kernel/locking/qspinlock.c

   1 /*
   2  * Queued spinlock
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * (C) Copyright 2013-2015 Hewlett-Packard Development Company, L.P.
  15  * (C) Copyright 2013-2014 Red Hat, Inc.
  16  * (C) Copyright 2015 Intel Corp.
  17  * (C) Copyright 2015 Hewlett-Packard Enterprise Development LP
  18  *
  19  * Authors: Waiman Long <waiman.long@hpe.com>
  20  *          Peter Zijlstra <peterz@infradead.org>
  21  */
  22
  23 #ifndef _GEN_PV_LOCK_SLOWPATH
  24
  25 #include <linux/smp.h>
  26 #include <linux/bug.h>
  27 #include <linux/cpumask.h>
  28 #include <linux/percpu.h>
  29 #include <linux/hardirq.h>
  30 #include <linux/mutex.h>
  31 #include <linux/prefetch.h>
  32 #include <asm/byteorder.h>
  33 #include <asm/qspinlock.h>
  34
  35 /*
  36  * The basic principle of a queue-based spinlock can best be understood
  37  * by studying a classic queue-based spinlock implementation called the
  38  * MCS lock. The paper below provides a good description for this kind
  39  * of lock.
  40  *
  41  * http://www.cise.ufl.edu/tr/DOC/REP-1992-71.pdf
  42  *
  43  * This queued spinlock implementation is based on the MCS lock, however to make
  44  * it fit the 4 bytes we assume spinlock_t to be, and preserve its existing
  45  * API, we must modify it somehow.
  46  *
  47  * In particular; where the traditional MCS lock consists of a tail pointer
  48  * (8 bytes) and needs the next pointer (another 8 bytes) of its own node to
  49  * unlock the next pending (next->locked), we compress both these: {tail,
  50  * next->locked} into a single u32 value.
  51  *
  52  * Since a spinlock disables recursion of its own context and there is a limit
  53  * to the contexts that can nest; namely: task, softirq, hardirq, nmi. As there
  54  * are at most 4 nesting levels, it can be encoded by a 2-bit number. Now
  55  * we can encode the tail by combining the 2-bit nesting level with the cpu
  56  * number. With one byte for the lock value and 3 bytes for the tail, only a
  57  * 32-bit word is now needed. Even though we only need 1 bit for the lock,
  58  * we extend it to a full byte to achieve better performance for architectures
  59  * that support atomic byte write.
  60  *
  61  * We also change the first spinner to spin on the lock bit instead of its
  62  * node; whereby avoiding the need to carry a node from lock to unlock, and
  63  * preserving existing lock API. This also makes the unlock code simpler and
  64  * faster.
  65  *
  66  * N.B. The current implementation only supports architectures that allow
  67  *      atomic operations on smaller 8-bit and 16-bit data types.
  68  *
  69  */
  70
  71 #include "mcs_spinlock.h"
  72
  73 #ifdef CONFIG_PARAVIRT_SPINLOCKS
  74 #define MAX_NODES       8
  75 #else
  76 #define MAX_NODES       4
  77 #endif
  78
  79 /*
  80  * Per-CPU queue node structures; we can never have more than 4 nested
  81  * contexts: task, softirq, hardirq, nmi.
  82  *
  83  * Exactly fits one 64-byte cacheline on a 64-bit architecture.
  84  *
  85  * PV doubles the storage and uses the second cacheline for PV state.
  86  */
  87 static DEFINE_PER_CPU_ALIGNED(struct mcs_spinlock, mcs_nodes[MAX_NODES]);
  88
  89 /*
  90  * We must be able to distinguish between no-tail and the tail at 0:0,
  91  * therefore increment the cpu number by one.
  92  */
  93
  94 static inline __pure u32 encode_tail(int cpu, int idx)
  95 {
  96         u32 tail;
  97
  98 #ifdef CONFIG_DEBUG_SPINLOCK
  99         BUG_ON(idx > 3);
 100 #endif
 101         tail  = (cpu + 1) << _Q_TAIL_CPU_OFFSET;
 102         tail |= idx << _Q_TAIL_IDX_OFFSET; /* assume < 4 */
 103
 104         return tail;
 105 }
 106
 107 static inline __pure struct mcs_spinlock *decode_tail(u32 tail)
 108 {
 109         int cpu = (tail >> _Q_TAIL_CPU_OFFSET) - 1;
 110         int idx = (tail &  _Q_TAIL_IDX_MASK) >> _Q_TAIL_IDX_OFFSET;
 111
 112         return per_cpu_ptr(&mcs_nodes[idx], cpu);
 113 }
 114
 115 #define _Q_LOCKED_PENDING_MASK (_Q_LOCKED_MASK | _Q_PENDING_MASK)
 116
 117 /*
 118  * By using the whole 2nd least significant byte for the pending bit, we
 119  * can allow better optimization of the lock acquisition for the pending
 120  * bit holder.
 121  *
 122  * This internal structure is also used by the set_locked function which
 123  * is not restricted to _Q_PENDING_BITS == 8.
 124  */
 125 struct __qspinlock {
 126         union {
 127                 atomic_t val;
 128 #ifdef __LITTLE_ENDIAN
 129                 struct {
 130                         u8      locked;
 131                         u8      pending;
 132                 };
 133                 struct {
 134                         u16     locked_pending;
 135                         u16     tail;
 136                 };
 137 #else
 138                 struct {
 139                         u16     tail;
 140                         u16     locked_pending;
 141                 };
 142                 struct {
 143                         u8      reserved[2];
 144                         u8      pending;
 145                         u8      locked;
 146                 };
 147 #endif
 148         };
 149 };
 150
 151 #if _Q_PENDING_BITS == 8
 152 /**
 153  * clear_pending_set_locked - take ownership and clear the pending bit.
 154  * @lock: Pointer to queued spinlock structure
 155  *
 156  * *,1,0 -> *,0,1
 157  *
 158  * Lock stealing is not allowed if this function is used.
 159  */
 160 static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
 161 {
 162         struct __qspinlock *l = (void *)lock;
 163
 164         WRITE_ONCE(l->locked_pending, _Q_LOCKED_VAL);
 165 }
 166
 167 /*
 168  * xchg_tail - Put in the new queue tail code word & retrieve previous one
 169  * @lock : Pointer to queued spinlock structure
 170  * @tail : The new queue tail code word
 171  * Return: The previous queue tail code word
 172  *
 173  * xchg(lock, tail)
 174  *
 175  * p,*,* -> n,*,* ; prev = xchg(lock, node)
 176  */
 177 static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
 178 {
 179         struct __qspinlock *l = (void *)lock;
 180
 181         /*
 182          * Use release semantics to make sure that the MCS node is properly
 183          * initialized before changing the tail code.
 184          */
 185         return (u32)xchg_release(&l->tail,
 186                                  tail >> _Q_TAIL_OFFSET) << _Q_TAIL_OFFSET;
 187 }
 188
 189 #else /* _Q_PENDING_BITS == 8 */
 190
 191 /**
 192  * clear_pending_set_locked - take ownership and clear the pending bit.
 193  * @lock: Pointer to queued spinlock structure
 194  *
 195  * *,1,0 -> *,0,1
 196  */
 197 static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
 198 {
 199         atomic_add(-_Q_PENDING_VAL + _Q_LOCKED_VAL, &lock->val);
 200 }
 201
 202 /**
 203  * xchg_tail - Put in the new queue tail code word & retrieve previous one
 204  * @lock : Pointer to queued spinlock structure
 205  * @tail : The new queue tail code word
 206  * Return: The previous queue tail code word
 207  *
 208  * xchg(lock, tail)
 209  *
 210  * p,*,* -> n,*,* ; prev = xchg(lock, node)
 211  */
 212 static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
 213 {
 214         u32 old, new, val = atomic_read(&lock->val);
 215
 216         for (;;) {
 217                 new = (val & _Q_LOCKED_PENDING_MASK) | tail;
 218                 /*
 219                  * Use release semantics to make sure that the MCS node is
 220                  * properly initialized before changing the tail code.
 221                  */
 222                 old = atomic_cmpxchg_release(&lock->val, val, new);
 223                 if (old == val)
 224                         break;
 225
 226                 val = old;
 227         }
 228         return old;
 229 }
 230 #endif /* _Q_PENDING_BITS == 8 */
 231
 232 /**
 233  * set_locked - Set the lock bit and own the lock
 234  * @lock: Pointer to queued spinlock structure
 235  *
 236  * *,*,0 -> *,0,1
 237  */
 238 static __always_inline void set_locked(struct qspinlock *lock)
 239 {
 240         struct __qspinlock *l = (void *)lock;
 241
 242         WRITE_ONCE(l->locked, _Q_LOCKED_VAL);
 243 }
 244
 245
 246 /*
 247  * Generate the native code for queued_spin_unlock_slowpath(); provide NOPs for
 248  * all the PV callbacks.
 249  */
 250
 251 static __always_inline void __pv_init_node(struct mcs_spinlock *node) { }
 252 static __always_inline void __pv_wait_node(struct mcs_spinlock *node,
 253                                            struct mcs_spinlock *prev) { }
 254 static __always_inline void __pv_kick_node(struct qspinlock *lock,
 255                                            struct mcs_spinlock *node) { }
 256 static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
 257                                                    struct mcs_spinlock *node)
 258                                                    { return 0; }
 259
 260 #define pv_enabled()            false
 261
 262 #define pv_init_node            __pv_init_node
 263 #define pv_wait_node            __pv_wait_node
 264 #define pv_kick_node            __pv_kick_node
 265 #define pv_wait_head_or_lock    __pv_wait_head_or_lock
 266
 267 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 268 #define queued_spin_lock_slowpath       native_queued_spin_lock_slowpath
 269 #endif
 270
 271 #endif /* _GEN_PV_LOCK_SLOWPATH */
 272
 273 /**
 274  * queued_spin_lock_slowpath - acquire the queued spinlock
 275  * @lock: Pointer to queued spinlock structure
 276  * @val: Current value of the queued spinlock 32-bit word
 277  *
 278  * (queue tail, pending bit, lock value)
 279  *
 280  *              fast     :    slow                                  :    unlock
 281  *                       :                                          :
 282  * uncontended  (0,0,0) -:--> (0,0,1) ------------------------------:--> (*,*,0)
 283  *                       :       | ^--------.------.             /  :
 284  *                       :       v           \      \            |  :
 285  * pending               :    (0,1,1) +--> (0,1,0)   \           |  :
 286  *                       :       | ^--'              |           |  :
 287  *                       :       v                   |           |  :
 288  * uncontended           :    (n,x,y) +--> (n,0,0) --'           |  :
 289  *   queue               :       | ^--'                          |  :
 290  *                       :       v                               |  :
 291  * contended             :    (*,x,y) +--> (*,0,0) ---> (*,0,1) -'  :
 292  *   queue               :         ^--'                             :
 293  */
 294 void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
 295 {
 296         struct mcs_spinlock *prev, *next, *node;
 297         u32 new, old, tail;
 298         int idx;
 299
 300         BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
 301
 302         if (pv_enabled())
 303                 goto queue;
 304
 305         if (virt_spin_lock(lock))
 306                 return;
 307
 308         /*
 309          * wait for in-progress pending->locked hand-overs
 310          *
 311          * 0,1,0 -> 0,0,1
 312          */
 313         if (val == _Q_PENDING_VAL) {
 314                 while ((val = atomic_read(&lock->val)) == _Q_PENDING_VAL)
 315                         cpu_relax();
 316         }
 317
 318         /*
 319          * trylock || pending
 320          *
 321          * 0,0,0 -> 0,0,1 ; trylock
 322          * 0,0,1 -> 0,1,1 ; pending
 323          */
 324         for (;;) {
 325                 /*
 326                  * If we observe any contention; queue.
 327                  */
 328                 if (val & ~_Q_LOCKED_MASK)
 329                         goto queue;
 330
 331                 new = _Q_LOCKED_VAL;
 332                 if (val == new)
 333                         new |= _Q_PENDING_VAL;
 334
 335                 /*
 336                  * Acquire semantic is required here as the function may
 337                  * return immediately if the lock was free.
 338                  */
 339                 old = atomic_cmpxchg_acquire(&lock->val, val, new);
 340                 if (old == val)
 341                         break;
 342
 343                 val = old;
 344         }
 345
 346         /*
 347          * we won the trylock
 348          */
 349         if (new == _Q_LOCKED_VAL)
 350                 return;
 351
 352         /*
 353          * we're pending, wait for the owner to go away.
 354          *
 355          * *,1,1 -> *,1,0
 356          *
 357          * this wait loop must be a load-acquire such that we match the
 358          * store-release that clears the locked bit and create lock
 359          * sequentiality; this is because not all clear_pending_set_locked()
 360          * implementations imply full barriers.
 361          */
 362         smp_cond_load_acquire(&lock->val.counter, !(VAL & _Q_LOCKED_MASK));
 363
 364         /*
 365          * take ownership and clear the pending bit.
 366          *
 367          * *,1,0 -> *,0,1
 368          */
 369         clear_pending_set_locked(lock);
 370         return;
 371
 372         /*
 373          * End of pending bit optimistic spinning and beginning of MCS
 374          * queuing.
 375          */
 376 queue:
 377         node = this_cpu_ptr(&mcs_nodes[0]);
 378         idx = node->count++;
 379         tail = encode_tail(smp_processor_id(), idx);
 380
 381         node += idx;
 382
 383         /*
 384          * Ensure that we increment the head node->count before initialising
 385          * the actual node. If the compiler is kind enough to reorder these
 386          * stores, then an IRQ could overwrite our assignments.
 387          */
 388         barrier();
 389
 390         node->locked = 0;
 391         node->next = NULL;
 392         pv_init_node(node);
 393
 394         /*
 395          * We touched a (possibly) cold cacheline in the per-cpu queue node;
 396          * attempt the trylock once more in the hope someone let go while we
 397          * weren't watching.
 398          */
 399         if (queued_spin_trylock(lock))
 400                 goto release;
 401
 402         /*
 403          * We have already touched the queueing cacheline; don't bother with
 404          * pending stuff.
 405          *
 406          * p,*,* -> n,*,*
 407          *
 408          * RELEASE, such that the stores to @node must be complete.
 409          */
 410         old = xchg_tail(lock, tail);
 411         next = NULL;
 412
 413         /*
 414          * if there was a previous node; link it and wait until reaching the
 415          * head of the waitqueue.
 416          */
 417         if (old & _Q_TAIL_MASK) {
 418                 prev = decode_tail(old);
 419                 /*
 420                  * The above xchg_tail() is also a load of @lock which generates,
 421                  * through decode_tail(), a pointer.
 422                  *
 423                  * The address dependency matches the RELEASE of xchg_tail()
 424                  * such that the access to @prev must happen after.
 425                  */
 426                 smp_read_barrier_depends();
 427
 428                 WRITE_ONCE(prev->next, node);
 429
 430                 pv_wait_node(node, prev);
 431                 arch_mcs_spin_lock_contended(&node->locked);
 432
 433                 /*
 434                  * While waiting for the MCS lock, the next pointer may have
 435                  * been set by another lock waiter. We optimistically load
 436                  * the next pointer & prefetch the cacheline for writing
 437                  * to reduce latency in the upcoming MCS unlock operation.
 438                  */
 439                 next = READ_ONCE(node->next);
 440                 if (next)
 441                         prefetchw(next);
 442         }
 443
 444         /*
 445          * we're at the head of the waitqueue, wait for the owner & pending to
 446          * go away.
 447          *
 448          * *,x,y -> *,0,0
 449          *
 450          * this wait loop must use a load-acquire such that we match the
 451          * store-release that clears the locked bit and create lock
 452          * sequentiality; this is because the set_locked() function below
 453          * does not imply a full barrier.
 454          *
 455          * The PV pv_wait_head_or_lock function, if active, will acquire
 456          * the lock and return a non-zero value. So we have to skip the
 457          * smp_cond_load_acquire() call. As the next PV queue head hasn't been
 458          * designated yet, there is no way for the locked value to become
 459          * _Q_SLOW_VAL. So both the set_locked() and the
 460          * atomic_cmpxchg_relaxed() calls will be safe.
 461          *
 462          * If PV isn't active, 0 will be returned instead.
 463          *
 464          */
 465         if ((val = pv_wait_head_or_lock(lock, node)))
 466                 goto locked;
 467
 468         val = smp_cond_load_acquire(&lock->val.counter, !(VAL & _Q_LOCKED_PENDING_MASK));
 469
 470 locked:
 471         /*
 472          * claim the lock:
 473          *
 474          * n,0,0 -> 0,0,1 : lock, uncontended
 475          * *,0,0 -> *,0,1 : lock, contended
 476          *
 477          * If the queue head is the only one in the queue (lock value == tail),
 478          * clear the tail code and grab the lock. Otherwise, we only need
 479          * to grab the lock.
 480          */
 481         for (;;) {
 482                 /* In the PV case we might already have _Q_LOCKED_VAL set */
 483                 if ((val & _Q_TAIL_MASK) != tail) {
 484                         set_locked(lock);
 485                         break;
 486                 }
 487                 /*
 488                  * The smp_cond_load_acquire() call above has provided the
 489                  * necessary acquire semantics required for locking. At most
 490                  * two iterations of this loop may be ran.
 491                  */
 492                 old = atomic_cmpxchg_relaxed(&lock->val, val, _Q_LOCKED_VAL);
 493                 if (old == val)
 494                         goto release;   /* No contention */
 495
 496                 val = old;
 497         }
 498
 499         /*
 500          * contended path; wait for next if not observed yet, release.
 501          */
 502         if (!next) {
 503                 while (!(next = READ_ONCE(node->next)))
 504                         cpu_relax();
 505         }
 506
 507         arch_mcs_spin_unlock_contended(&next->locked);
 508         pv_kick_node(lock, next);
 509
 510 release:
 511         /*
 512          * release the node
 513          */
 514         __this_cpu_dec(mcs_nodes[0].count);
 515 }
 516 EXPORT_SYMBOL(queued_spin_lock_slowpath);
 517
 518 /*
 519  * Generate the paravirt code for queued_spin_unlock_slowpath().
 520  */
 521 #if !defined(_GEN_PV_LOCK_SLOWPATH) && defined(CONFIG_PARAVIRT_SPINLOCKS)
 522 #define _GEN_PV_LOCK_SLOWPATH
 523
 524 #undef  pv_enabled
 525 #define pv_enabled()    true
 526
 527 #undef pv_init_node
 528 #undef pv_wait_node
 529 #undef pv_kick_node
 530 #undef pv_wait_head_or_lock
 531
 532 #undef  queued_spin_lock_slowpath
 533 #define queued_spin_lock_slowpath       __pv_queued_spin_lock_slowpath
 534
 535 #include "qspinlock_paravirt.h"
 536 #include "qspinlock.c"
 537
 538 #endif