                 } else {
 180                         atomic64_add(v, &lock->state.counter);
 181                 }
 182         } else {
 183                 v = READ_ONCE(lock->state.v);
 184                 do {
 185                         new.v = old.v = v;
 186
 187                         if (!(old.v & l[type].lock_fail)) {
 188                                 new.v += l[type].lock_val;
 189
 190                                 if (type == SIX_LOCK_write)
 191                                         new.write_locking = 0;
 192                         } else if (!try && !(new.waiters & (1 << type)))
 193                                 new.waiters |= 1 << type;
 194                         else
 195                                 break; /* waiting bit already set */
 196                 } while ((v = atomic64_cmpxchg_acquire(&lock->state.counter,
 197                                         old.v, new.v)) != old.v);
 198
 199                 ret = !(old.v & l[type].lock_fail);
 200
 201                 EBUG_ON(ret && !(lock->state.v & l[type].held_mask));
 202         }
 203
 204         if (ret > 0)
 205                 six_set_owner(lock, type, old, task);
 206
 207         EBUG_ON(type == SIX_LOCK_write && (try || ret > 0) && (lock->state.write_locking));
 208
 209         return ret;
 210 }
 211
 212 static inline void __six_lock_wakeup(struct six_lock *lock, enum six_lock_type lock_type)
 213 {
 214         struct six_lock_waiter *w, *next;
 215         struct task_struct *task;
 216         bool saw_one;
 217         int ret;
 218 again:
 219         ret = 0;
 220         saw_one = false;
 221         raw_spin_lock(&lock->wait_lock);
 222
 223         list_for_each_entry_safe(w, next, &lock->wait_list, list) {
 224                 if (w->lock_want != lock_type)
 225                         continue;
 226
 227                 if (saw_one && lock_type != SIX_LOCK_read)
 228                         goto unlock;
 229                 saw_one = true;
 230
 231                 ret = __do_six_trylock_type(lock, lock_type, w->task, false);
 232                 if (ret <= 0)
 233                         goto unlock;
 234
 235                 __list_del(w->list.prev, w->list.next);
 236                 task = w->task;
 237                 /*
 238                  * Do no writes to @w besides setting lock_acquired - otherwise
 239                  * we would need a memory barrier:
 240                  */
 241                 barrier();
 242                 w->lock_acquired = true;
 243                 wake_up_process(task);
 244         }
 245
 246         clear_bit(waitlist_bitnr(lock_type), (unsigned long *) &lock->state.v);
 247 unlock:
 248         raw_spin_unlock(&lock->wait_lock);
 249
 250         if (ret < 0) {
 251                 lock_type = -ret - 1;
 252                 goto again;
 253         }
 254 }
 255
 256 static inline void six_lock_wakeup(struct six_lock *lock,
 257                                    union six_lock_state state,
 258                                    enum six_lock_type lock_type)
 259 {
 260         if (lock_type == SIX_LOCK_write && state.read_lock)
 261                 return;
 262
 263         if (!(state.waiters & (1 << lock_type)))
 264                 return;
 265
 266         __six_lock_wakeup(lock, lock_type);
 267 }
 268
 269 static bool do_six_trylock_type(struct six_lock *lock,
 270                                 enum six_lock_type type,
 271                                 bool try)
 272 {
 273         int ret;
 274
 275         ret = __do_six_trylock_type(lock, type, current, try);
 276         if (ret < 0)
 277                 __six_lock_wakeup(lock, -ret - 1);
 278
 279         return ret > 0;
 280 }
 281
 282 __always_inline __flatten
 283 static bool __six_trylock_type(struct six_lock *lock, enum six_lock_type type,
 284                                unsigned long ip)
 285 {
 286         if (!do_six_trylock_type(lock, type, true))
 287                 return false;
 288
 289         if (type != SIX_LOCK_write)
 290                 six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
 291         return true;
 292 }
 293
 294 __always_inline __flatten
 295 static bool __six_relock_type(struct six_lock *lock, enum six_lock_type type,
 296                               unsigned seq, unsigned long ip)
 297 {
 298         const struct six_lock_vals l[] = LOCK_VALS;
 299         union six_lock_state old;
 300         u64 v;
 301
 302         EBUG_ON(type == SIX_LOCK_write);
 303
 304         if (type == SIX_LOCK_read &&
 305             lock->readers) {
 306                 bool ret;
 307
 308                 preempt_disable();
 309                 this_cpu_inc(*lock->readers);
 310
 311                 smp_mb();
 312
 313                 old.v = READ_ONCE(lock->state.v);
 314                 ret = !(old.v & l[type].lock_fail) && old.seq == seq;
 315
 316                 this_cpu_sub(*lock->readers, !ret);
 317                 preempt_enable();
 318
 319                 /*
 320                  * Similar to the lock path, we may have caused a spurious write
 321                  * lock fail and need to issue a wakeup:
 322                  */
 323                 if (old.write_locking)
 324                         six_lock_wakeup(lock, old, SIX_LOCK_write);
 325
 326                 if (ret)
 327                         six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
 328
 329                 return ret;
 330         }
 331
 332         v = READ_ONCE(lock->state.v);
 333         do {
 334                 old.v = v;
 335
 336                 if (old.seq != seq || old.v & l[type].lock_fail)
 337                         return false;
 338         } while ((v = atomic64_cmpxchg_acquire(&lock->state.counter,
 339                                 old.v,
 340                                 old.v + l[type].lock_val)) != old.v);
 341
 342         six_set_owner(lock, type, old, current);
 343         if (type != SIX_LOCK_write)
 344                 six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
 345         return true;
 346 }
 347
 348 #ifdef CONFIG_LOCK_SPIN_ON_OWNER
 349
 350 static inline bool six_can_spin_on_owner(struct six_lock *lock)
 351 {
 352         struct task_struct *owner;
 353         bool ret;
 354
 355         if (need_resched())
 356                 return false;
 357
 358         rcu_read_lock();
 359         owner = READ_ONCE(lock->owner);
 360         ret = !owner || owner_on_cpu(owner);
 361         rcu_read_unlock();
 362
 363         return ret;
 364 }
 365
 366 static inline void six_set_nospin(struct six_lock *lock)
 367 {
 368         union six_lock_state old, new;
 369         u64 v = READ_ONCE(lock->state.v);
 370
 371         do {
 372                 new.v = old.v = v;
 373                 new.nospin = true;
 374         } while ((v = atomic64_cmpxchg(&lock->state.counter, old.v, new.v)) != old.v);
 375 }
 376
 377 static inline bool six_spin_on_owner(struct six_lock *lock,
 378                                      struct task_struct *owner,
 379                                      u64 end_time)
 380 {
 381         bool ret = true;
 382         unsigned loop = 0;
 383
 384         rcu_read_lock();
 385         while (lock->owner == owner) {
 386                 /*
 387                  * Ensure we emit the owner->on_cpu, dereference _after_
 388                  * checking lock->owner still matches owner. If that fails,
 389                  * owner might point to freed memory. If it still matches,
 390                  * the rcu_read_lock() ensures the memory stays valid.
 391                  */
 392                 barrier();
 393
 394                 if (!owner_on_cpu(owner) || need_resched()) {
 395                         ret = false;
 396                         break;
 397                 }
 398
 399                 if (!(++loop & 0xf) && (time_after64(sched_clock(), end_time))) {
 400                         six_set_nospin(lock);
 401                         ret = false;
 402                         break;
 403                 }
 404
 405                 cpu_relax();
 406         }
 407         rcu_read_unlock();
 408
 409         return ret;
 410 }
 411
 412 static inline bool six_optimistic_spin(struct six_lock *lock, enum six_lock_type type)
 413 {
 414         struct task_struct *task = current;
 415         u64 end_time;
 416
 417         if (type == SIX_LOCK_write)
 418                 return false;
 419
 420         preempt_disable();
 421         if (!six_can_spin_on_owner(lock))
 422                 goto fail;
 423
 424         if (!osq_lock(&lock->osq))
 425                 goto fail;
 426
 427         end_time = sched_clock() + 10 * NSEC_PER_USEC;
 428
 429         while (1) {
 430                 struct task_struct *owner;
 431
 432                 /*
 433                  * If there's an owner, wait for it to either
 434                  * release the lock or go to sleep.
 435                  */
 436                 owner = READ_ONCE(lock->owner);
 437                 if (owner && !six_spin_on_owner(lock, owner, end_time))
 438                         break;
 439
 440                 if (do_six_trylock_type(lock, type, false)) {
 441                         osq_unlock(&lock->osq);
 442                         preempt_enable();
 443                         return true;
 444                 }
 445
 446                 /*
 447                  * When there's no owner, we might have preempted between the
 448                  * owner acquiring the lock and setting the owner field. If
 449                  * we're an RT task that will live-lock because we won't let
 450                  * the owner complete.
 451                  */
 452                 if (!owner && (need_resched() || rt_task(task)))
 453                         break;
 454
 455                 /*
 456                  * The cpu_relax() call is a compiler barrier which forces
 457                  * everything in this loop to be re-loaded. We don't need
 458                  * memory barriers as we'll eventually observe the right
 459                  * values at the cost of a few extra spins.
 460                  */
 461                 cpu_relax();
 462         }
 463
 464         osq_unlock(&lock->osq);
 465 fail:
 466         preempt_enable();
 467
 468         /*
 469          * If we fell out of the spin path because of need_resched(),
 470          * reschedule now, before we try-lock again. This avoids getting
 471          * scheduled out right after we obtained the lock.
 472          */
 473         if (need_resched())
 474                 schedule();
 475
 476         return false;
 477 }
 478
 479 #else /* CONFIG_LOCK_SPIN_ON_OWNER */
 480
 481 static inline bool six_optimistic_spin(struct six_lock *lock, enum six_lock_type type)
 482 {
 483         return false;
 484 }
 485
 486 #endif
 487
 488 noinline
 489 static int __six_lock_type_slowpath(struct six_lock *lock, enum six_lock_type type,
 490                                     struct six_lock_waiter *wait,
 491                                     six_lock_should_sleep_fn should_sleep_fn, void *p,
 492                                     unsigned long ip)
 493 {
 494         union six_lock_state old;
 495         int ret = 0;
 496
 497         if (type == SIX_LOCK_write) {
 498                 EBUG_ON(lock->state.write_locking);
 499                 atomic64_add(__SIX_VAL(write_locking, 1), &lock->state.counter);
 500                 smp_mb__after_atomic();
 501         }
 502
 503         trace_contention_begin(lock, 0);
 504         lock_contended(&lock->dep_map, ip);
 505
 506         if (six_optimistic_spin(lock, type))
 507                 goto out;
 508
 509         wait->task              = current;
 510         wait->lock_want         = type;
 511         wait->lock_acquired     = false;
 512
 513         raw_spin_lock(&lock->wait_lock);
 514         if (!(lock->state.waiters & (1 << type)))
 515                 set_bit(waitlist_bitnr(type), (unsigned long *) &lock->state.v);
 516         /*
 517          * Retry taking the lock after taking waitlist lock, have raced with an
 518          * unlock:
 519          */
 520         ret = __do_six_trylock_type(lock, type, current, false);
 521         if (ret <= 0) {
 522                 wait->start_time = local_clock();
 523
 524                 if (!list_empty(&lock->wait_list)) {
 525                         struct six_lock_waiter *last =
 526                                 list_last_entry(&lock->wait_list,
 527                                         struct six_lock_waiter, list);
 528
 529                         if (time_before_eq64(wait->start_time, last->start_time))
 530                                 wait->start_time = last->start_time + 1;
 531                 }
 532
 533                 list_add_tail(&wait->list, &lock->wait_list);
 534         }
 535         raw_spin_unlock(&lock->wait_lock);
 536
 537         if (unlikely(ret > 0)) {
 538                 ret = 0;
 539                 goto out;
 540         }
 541
 542         if (unlikely(ret < 0)) {
 543                 __six_lock_wakeup(lock, -ret - 1);
 544                 ret = 0;
 545         }
 546
 547         while (1) {
 548                 set_current_state(TASK_UNINTERRUPTIBLE);
 549
 550                 if (wait->lock_acquired)
 551                         break;
 552
 553                 ret = should_sleep_fn ? should_sleep_fn(lock, p) : 0;
 554                 if (unlikely(ret)) {
 555                         raw_spin_lock(&lock->wait_lock);
 556                         if (!wait->lock_acquired)
 557                                 list_del(&wait->list);
 558                         raw_spin_unlock(&lock->wait_lock);
 559
 560                         if (wait->lock_acquired)
 561                                 do_six_unlock_type(lock, type);
 562                         break;
 563                 }
 564
 565                 schedule();
 566         }
 567
 568         __set_current_state(TASK_RUNNING);
 569 out:
 570         if (ret && type == SIX_LOCK_write && lock->state.write_locking) {
 571                 old.v = atomic64_sub_return(__SIX_VAL(write_locking, 1),
 572                                             &lock->state.counter);
 573                 six_lock_wakeup(lock, old, SIX_LOCK_read);
 574         }
 575         trace_contention_end(lock, 0);
 576
 577         return ret;
 578 }
 579
 580 __always_inline __flatten
 581 static int __six_lock_type_waiter(struct six_lock *lock, enum six_lock_type type,
 582                          struct six_lock_waiter *wait,
 583                          six_lock_should_sleep_fn should_sleep_fn, void *p,
 584                          unsigned long ip)
 585 {
 586         int ret;
 587
 588         wait->start_time = 0;
 589
 590         if (type != SIX_LOCK_write)
 591                 six_acquire(&lock->dep_map, 0, type == SIX_LOCK_read, ip);
 592
 593         ret = do_six_trylock_type(lock, type, true) ? 0
 594                 : __six_lock_type_slowpath(lock, type, wait, should_sleep_fn, p, ip);
 595
 596         if (ret && type != SIX_LOCK_write)
 597                 six_release(&lock->dep_map, ip);
 598         if (!ret)
 599                 lock_acquired(&lock->dep_map, ip);
 600
 601         return ret;
 602 }
 603
 604 __always_inline
 605 static int __six_lock_type(struct six_lock *lock, enum six_lock_type type,
 606                            six_lock_should_sleep_fn should_sleep_fn, void *p,
 607                            unsigned long ip)
 608 {
 609         struct six_lock_waiter wait;
 610
 611         return __six_lock_type_waiter(lock, type, &wait, should_sleep_fn, p, ip);
 612 }
 613
 614 __always_inline __flatten
 615 static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type)
 616 {
 617         const struct six_lock_vals l[] = LOCK_VALS;
 618         union six_lock_state state;
 619
 620         if (type == SIX_LOCK_intent)
 621                 lock->owner = NULL;
 622
 623         if (type == SIX_LOCK_read &&
 624             lock->readers) {
 625                 smp_mb(); /* unlock barrier */
 626                 this_cpu_dec(*lock->readers);
 627                 smp_mb(); /* between unlocking and checking for waiters */
 628                 state.v = READ_ONCE(lock->state.v);
 629         } else {
 630                 u64 v = l[type].unlock_val;
 631
 632                 if (type != SIX_LOCK_read)
 633                         v -= lock->state.v & __SIX_VAL(nospin, 1);
 634
 635                 EBUG_ON(!(lock->state.v & l[type].held_mask));
 636                 state.v = atomic64_add_return_release(v, &lock->state.counter);
 637         }
 638
 639         six_lock_wakeup(lock, state, l[type].unlock_wakeup);
 640 }
 641
 642 __always_inline __flatten
 643 static void __six_unlock_type(struct six_lock *lock, enum six_lock_type type,
 644                               unsigned long ip)
 645 {
 646         EBUG_ON(type == SIX_LOCK_write &&
 647                 !(lock->state.v & __SIX_LOCK_HELD_intent));
 648         EBUG_ON((type == SIX_LOCK_write ||
 649                  type == SIX_LOCK_intent) &&
 650                 lock->owner != current);
 651
 652         if (type != SIX_LOCK_write)
 653                 six_release(&lock->dep_map, ip);
 654
 655         if (type == SIX_LOCK_intent &&
 656             lock->intent_lock_recurse) {
 657                 --lock->intent_lock_recurse;
 658                 return;
 659         }
 660
 661         do_six_unlock_type(lock, type);
 662 }
 663
 664 #define __SIX_LOCK(type)                                                \
 665 bool six_trylock_ip_##type(struct six_lock *lock, unsigned long ip)     \
 666 {                                                                       \
 667         return __six_trylock_type(lock, SIX_LOCK_##type, ip);           \
 668 }                                                                       \
 669 EXPORT_SYMBOL_GPL(six_trylock_ip_##type);                               \
 670                                                                         \
 671 bool six_relock_ip_##type(struct six_lock *lock, u32 seq, unsigned long ip)\
 672 {                                                                       \
 673         return __six_relock_type(lock, SIX_LOCK_##type, seq, ip);       \
 674 }                                                                       \
 675 EXPORT_SYMBOL_GPL(six_relock_ip_##type);                                \
 676                                                                         \
 677 int six_lock_ip_##type(struct six_lock *lock,                           \
 678                     six_lock_should_sleep_fn should_sleep_fn, void *p,  \
 679                     unsigned long ip)                                   \
 680 {                                                                       \
 681         return __six_lock_type(lock, SIX_LOCK_##type, should_sleep_fn, p, ip);\
 682 }                                                                       \
 683 EXPORT_SYMBOL_GPL(six_lock_ip_##type);                                  \
 684                                                                         \
 685 int six_lock_ip_waiter_##type(struct six_lock *lock,                    \
 686                            struct six_lock_waiter *wait,                \
 687                            six_lock_should_sleep_fn should_sleep_fn, void *p,\
 688                            unsigned long ip)                            \
 689 {                                                                       \
 690         return __six_lock_type_waiter(lock, SIX_LOCK_##type, wait, should_sleep_fn, p, ip);\
 691 }                                                                       \
 692 EXPORT_SYMBOL_GPL(six_lock_ip_waiter_##type);                           \
 693                                                                         \
 694 void six_unlock_ip_##type(struct six_lock *lock, unsigned long ip)      \
 695 {                                                                       \
 696         __six_unlock_type(lock, SIX_LOCK_##type, ip);                   \
 697 }                                                                       \
 698 EXPORT_SYMBOL_GPL(six_unlock_ip_##type);
 699
 700 __SIX_LOCK(read)
 701 __SIX_LOCK(intent)
 702 __SIX_LOCK(write)
 703
 704 #undef __SIX_LOCK
 705
 706 /* Convert from intent to read: */
 707 void six_lock_downgrade(struct six_lock *lock)
 708 {
 709         six_lock_increment(lock, SIX_LOCK_read);
 710         six_unlock_intent(lock);
 711 }
 712 EXPORT_SYMBOL_GPL(six_lock_downgrade);
 713
 714 bool six_lock_tryupgrade(struct six_lock *lock)
 715 {
 716         union six_lock_state old, new;
 717         u64 v = READ_ONCE(lock->state.v);
 718
 719         do {
 720                 new.v = old.v = v;
 721
 722                 if (new.intent_lock)
 723                         return false;
 724
 725                 if (!lock->readers) {
 726                         EBUG_ON(!new.read_lock);
 727                         new.read_lock--;
 728                 }
 729
 730                 new.intent_lock = 1;
 731         } while ((v = atomic64_cmpxchg_acquire(&lock->state.counter,
 732                                 old.v, new.v)) != old.v);
 733
 734         if (lock->readers)
 735                 this_cpu_dec(*lock->readers);
 736
 737         six_set_owner(lock, SIX_LOCK_intent, old, current);
 738
 739         return true;
 740 }
 741 EXPORT_SYMBOL_GPL(six_lock_tryupgrade);
 742
 743 bool six_trylock_convert(struct six_lock *lock,
 744                          enum six_lock_type from,
 745                          enum six_lock_type to)
 746 {
 747         EBUG_ON(to == SIX_LOCK_write || from == SIX_LOCK_write);
 748
 749         if (to == from)
 750                 return true;
 751
 752         if (to == SIX_LOCK_read) {
 753                 six_lock_downgrade(lock);
 754                 return true;
 755         } else {
 756                 return six_lock_tryupgrade(lock);
 757         }
 758 }
 759 EXPORT_SYMBOL_GPL(six_trylock_convert);
 760
 761 /*
 762  * Increment read/intent lock count, assuming we already have it read or intent
 763  * locked:
 764  */
 765 void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
 766 {
 767         const struct six_lock_vals l[] = LOCK_VALS;
 768
 769         six_acquire(&lock->dep_map, 0, type == SIX_LOCK_read, _RET_IP_);
 770
 771         /* XXX: assert already locked, and that we don't overflow: */
 772
 773         switch (type) {
 774         case SIX_LOCK_read:
 775                 if (lock->readers) {
 776                         this_cpu_inc(*lock->readers);
 777                 } else {
 778                         EBUG_ON(!lock->state.read_lock &&
 779                                 !lock->state.intent_lock);
 780                         atomic64_add(l[type].lock_val, &lock->state.counter);
 781                 }
 782                 break;
 783         case SIX_LOCK_intent:
 784                 EBUG_ON(!lock->state.intent_lock);
 785                 lock->intent_lock_recurse++;
 786                 break;
 787         case SIX_LOCK_write:
 788                 BUG();
 789                 break;
 790         }
 791 }
 792 EXPORT_SYMBOL_GPL(six_lock_increment);
 793
 794 void six_lock_wakeup_all(struct six_lock *lock)
 795 {
 796         union six_lock_state state = lock->state;
 797         struct six_lock_waiter *w;
 798
 799         six_lock_wakeup(lock, state, SIX_LOCK_read);
 800         six_lock_wakeup(lock, state, SIX_LOCK_intent);
 801         six_lock_wakeup(lock, state, SIX_LOCK_write);
 802
 803         raw_spin_lock(&lock->wait_lock);
 804         list_for_each_entry(w, &lock->wait_list, list)
 805                 wake_up_process(w->task);
 806         raw_spin_unlock(&lock->wait_lock);
 807 }
 808 EXPORT_SYMBOL_GPL(six_lock_wakeup_all);
 809
 810 void six_lock_pcpu_free(struct six_lock *lock)
 811 {
 812         BUG_ON(lock->readers && pcpu_read_count(lock));
 813         BUG_ON(lock->state.read_lock);
 814
 815         free_percpu(lock->readers);
 816         lock->readers = NULL;
 817 }
 818 EXPORT_SYMBOL_GPL(six_lock_pcpu_free);
 819
 820 void six_lock_pcpu_alloc(struct six_lock *lock)
 821 {
 822 #ifdef __KERNEL__
 823         if (!lock->readers)
 824                 lock->readers = alloc_percpu(unsigned);
 825 #endif
 826 }
 827 EXPORT_SYMBOL_GPL(six_lock_pcpu_alloc);
 828
 829 /*
 830  * Returns lock held counts, for both read and intent
 831  */
 832 struct six_lock_count six_lock_counts(struct six_lock *lock)
 833 {
 834         struct six_lock_count ret;
 835
 836         ret.n[SIX_LOCK_read]    = !lock->readers
 837                 ? lock->state.read_lock
 838                 : pcpu_read_count(lock);
 839         ret.n[SIX_LOCK_intent]  = lock->state.intent_lock + lock->intent_lock_recurse;
 840         ret.n[SIX_LOCK_write]   = lock->state.seq & 1;
 841
 842         return ret;
 843 }
 844 EXPORT_SYMBOL_GPL(six_lock_counts);