git.sesse.net Git - bcachefs-tools-debian/blob - c_src/libbcachefs/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/sched/task.h>
  12 #include <linux/slab.h>
  13
  14 #include <trace/events/lock.h>
  15
  16 #include "six.h"
  17
  18 #ifdef DEBUG
  19 #define EBUG_ON(cond)                   BUG_ON(cond)
  20 #else
  21 #define EBUG_ON(cond)                   do {} while (0)
  22 #endif
  23
  24 #define six_acquire(l, t, r, ip)        lock_acquire(l, 0, t, r, 1, NULL, ip)
  25 #define six_release(l, ip)              lock_release(l, ip)
  26
  27 static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type);
  28
  29 #define SIX_LOCK_HELD_read_OFFSET       0
  30 #define SIX_LOCK_HELD_read              ~(~0U << 26)
  31 #define SIX_LOCK_HELD_intent            (1U << 26)
  32 #define SIX_LOCK_HELD_write             (1U << 27)
  33 #define SIX_LOCK_WAITING_read           (1U << (28 + SIX_LOCK_read))
  34 #define SIX_LOCK_WAITING_write          (1U << (28 + SIX_LOCK_write))
  35 #define SIX_LOCK_NOSPIN                 (1U << 31)
  36
  37 struct six_lock_vals {
  38         /* Value we add to the lock in order to take the lock: */
  39         u32                     lock_val;
  40
  41         /* If the lock has this value (used as a mask), taking the lock fails: */
  42         u32                     lock_fail;
  43
  44         /* Mask that indicates lock is held for this type: */
  45         u32                     held_mask;
  46
  47         /* Waitlist we wakeup when releasing the lock: */
  48         enum six_lock_type      unlock_wakeup;
  49 };
  50
  51 static const struct six_lock_vals l[] = {
  52         [SIX_LOCK_read] = {
  53                 .lock_val       = 1U << SIX_LOCK_HELD_read_OFFSET,
  54                 .lock_fail      = SIX_LOCK_HELD_write,
  55                 .held_mask      = SIX_LOCK_HELD_read,
  56                 .unlock_wakeup  = SIX_LOCK_write,
  57         },
  58         [SIX_LOCK_intent] = {
  59                 .lock_val       = SIX_LOCK_HELD_intent,
  60                 .lock_fail      = SIX_LOCK_HELD_intent,
  61                 .held_mask      = SIX_LOCK_HELD_intent,
  62                 .unlock_wakeup  = SIX_LOCK_intent,
  63         },
  64         [SIX_LOCK_write] = {
  65                 .lock_val       = SIX_LOCK_HELD_write,
  66                 .lock_fail      = SIX_LOCK_HELD_read,
  67                 .held_mask      = SIX_LOCK_HELD_write,
  68                 .unlock_wakeup  = SIX_LOCK_read,
  69         },
  70 };
  71
  72 static inline void six_set_bitmask(struct six_lock *lock, u32 mask)
  73 {
  74         if ((atomic_read(&lock->state) & mask) != mask)
  75                 atomic_or(mask, &lock->state);
  76 }
  77
  78 static inline void six_clear_bitmask(struct six_lock *lock, u32 mask)
  79 {
  80         if (atomic_read(&lock->state) & mask)
  81                 atomic_and(~mask, &lock->state);
  82 }
  83
  84 static inline void six_set_owner(struct six_lock *lock, enum six_lock_type type,
  85                                  u32 old, struct task_struct *owner)
  86 {
  87         if (type != SIX_LOCK_intent)
  88                 return;
  89
  90         if (!(old & SIX_LOCK_HELD_intent)) {
  91                 EBUG_ON(lock->owner);
  92                 lock->owner = owner;
  93         } else {
  94                 EBUG_ON(lock->owner != current);
  95         }
  96 }
  97
  98 static inline unsigned pcpu_read_count(struct six_lock *lock)
  99 {
 100         unsigned read_count = 0;
 101         int cpu;
 102
 103         for_each_possible_cpu(cpu)
 104                 read_count += *per_cpu_ptr(lock->readers, cpu);
 105         return read_count;
 106 }
 107
 108 /*
 109  * __do_six_trylock() - main trylock routine
 110  *
 111  * Returns 1 on success, 0 on failure
 112  *
 113  * In percpu reader mode, a failed trylock may cause a spurious trylock failure
 114  * for anoter thread taking the competing lock type, and we may havve to do a
 115  * wakeup: when a wakeup is required, we return -1 - wakeup_type.
 116  */
 117 static int __do_six_trylock(struct six_lock *lock, enum six_lock_type type,
 118                             struct task_struct *task, bool try)
 119 {
 120         int ret;
 121         u32 old;
 122
 123         EBUG_ON(type == SIX_LOCK_write && lock->owner != task);
 124         EBUG_ON(type == SIX_LOCK_write &&
 125                 (try != !(atomic_read(&lock->state) & SIX_LOCK_HELD_write)));
 126
 127         /*
 128          * Percpu reader mode:
 129          *
 130          * The basic idea behind this algorithm is that you can implement a lock
 131          * between two threads without any atomics, just memory barriers:
 132          *
 133          * For two threads you'll need two variables, one variable for "thread a
 134          * has the lock" and another for "thread b has the lock".
 135          *
 136          * To take the lock, a thread sets its variable indicating that it holds
 137          * the lock, then issues a full memory barrier, then reads from the
 138          * other thread's variable to check if the other thread thinks it has
 139          * the lock. If we raced, we backoff and retry/sleep.
 140          *
 141          * Failure to take the lock may cause a spurious trylock failure in
 142          * another thread, because we temporarily set the lock to indicate that
 143          * we held it. This would be a problem for a thread in six_lock(), when
 144          * they are calling trylock after adding themself to the waitlist and
 145          * prior to sleeping.
 146          *
 147          * Therefore, if we fail to get the lock, and there were waiters of the
 148          * type we conflict with, we will have to issue a wakeup.
 149          *
 150          * Since we may be called under wait_lock (and by the wakeup code
 151          * itself), we return that the wakeup has to be done instead of doing it
 152          * here.
 153          */
 154         if (type == SIX_LOCK_read && lock->readers) {
 155                 preempt_disable();
 156                 this_cpu_inc(*lock->readers); /* signal that we own lock */
 157
 158                 smp_mb();
 159
 160                 old = atomic_read(&lock->state);
 161                 ret = !(old & l[type].lock_fail);
 162
 163                 this_cpu_sub(*lock->readers, !ret);
 164                 preempt_enable();
 165
 166                 if (!ret) {
 167                         smp_mb();
 168                         if (atomic_read(&lock->state) & SIX_LOCK_WAITING_write)
 169                                 ret = -1 - SIX_LOCK_write;
 170                 }
 171         } else if (type == SIX_LOCK_write && lock->readers) {
 172                 if (try) {
 173                         atomic_add(SIX_LOCK_HELD_write, &lock->state);
 174                         smp_mb__after_atomic();
 175                 }
 176
 177                 ret = !pcpu_read_count(lock);
 178
 179                 if (try && !ret) {
 180                         old = atomic_sub_return(SIX_LOCK_HELD_write, &lock->state);
 181                         if (old & SIX_LOCK_WAITING_read)
 182                                 ret = -1 - SIX_LOCK_read;
 183                 }
 184         } else {
 185                 old = atomic_read(&lock->state);
 186                 do {
 187                         ret = !(old & l[type].lock_fail);
 188                         if (!ret || (type == SIX_LOCK_write && !try)) {
 189                                 smp_mb();
 190                                 break;
 191                         }
 192                 } while (!atomic_try_cmpxchg_acquire(&lock->state, &old, old + l[type].lock_val));
 193
 194                 EBUG_ON(ret && !(atomic_read(&lock->state) & l[type].held_mask));
 195         }
 196
 197         if (ret > 0)
 198                 six_set_owner(lock, type, old, task);
 199
 200         EBUG_ON(type == SIX_LOCK_write && try && ret <= 0 &&
 201                 (atomic_read(&lock->state) & SIX_LOCK_HELD_write));
 202
 203         return ret;
 204 }
 205
 206 static void __six_lock_wakeup(struct six_lock *lock, enum six_lock_type lock_type)
 207 {
 208         struct six_lock_waiter *w, *next;
 209         struct task_struct *task;
 210         bool saw_one;
 211         int ret;
 212 again:
 213         ret = 0;
 214         saw_one = false;
 215         raw_spin_lock(&lock->wait_lock);
 216
 217         list_for_each_entry_safe(w, next, &lock->wait_list, list) {
 218                 if (w->lock_want != lock_type)
 219                         continue;
 220
 221                 if (saw_one && lock_type != SIX_LOCK_read)
 222                         goto unlock;
 223                 saw_one = true;
 224
 225                 ret = __do_six_trylock(lock, lock_type, w->task, false);
 226                 if (ret <= 0)
 227                         goto unlock;
 228
 229                 /*
 230                  * Similar to percpu_rwsem_wake_function(), we need to guard
 231                  * against the wakee noticing w->lock_acquired, returning, and
 232                  * then exiting before we do the wakeup:
 233                  */
 234                 task = get_task_struct(w->task);
 235                 __list_del(w->list.prev, w->list.next);
 236                 /*
 237                  * The release barrier here ensures the ordering of the
 238                  * __list_del before setting w->lock_acquired; @w is on the
 239                  * stack of the thread doing the waiting and will be reused
 240                  * after it sees w->lock_acquired with no other locking:
 241                  * pairs with smp_load_acquire() in six_lock_slowpath()
 242                  */
 243                 smp_store_release(&w->lock_acquired, true);
 244                 wake_up_process(task);
 245                 put_task_struct(task);
 246         }
 247
 248         six_clear_bitmask(lock, SIX_LOCK_WAITING_read << lock_type);
 249 unlock:
 250         raw_spin_unlock(&lock->wait_lock);
 251
 252         if (ret < 0) {
 253                 lock_type = -ret - 1;
 254                 goto again;
 255         }
 256 }
 257
 258 __always_inline
 259 static void six_lock_wakeup(struct six_lock *lock, u32 state,
 260                             enum six_lock_type lock_type)
 261 {
 262         if (lock_type == SIX_LOCK_write && (state & SIX_LOCK_HELD_read))
 263                 return;
 264
 265         if (!(state & (SIX_LOCK_WAITING_read << lock_type)))
 266                 return;
 267
 268         __six_lock_wakeup(lock, lock_type);
 269 }
 270
 271 __always_inline
 272 static bool do_six_trylock(struct six_lock *lock, enum six_lock_type type, bool try)
 273 {
 274         int ret;
 275
 276         ret = __do_six_trylock(lock, type, current, try);
 277         if (ret < 0)
 278                 __six_lock_wakeup(lock, -ret - 1);
 279
 280         return ret > 0;
 281 }
 282
 283 /**
 284  * six_trylock_ip - attempt to take a six lock without blocking
 285  * @lock:       lock to take
 286  * @type:       SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
 287  * @ip:         ip parameter for lockdep/lockstat, i.e. _THIS_IP_
 288  *
 289  * Return: true on success, false on failure.
 290  */
 291 bool six_trylock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip)
 292 {
 293         if (!do_six_trylock(lock, type, true))
 294                 return false;
 295
 296         if (type != SIX_LOCK_write)
 297                 six_acquire(&lock->dep_map, 1, type == SIX_LOCK_read, ip);
 298         return true;
 299 }
 300 EXPORT_SYMBOL_GPL(six_trylock_ip);
 301
 302 /**
 303  * six_relock_ip - attempt to re-take a lock that was held previously
 304  * @lock:       lock to take
 305  * @type:       SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
 306  * @seq:        lock sequence number obtained from six_lock_seq() while lock was
 307  *              held previously
 308  * @ip:         ip parameter for lockdep/lockstat, i.e. _THIS_IP_
 309  *
 310  * Return: true on success, false on failure.
 311  */
 312 bool six_relock_ip(struct six_lock *lock, enum six_lock_type type,
 313                    unsigned seq, unsigned long ip)
 314 {
 315         if (six_lock_seq(lock) != seq || !six_trylock_ip(lock, type, ip))
 316                 return false;
 317
 318         if (six_lock_seq(lock) != seq) {
 319                 six_unlock_ip(lock, type, ip);
 320                 return false;
 321         }
 322
 323         return true;
 324 }
 325 EXPORT_SYMBOL_GPL(six_relock_ip);
 326
 327 #ifdef CONFIG_BCACHEFS_SIX_OPTIMISTIC_SPIN
 328
 329 static inline bool six_owner_running(struct six_lock *lock)
 330 {
 331         /*
 332          * When there's no owner, we might have preempted between the owner
 333          * acquiring the lock and setting the owner field. If we're an RT task
 334          * that will live-lock because we won't let the owner complete.
 335          */
 336         rcu_read_lock();
 337         struct task_struct *owner = READ_ONCE(lock->owner);
 338         bool ret = owner ? owner_on_cpu(owner) : !rt_task(current);
 339         rcu_read_unlock();
 340
 341         return ret;
 342 }
 343
 344 static inline bool six_optimistic_spin(struct six_lock *lock,
 345                                        struct six_lock_waiter *wait,
 346                                        enum six_lock_type type)
 347 {
 348         unsigned loop = 0;
 349         u64 end_time;
 350
 351         if (type == SIX_LOCK_write)
 352                 return false;
 353
 354         if (lock->wait_list.next != &wait->list)
 355                 return false;
 356
 357         if (atomic_read(&lock->state) & SIX_LOCK_NOSPIN)
 358                 return false;
 359
 360         preempt_disable();
 361         end_time = sched_clock() + 10 * NSEC_PER_USEC;
 362
 363         while (!need_resched() && six_owner_running(lock)) {
 364                 /*
 365                  * Ensures that writes to the waitlist entry happen after we see
 366                  * wait->lock_acquired: pairs with the smp_store_release in
 367                  * __six_lock_wakeup
 368                  */
 369                 if (smp_load_acquire(&wait->lock_acquired)) {
 370                         preempt_enable();
 371                         return true;
 372                 }
 373
 374                 if (!(++loop & 0xf) && (time_after64(sched_clock(), end_time))) {
 375                         six_set_bitmask(lock, SIX_LOCK_NOSPIN);
 376                         break;
 377                 }
 378
 379                 /*
 380                  * The cpu_relax() call is a compiler barrier which forces
 381                  * everything in this loop to be re-loaded. We don't need
 382                  * memory barriers as we'll eventually observe the right
 383                  * values at the cost of a few extra spins.
 384                  */
 385                 cpu_relax();
 386         }
 387
 388         preempt_enable();
 389         return false;
 390 }
 391
 392 #else /* CONFIG_LOCK_SPIN_ON_OWNER */
 393
 394 static inline bool six_optimistic_spin(struct six_lock *lock,
 395                                        struct six_lock_waiter *wait,
 396                                        enum six_lock_type type)
 397 {
 398         return false;
 399 }
 400
 401 #endif
 402
 403 noinline
 404 static int six_lock_slowpath(struct six_lock *lock, enum six_lock_type type,
 405                              struct six_lock_waiter *wait,
 406                              six_lock_should_sleep_fn should_sleep_fn, void *p,
 407                              unsigned long ip)
 408 {
 409         int ret = 0;
 410
 411         if (type == SIX_LOCK_write) {
 412                 EBUG_ON(atomic_read(&lock->state) & SIX_LOCK_HELD_write);
 413                 atomic_add(SIX_LOCK_HELD_write, &lock->state);
 414                 smp_mb__after_atomic();
 415         }
 416
 417         trace_contention_begin(lock, 0);
 418         lock_contended(&lock->dep_map, ip);
 419
 420         wait->task              = current;
 421         wait->lock_want         = type;
 422         wait->lock_acquired     = false;
 423
 424         raw_spin_lock(&lock->wait_lock);
 425         six_set_bitmask(lock, SIX_LOCK_WAITING_read << type);
 426         /*
 427          * Retry taking the lock after taking waitlist lock, in case we raced
 428          * with an unlock:
 429          */
 430         ret = __do_six_trylock(lock, type, current, false);
 431         if (ret <= 0) {
 432                 wait->start_time = local_clock();
 433
 434                 if (!list_empty(&lock->wait_list)) {
 435                         struct six_lock_waiter *last =
 436                                 list_last_entry(&lock->wait_list,
 437                                         struct six_lock_waiter, list);
 438
 439                         if (time_before_eq64(wait->start_time, last->start_time))
 440                                 wait->start_time = last->start_time + 1;
 441                 }
 442
 443                 list_add_tail(&wait->list, &lock->wait_list);
 444         }
 445         raw_spin_unlock(&lock->wait_lock);
 446
 447         if (unlikely(ret > 0)) {
 448                 ret = 0;
 449                 goto out;
 450         }
 451
 452         if (unlikely(ret < 0)) {
 453                 __six_lock_wakeup(lock, -ret - 1);
 454                 ret = 0;
 455         }
 456
 457         if (six_optimistic_spin(lock, wait, type))
 458                 goto out;
 459
 460         while (1) {
 461                 set_current_state(TASK_UNINTERRUPTIBLE);
 462
 463                 /*
 464                  * Ensures that writes to the waitlist entry happen after we see
 465                  * wait->lock_acquired: pairs with the smp_store_release in
 466                  * __six_lock_wakeup
 467                  */
 468                 if (smp_load_acquire(&wait->lock_acquired))
 469                         break;
 470
 471                 ret = should_sleep_fn ? should_sleep_fn(lock, p) : 0;
 472                 if (unlikely(ret)) {
 473                         bool acquired;
 474
 475                         /*
 476                          * If should_sleep_fn() returns an error, we are
 477                          * required to return that error even if we already
 478                          * acquired the lock - should_sleep_fn() might have
 479                          * modified external state (e.g. when the deadlock cycle
 480                          * detector in bcachefs issued a transaction restart)
 481                          */
 482                         raw_spin_lock(&lock->wait_lock);
 483                         acquired = wait->lock_acquired;
 484                         if (!acquired)
 485                                 list_del(&wait->list);
 486                         raw_spin_unlock(&lock->wait_lock);
 487
 488                         if (unlikely(acquired))
 489                                 do_six_unlock_type(lock, type);
 490                         break;
 491                 }
 492
 493                 schedule();
 494         }
 495
 496         __set_current_state(TASK_RUNNING);
 497 out:
 498         if (ret && type == SIX_LOCK_write) {
 499                 six_clear_bitmask(lock, SIX_LOCK_HELD_write);
 500                 six_lock_wakeup(lock, atomic_read(&lock->state), SIX_LOCK_read);
 501         }
 502         trace_contention_end(lock, 0);
 503
 504         return ret;
 505 }
 506
 507 /**
 508  * six_lock_ip_waiter - take a lock, with full waitlist interface
 509  * @lock:       lock to take
 510  * @type:       SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
 511  * @wait:       pointer to wait object, which will be added to lock's waitlist
 512  * @should_sleep_fn: callback run after adding to waitlist, immediately prior
 513  *              to scheduling
 514  * @p:          passed through to @should_sleep_fn
 515  * @ip:         ip parameter for lockdep/lockstat, i.e. _THIS_IP_
 516  *
 517  * This is the most general six_lock() variant, with parameters to support full
 518  * cycle detection for deadlock avoidance.
 519  *
 520  * The code calling this function must implement tracking of held locks, and the
 521  * @wait object should be embedded into the struct that tracks held locks -
 522  * which must also be accessible in a thread-safe way.
 523  *
 524  * @should_sleep_fn should invoke the cycle detector; it should walk each
 525  * lock's waiters, and for each waiter recursively walk their held locks.
 526  *
 527  * When this function must block, @wait will be added to @lock's waitlist before
 528  * calling trylock, and before calling @should_sleep_fn, and @wait will not be
 529  * removed from the lock waitlist until the lock has been successfully acquired,
 530  * or we abort.
 531  *
 532  * @wait.start_time will be monotonically increasing for any given waitlist, and
 533  * thus may be used as a loop cursor.
 534  *
 535  * Return: 0 on success, or the return code from @should_sleep_fn on failure.
 536  */
 537 int six_lock_ip_waiter(struct six_lock *lock, enum six_lock_type type,
 538                        struct six_lock_waiter *wait,
 539                        six_lock_should_sleep_fn should_sleep_fn, void *p,
 540                        unsigned long ip)
 541 {
 542         int ret;
 543
 544         wait->start_time = 0;
 545
 546         if (type != SIX_LOCK_write)
 547                 six_acquire(&lock->dep_map, 0, type == SIX_LOCK_read, ip);
 548
 549         ret = do_six_trylock(lock, type, true) ? 0
 550                 : six_lock_slowpath(lock, type, wait, should_sleep_fn, p, ip);
 551
 552         if (ret && type != SIX_LOCK_write)
 553                 six_release(&lock->dep_map, ip);
 554         if (!ret)
 555                 lock_acquired(&lock->dep_map, ip);
 556
 557         return ret;
 558 }
 559 EXPORT_SYMBOL_GPL(six_lock_ip_waiter);
 560
 561 __always_inline
 562 static void do_six_unlock_type(struct six_lock *lock, enum six_lock_type type)
 563 {
 564         u32 state;
 565
 566         if (type == SIX_LOCK_intent)
 567                 lock->owner = NULL;
 568
 569         if (type == SIX_LOCK_read &&
 570             lock->readers) {
 571                 smp_mb(); /* unlock barrier */
 572                 this_cpu_dec(*lock->readers);
 573                 smp_mb(); /* between unlocking and checking for waiters */
 574                 state = atomic_read(&lock->state);
 575         } else {
 576                 u32 v = l[type].lock_val;
 577
 578                 if (type != SIX_LOCK_read)
 579                         v += atomic_read(&lock->state) & SIX_LOCK_NOSPIN;
 580
 581                 EBUG_ON(!(atomic_read(&lock->state) & l[type].held_mask));
 582                 state = atomic_sub_return_release(v, &lock->state);
 583         }
 584
 585         six_lock_wakeup(lock, state, l[type].unlock_wakeup);
 586 }
 587
 588 /**
 589  * six_unlock_ip - drop a six lock
 590  * @lock:       lock to unlock
 591  * @type:       SIX_LOCK_read, SIX_LOCK_intent, or SIX_LOCK_write
 592  * @ip:         ip parameter for lockdep/lockstat, i.e. _THIS_IP_
 593  *
 594  * When a lock is held multiple times (because six_lock_incement()) was used),
 595  * this decrements the 'lock held' counter by one.
 596  *
 597  * For example:
 598  * six_lock_read(&foo->lock);                           read count 1
 599  * six_lock_increment(&foo->lock, SIX_LOCK_read);       read count 2
 600  * six_lock_unlock(&foo->lock, SIX_LOCK_read);          read count 1
 601  * six_lock_unlock(&foo->lock, SIX_LOCK_read);          read count 0
 602  */
 603 void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long ip)
 604 {
 605         EBUG_ON(type == SIX_LOCK_write &&
 606                 !(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
 607         EBUG_ON((type == SIX_LOCK_write ||
 608                  type == SIX_LOCK_intent) &&
 609                 lock->owner != current);
 610
 611         if (type != SIX_LOCK_write)
 612                 six_release(&lock->dep_map, ip);
 613         else
 614                 lock->seq++;
 615
 616         if (type == SIX_LOCK_intent &&
 617             lock->intent_lock_recurse) {
 618                 --lock->intent_lock_recurse;
 619                 return;
 620         }
 621
 622         do_six_unlock_type(lock, type);
 623 }
 624 EXPORT_SYMBOL_GPL(six_unlock_ip);
 625
 626 /**
 627  * six_lock_downgrade - convert an intent lock to a read lock
 628  * @lock:       lock to dowgrade
 629  *
 630  * @lock will have read count incremented and intent count decremented
 631  */
 632 void six_lock_downgrade(struct six_lock *lock)
 633 {
 634         six_lock_increment(lock, SIX_LOCK_read);
 635         six_unlock_intent(lock);
 636 }
 637 EXPORT_SYMBOL_GPL(six_lock_downgrade);
 638
 639 /**
 640  * six_lock_tryupgrade - attempt to convert read lock to an intent lock
 641  * @lock:       lock to upgrade
 642  *
 643  * On success, @lock will have intent count incremented and read count
 644  * decremented
 645  *
 646  * Return: true on success, false on failure
 647  */
 648 bool six_lock_tryupgrade(struct six_lock *lock)
 649 {
 650         u32 old = atomic_read(&lock->state), new;
 651
 652         do {
 653                 new = old;
 654
 655                 if (new & SIX_LOCK_HELD_intent)
 656                         return false;
 657
 658                 if (!lock->readers) {
 659                         EBUG_ON(!(new & SIX_LOCK_HELD_read));
 660                         new -= l[SIX_LOCK_read].lock_val;
 661                 }
 662
 663                 new |= SIX_LOCK_HELD_intent;
 664         } while (!atomic_try_cmpxchg_acquire(&lock->state, &old, new));
 665
 666         if (lock->readers)
 667                 this_cpu_dec(*lock->readers);
 668
 669         six_set_owner(lock, SIX_LOCK_intent, old, current);
 670
 671         return true;
 672 }
 673 EXPORT_SYMBOL_GPL(six_lock_tryupgrade);
 674
 675 /**
 676  * six_trylock_convert - attempt to convert a held lock from one type to another
 677  * @lock:       lock to upgrade
 678  * @from:       SIX_LOCK_read or SIX_LOCK_intent
 679  * @to:         SIX_LOCK_read or SIX_LOCK_intent
 680  *
 681  * On success, @lock will have intent count incremented and read count
 682  * decremented
 683  *
 684  * Return: true on success, false on failure
 685  */
 686 bool six_trylock_convert(struct six_lock *lock,
 687                          enum six_lock_type from,
 688                          enum six_lock_type to)
 689 {
 690         EBUG_ON(to == SIX_LOCK_write || from == SIX_LOCK_write);
 691
 692         if (to == from)
 693                 return true;
 694
 695         if (to == SIX_LOCK_read) {
 696                 six_lock_downgrade(lock);
 697                 return true;
 698         } else {
 699                 return six_lock_tryupgrade(lock);
 700         }
 701 }
 702 EXPORT_SYMBOL_GPL(six_trylock_convert);
 703
 704 /**
 705  * six_lock_increment - increase held lock count on a lock that is already held
 706  * @lock:       lock to increment
 707  * @type:       SIX_LOCK_read or SIX_LOCK_intent
 708  *
 709  * @lock must already be held, with a lock type that is greater than or equal to
 710  * @type
 711  *
 712  * A corresponding six_unlock_type() call will be required for @lock to be fully
 713  * unlocked.
 714  */
 715 void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
 716 {
 717         six_acquire(&lock->dep_map, 0, type == SIX_LOCK_read, _RET_IP_);
 718
 719         /* XXX: assert already locked, and that we don't overflow: */
 720
 721         switch (type) {
 722         case SIX_LOCK_read:
 723                 if (lock->readers) {
 724                         this_cpu_inc(*lock->readers);
 725                 } else {
 726                         EBUG_ON(!(atomic_read(&lock->state) &
 727                                   (SIX_LOCK_HELD_read|
 728                                    SIX_LOCK_HELD_intent)));
 729                         atomic_add(l[type].lock_val, &lock->state);
 730                 }
 731                 break;
 732         case SIX_LOCK_intent:
 733                 EBUG_ON(!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
 734                 lock->intent_lock_recurse++;
 735                 break;
 736         case SIX_LOCK_write:
 737                 BUG();
 738                 break;
 739         }
 740 }
 741 EXPORT_SYMBOL_GPL(six_lock_increment);
 742
 743 /**
 744  * six_lock_wakeup_all - wake up all waiters on @lock
 745  * @lock:       lock to wake up waiters for
 746  *
 747  * Wakeing up waiters will cause them to re-run should_sleep_fn, which may then
 748  * abort the lock operation.
 749  *
 750  * This function is never needed in a bug-free program; it's only useful in
 751  * debug code, e.g. to determine if a cycle detector is at fault.
 752  */
 753 void six_lock_wakeup_all(struct six_lock *lock)
 754 {
 755         u32 state = atomic_read(&lock->state);
 756         struct six_lock_waiter *w;
 757
 758         six_lock_wakeup(lock, state, SIX_LOCK_read);
 759         six_lock_wakeup(lock, state, SIX_LOCK_intent);
 760         six_lock_wakeup(lock, state, SIX_LOCK_write);
 761
 762         raw_spin_lock(&lock->wait_lock);
 763         list_for_each_entry(w, &lock->wait_list, list)
 764                 wake_up_process(w->task);
 765         raw_spin_unlock(&lock->wait_lock);
 766 }
 767 EXPORT_SYMBOL_GPL(six_lock_wakeup_all);
 768
 769 /**
 770  * six_lock_counts - return held lock counts, for each lock type
 771  * @lock:       lock to return counters for
 772  *
 773  * Return: the number of times a lock is held for read, intent and write.
 774  */
 775 struct six_lock_count six_lock_counts(struct six_lock *lock)
 776 {
 777         struct six_lock_count ret;
 778
 779         ret.n[SIX_LOCK_read]    = !lock->readers
 780                 ? atomic_read(&lock->state) & SIX_LOCK_HELD_read
 781                 : pcpu_read_count(lock);
 782         ret.n[SIX_LOCK_intent]  = !!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent) +
 783                 lock->intent_lock_recurse;
 784         ret.n[SIX_LOCK_write]   = !!(atomic_read(&lock->state) & SIX_LOCK_HELD_write);
 785
 786         return ret;
 787 }
 788 EXPORT_SYMBOL_GPL(six_lock_counts);
 789
 790 /**
 791  * six_lock_readers_add - directly manipulate reader count of a lock
 792  * @lock:       lock to add/subtract readers for
 793  * @nr:         reader count to add/subtract
 794  *
 795  * When an upper layer is implementing lock reentrency, we may have both read
 796  * and intent locks on the same lock.
 797  *
 798  * When we need to take a write lock, the read locks will cause self-deadlock,
 799  * because six locks themselves do not track which read locks are held by the
 800  * current thread and which are held by a different thread - it does no
 801  * per-thread tracking of held locks.
 802  *
 803  * The upper layer that is tracking held locks may however, if trylock() has
 804  * failed, count up its own read locks, subtract them, take the write lock, and
 805  * then re-add them.
 806  *
 807  * As in any other situation when taking a write lock, @lock must be held for
 808  * intent one (or more) times, so @lock will never be left unlocked.
 809  */
 810 void six_lock_readers_add(struct six_lock *lock, int nr)
 811 {
 812         if (lock->readers) {
 813                 this_cpu_add(*lock->readers, nr);
 814         } else {
 815                 EBUG_ON((int) (atomic_read(&lock->state) & SIX_LOCK_HELD_read) + nr < 0);
 816                 /* reader count starts at bit 0 */
 817                 atomic_add(nr, &lock->state);
 818         }
 819 }
 820 EXPORT_SYMBOL_GPL(six_lock_readers_add);
 821
 822 /**
 823  * six_lock_exit - release resources held by a lock prior to freeing
 824  * @lock:       lock to exit
 825  *
 826  * When a lock was initialized in percpu mode (SIX_OLCK_INIT_PCPU), this is
 827  * required to free the percpu read counts.
 828  */
 829 void six_lock_exit(struct six_lock *lock)
 830 {
 831         WARN_ON(lock->readers && pcpu_read_count(lock));
 832         WARN_ON(atomic_read(&lock->state) & SIX_LOCK_HELD_read);
 833
 834         free_percpu(lock->readers);
 835         lock->readers = NULL;
 836 }
 837 EXPORT_SYMBOL_GPL(six_lock_exit);
 838
 839 void __six_lock_init(struct six_lock *lock, const char *name,
 840                      struct lock_class_key *key, enum six_lock_init_flags flags)
 841 {
 842         atomic_set(&lock->state, 0);
 843         raw_spin_lock_init(&lock->wait_lock);
 844         INIT_LIST_HEAD(&lock->wait_list);
 845 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 846         debug_check_no_locks_freed((void *) lock, sizeof(*lock));
 847         lockdep_init_map(&lock->dep_map, name, key, 0);
 848 #endif
 849
 850         /*
 851          * Don't assume that we have real percpu variables available in
 852          * userspace:
 853          */
 854 #ifdef __KERNEL__
 855         if (flags & SIX_LOCK_INIT_PCPU) {
 856                 /*
 857                  * We don't return an error here on memory allocation failure
 858                  * since percpu is an optimization, and locks will work with the
 859                  * same semantics in non-percpu mode: callers can check for
 860                  * failure if they wish by checking lock->readers, but generally
 861                  * will not want to treat it as an error.
 862                  */
 863                 lock->readers = alloc_percpu(unsigned);
 864         }
 865 #endif
 866 }
 867 EXPORT_SYMBOL_GPL(__six_lock_init);