Linux Audio

Check our new training course

Loading...
v4.10.11
 
  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 <asm/byteorder.h>
 32#include <asm/qspinlock.h>
 33
 34/*
 
 
 
 
 
 35 * The basic principle of a queue-based spinlock can best be understood
 36 * by studying a classic queue-based spinlock implementation called the
 37 * MCS lock. The paper below provides a good description for this kind
 38 * of lock.
 39 *
 40 * http://www.cise.ufl.edu/tr/DOC/REP-1992-71.pdf
 41 *
 42 * This queued spinlock implementation is based on the MCS lock, however to make
 43 * it fit the 4 bytes we assume spinlock_t to be, and preserve its existing
 44 * API, we must modify it somehow.
 
 45 *
 46 * In particular; where the traditional MCS lock consists of a tail pointer
 47 * (8 bytes) and needs the next pointer (another 8 bytes) of its own node to
 48 * unlock the next pending (next->locked), we compress both these: {tail,
 49 * next->locked} into a single u32 value.
 50 *
 51 * Since a spinlock disables recursion of its own context and there is a limit
 52 * to the contexts that can nest; namely: task, softirq, hardirq, nmi. As there
 53 * are at most 4 nesting levels, it can be encoded by a 2-bit number. Now
 54 * we can encode the tail by combining the 2-bit nesting level with the cpu
 55 * number. With one byte for the lock value and 3 bytes for the tail, only a
 56 * 32-bit word is now needed. Even though we only need 1 bit for the lock,
 57 * we extend it to a full byte to achieve better performance for architectures
 58 * that support atomic byte write.
 59 *
 60 * We also change the first spinner to spin on the lock bit instead of its
 61 * node; whereby avoiding the need to carry a node from lock to unlock, and
 62 * preserving existing lock API. This also makes the unlock code simpler and
 63 * faster.
 64 *
 65 * N.B. The current implementation only supports architectures that allow
 66 *      atomic operations on smaller 8-bit and 16-bit data types.
 67 *
 68 */
 69
 70#include "mcs_spinlock.h"
 
 71
 
 
 
 
 
 
 
 
 
 
 
 
 72#ifdef CONFIG_PARAVIRT_SPINLOCKS
 73#define MAX_NODES	8
 74#else
 75#define MAX_NODES	4
 
 
 
 
 
 
 
 
 
 
 
 76#endif
 77
 78/*
 79 * Per-CPU queue node structures; we can never have more than 4 nested
 80 * contexts: task, softirq, hardirq, nmi.
 81 *
 82 * Exactly fits one 64-byte cacheline on a 64-bit architecture.
 83 *
 84 * PV doubles the storage and uses the second cacheline for PV state.
 85 */
 86static DEFINE_PER_CPU_ALIGNED(struct mcs_spinlock, mcs_nodes[MAX_NODES]);
 87
 88/*
 89 * We must be able to distinguish between no-tail and the tail at 0:0,
 90 * therefore increment the cpu number by one.
 91 */
 92
 93static inline __pure u32 encode_tail(int cpu, int idx)
 94{
 95	u32 tail;
 96
 97#ifdef CONFIG_DEBUG_SPINLOCK
 98	BUG_ON(idx > 3);
 99#endif
100	tail  = (cpu + 1) << _Q_TAIL_CPU_OFFSET;
101	tail |= idx << _Q_TAIL_IDX_OFFSET; /* assume < 4 */
102
103	return tail;
104}
105
106static inline __pure struct mcs_spinlock *decode_tail(u32 tail)
107{
108	int cpu = (tail >> _Q_TAIL_CPU_OFFSET) - 1;
109	int idx = (tail &  _Q_TAIL_IDX_MASK) >> _Q_TAIL_IDX_OFFSET;
110
111	return per_cpu_ptr(&mcs_nodes[idx], cpu);
112}
113
114#define _Q_LOCKED_PENDING_MASK (_Q_LOCKED_MASK | _Q_PENDING_MASK)
 
 
 
 
115
116/*
117 * By using the whole 2nd least significant byte for the pending bit, we
118 * can allow better optimization of the lock acquisition for the pending
119 * bit holder.
120 *
121 * This internal structure is also used by the set_locked function which
122 * is not restricted to _Q_PENDING_BITS == 8.
123 */
124struct __qspinlock {
125	union {
126		atomic_t val;
127#ifdef __LITTLE_ENDIAN
128		struct {
129			u8	locked;
130			u8	pending;
131		};
132		struct {
133			u16	locked_pending;
134			u16	tail;
135		};
136#else
137		struct {
138			u16	tail;
139			u16	locked_pending;
140		};
141		struct {
142			u8	reserved[2];
143			u8	pending;
144			u8	locked;
145		};
146#endif
147	};
148};
149
150#if _Q_PENDING_BITS == 8
151/**
 
 
 
 
 
 
 
 
 
 
 
152 * clear_pending_set_locked - take ownership and clear the pending bit.
153 * @lock: Pointer to queued spinlock structure
154 *
155 * *,1,0 -> *,0,1
156 *
157 * Lock stealing is not allowed if this function is used.
158 */
159static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
160{
161	struct __qspinlock *l = (void *)lock;
162
163	WRITE_ONCE(l->locked_pending, _Q_LOCKED_VAL);
164}
165
166/*
167 * xchg_tail - Put in the new queue tail code word & retrieve previous one
168 * @lock : Pointer to queued spinlock structure
169 * @tail : The new queue tail code word
170 * Return: The previous queue tail code word
171 *
172 * xchg(lock, tail)
173 *
174 * p,*,* -> n,*,* ; prev = xchg(lock, node)
175 */
176static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
177{
178	struct __qspinlock *l = (void *)lock;
179
180	/*
181	 * Use release semantics to make sure that the MCS node is properly
182	 * initialized before changing the tail code.
183	 */
184	return (u32)xchg_release(&l->tail,
185				 tail >> _Q_TAIL_OFFSET) << _Q_TAIL_OFFSET;
186}
187
188#else /* _Q_PENDING_BITS == 8 */
189
190/**
 
 
 
 
 
 
 
 
 
 
 
191 * clear_pending_set_locked - take ownership and clear the pending bit.
192 * @lock: Pointer to queued spinlock structure
193 *
194 * *,1,0 -> *,0,1
195 */
196static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
197{
198	atomic_add(-_Q_PENDING_VAL + _Q_LOCKED_VAL, &lock->val);
199}
200
201/**
202 * xchg_tail - Put in the new queue tail code word & retrieve previous one
203 * @lock : Pointer to queued spinlock structure
204 * @tail : The new queue tail code word
205 * Return: The previous queue tail code word
206 *
207 * xchg(lock, tail)
208 *
209 * p,*,* -> n,*,* ; prev = xchg(lock, node)
210 */
211static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
212{
213	u32 old, new, val = atomic_read(&lock->val);
214
215	for (;;) {
216		new = (val & _Q_LOCKED_PENDING_MASK) | tail;
217		/*
218		 * Use release semantics to make sure that the MCS node is
219		 * properly initialized before changing the tail code.
 
220		 */
221		old = atomic_cmpxchg_release(&lock->val, val, new);
222		if (old == val)
223			break;
224
225		val = old;
226	}
227	return old;
228}
229#endif /* _Q_PENDING_BITS == 8 */
230
231/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
232 * set_locked - Set the lock bit and own the lock
233 * @lock: Pointer to queued spinlock structure
234 *
235 * *,*,0 -> *,0,1
236 */
237static __always_inline void set_locked(struct qspinlock *lock)
238{
239	struct __qspinlock *l = (void *)lock;
240
241	WRITE_ONCE(l->locked, _Q_LOCKED_VAL);
242}
243
244
245/*
246 * Generate the native code for queued_spin_unlock_slowpath(); provide NOPs for
247 * all the PV callbacks.
248 */
249
250static __always_inline void __pv_init_node(struct mcs_spinlock *node) { }
251static __always_inline void __pv_wait_node(struct mcs_spinlock *node,
252					   struct mcs_spinlock *prev) { }
253static __always_inline void __pv_kick_node(struct qspinlock *lock,
254					   struct mcs_spinlock *node) { }
255static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
256						   struct mcs_spinlock *node)
257						   { return 0; }
258
259#define pv_enabled()		false
260
261#define pv_init_node		__pv_init_node
262#define pv_wait_node		__pv_wait_node
263#define pv_kick_node		__pv_kick_node
264#define pv_wait_head_or_lock	__pv_wait_head_or_lock
265
266#ifdef CONFIG_PARAVIRT_SPINLOCKS
267#define queued_spin_lock_slowpath	native_queued_spin_lock_slowpath
268#endif
269
270/*
271 * Various notes on spin_is_locked() and spin_unlock_wait(), which are
272 * 'interesting' functions:
273 *
274 * PROBLEM: some architectures have an interesting issue with atomic ACQUIRE
275 * operations in that the ACQUIRE applies to the LOAD _not_ the STORE (ARM64,
276 * PPC). Also qspinlock has a similar issue per construction, the setting of
277 * the locked byte can be unordered acquiring the lock proper.
278 *
279 * This gets to be 'interesting' in the following cases, where the /should/s
280 * end up false because of this issue.
281 *
282 *
283 * CASE 1:
284 *
285 * So the spin_is_locked() correctness issue comes from something like:
286 *
287 *   CPU0				CPU1
288 *
289 *   global_lock();			local_lock(i)
290 *     spin_lock(&G)			  spin_lock(&L[i])
291 *     for (i)				  if (!spin_is_locked(&G)) {
292 *       spin_unlock_wait(&L[i]);	    smp_acquire__after_ctrl_dep();
293 *					    return;
294 *					  }
295 *					  // deal with fail
296 *
297 * Where it is important CPU1 sees G locked or CPU0 sees L[i] locked such
298 * that there is exclusion between the two critical sections.
299 *
300 * The load from spin_is_locked(&G) /should/ be constrained by the ACQUIRE from
301 * spin_lock(&L[i]), and similarly the load(s) from spin_unlock_wait(&L[i])
302 * /should/ be constrained by the ACQUIRE from spin_lock(&G).
303 *
304 * Similarly, later stuff is constrained by the ACQUIRE from CTRL+RMB.
305 *
306 *
307 * CASE 2:
308 *
309 * For spin_unlock_wait() there is a second correctness issue, namely:
310 *
311 *   CPU0				CPU1
312 *
313 *   flag = set;
314 *   smp_mb();				spin_lock(&l)
315 *   spin_unlock_wait(&l);		if (!flag)
316 *					  // add to lockless list
317 *					spin_unlock(&l);
318 *   // iterate lockless list
319 *
320 * Which wants to ensure that CPU1 will stop adding bits to the list and CPU0
321 * will observe the last entry on the list (if spin_unlock_wait() had ACQUIRE
322 * semantics etc..)
323 *
324 * Where flag /should/ be ordered against the locked store of l.
325 */
326
327/*
328 * queued_spin_lock_slowpath() can (load-)ACQUIRE the lock before
329 * issuing an _unordered_ store to set _Q_LOCKED_VAL.
330 *
331 * This means that the store can be delayed, but no later than the
332 * store-release from the unlock. This means that simply observing
333 * _Q_LOCKED_VAL is not sufficient to determine if the lock is acquired.
334 *
335 * There are two paths that can issue the unordered store:
336 *
337 *  (1) clear_pending_set_locked():	*,1,0 -> *,0,1
338 *
339 *  (2) set_locked():			t,0,0 -> t,0,1 ; t != 0
340 *      atomic_cmpxchg_relaxed():	t,0,0 -> 0,0,1
341 *
342 * However, in both cases we have other !0 state we've set before to queue
343 * ourseves:
344 *
345 * For (1) we have the atomic_cmpxchg_acquire() that set _Q_PENDING_VAL, our
346 * load is constrained by that ACQUIRE to not pass before that, and thus must
347 * observe the store.
348 *
349 * For (2) we have a more intersting scenario. We enqueue ourselves using
350 * xchg_tail(), which ends up being a RELEASE. This in itself is not
351 * sufficient, however that is followed by an smp_cond_acquire() on the same
352 * word, giving a RELEASE->ACQUIRE ordering. This again constrains our load and
353 * guarantees we must observe that store.
354 *
355 * Therefore both cases have other !0 state that is observable before the
356 * unordered locked byte store comes through. This means we can use that to
357 * wait for the lock store, and then wait for an unlock.
358 */
359#ifndef queued_spin_unlock_wait
360void queued_spin_unlock_wait(struct qspinlock *lock)
361{
362	u32 val;
363
364	for (;;) {
365		val = atomic_read(&lock->val);
366
367		if (!val) /* not locked, we're done */
368			goto done;
369
370		if (val & _Q_LOCKED_MASK) /* locked, go wait for unlock */
371			break;
372
373		/* not locked, but pending, wait until we observe the lock */
374		cpu_relax();
375	}
376
377	/* any unlock is good */
378	while (atomic_read(&lock->val) & _Q_LOCKED_MASK)
379		cpu_relax();
380
381done:
382	smp_acquire__after_ctrl_dep();
383}
384EXPORT_SYMBOL(queued_spin_unlock_wait);
385#endif
386
387#endif /* _GEN_PV_LOCK_SLOWPATH */
388
389/**
390 * queued_spin_lock_slowpath - acquire the queued spinlock
391 * @lock: Pointer to queued spinlock structure
392 * @val: Current value of the queued spinlock 32-bit word
393 *
394 * (queue tail, pending bit, lock value)
395 *
396 *              fast     :    slow                                  :    unlock
397 *                       :                                          :
398 * uncontended  (0,0,0) -:--> (0,0,1) ------------------------------:--> (*,*,0)
399 *                       :       | ^--------.------.             /  :
400 *                       :       v           \      \            |  :
401 * pending               :    (0,1,1) +--> (0,1,0)   \           |  :
402 *                       :       | ^--'              |           |  :
403 *                       :       v                   |           |  :
404 * uncontended           :    (n,x,y) +--> (n,0,0) --'           |  :
405 *   queue               :       | ^--'                          |  :
406 *                       :       v                               |  :
407 * contended             :    (*,x,y) +--> (*,0,0) ---> (*,0,1) -'  :
408 *   queue               :         ^--'                             :
409 */
410void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
411{
412	struct mcs_spinlock *prev, *next, *node;
413	u32 new, old, tail;
414	int idx;
415
416	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
417
418	if (pv_enabled())
419		goto queue;
420
421	if (virt_spin_lock(lock))
422		return;
423
424	/*
425	 * wait for in-progress pending->locked hand-overs
 
426	 *
427	 * 0,1,0 -> 0,0,1
428	 */
429	if (val == _Q_PENDING_VAL) {
430		while ((val = atomic_read(&lock->val)) == _Q_PENDING_VAL)
431			cpu_relax();
 
432	}
433
434	/*
 
 
 
 
 
 
435	 * trylock || pending
436	 *
437	 * 0,0,0 -> 0,0,1 ; trylock
438	 * 0,0,1 -> 0,1,1 ; pending
439	 */
440	for (;;) {
441		/*
442		 * If we observe any contention; queue.
443		 */
444		if (val & ~_Q_LOCKED_MASK)
445			goto queue;
446
447		new = _Q_LOCKED_VAL;
448		if (val == new)
449			new |= _Q_PENDING_VAL;
 
 
 
 
 
450
451		/*
452		 * Acquire semantic is required here as the function may
453		 * return immediately if the lock was free.
454		 */
455		old = atomic_cmpxchg_acquire(&lock->val, val, new);
456		if (old == val)
457			break;
458
459		val = old;
460	}
461
462	/*
463	 * we won the trylock
464	 */
465	if (new == _Q_LOCKED_VAL)
466		return;
467
468	/*
469	 * we're pending, wait for the owner to go away.
470	 *
471	 * *,1,1 -> *,1,0
472	 *
473	 * this wait loop must be a load-acquire such that we match the
474	 * store-release that clears the locked bit and create lock
475	 * sequentiality; this is because not all clear_pending_set_locked()
476	 * implementations imply full barriers.
 
477	 */
478	smp_cond_load_acquire(&lock->val.counter, !(VAL & _Q_LOCKED_MASK));
 
479
480	/*
481	 * take ownership and clear the pending bit.
482	 *
483	 * *,1,0 -> *,0,1
484	 */
485	clear_pending_set_locked(lock);
 
486	return;
487
488	/*
489	 * End of pending bit optimistic spinning and beginning of MCS
490	 * queuing.
491	 */
492queue:
493	node = this_cpu_ptr(&mcs_nodes[0]);
 
 
494	idx = node->count++;
495	tail = encode_tail(smp_processor_id(), idx);
496
497	node += idx;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
498	node->locked = 0;
499	node->next = NULL;
500	pv_init_node(node);
501
502	/*
503	 * We touched a (possibly) cold cacheline in the per-cpu queue node;
504	 * attempt the trylock once more in the hope someone let go while we
505	 * weren't watching.
506	 */
507	if (queued_spin_trylock(lock))
508		goto release;
509
510	/*
 
 
 
 
 
 
 
 
511	 * We have already touched the queueing cacheline; don't bother with
512	 * pending stuff.
513	 *
514	 * p,*,* -> n,*,*
515	 *
516	 * RELEASE, such that the stores to @node must be complete.
517	 */
518	old = xchg_tail(lock, tail);
519	next = NULL;
520
521	/*
522	 * if there was a previous node; link it and wait until reaching the
523	 * head of the waitqueue.
524	 */
525	if (old & _Q_TAIL_MASK) {
526		prev = decode_tail(old);
527		/*
528		 * The above xchg_tail() is also a load of @lock which generates,
529		 * through decode_tail(), a pointer.
530		 *
531		 * The address dependency matches the RELEASE of xchg_tail()
532		 * such that the access to @prev must happen after.
533		 */
534		smp_read_barrier_depends();
535
 
536		WRITE_ONCE(prev->next, node);
537
538		pv_wait_node(node, prev);
539		arch_mcs_spin_lock_contended(&node->locked);
540
541		/*
542		 * While waiting for the MCS lock, the next pointer may have
543		 * been set by another lock waiter. We optimistically load
544		 * the next pointer & prefetch the cacheline for writing
545		 * to reduce latency in the upcoming MCS unlock operation.
546		 */
547		next = READ_ONCE(node->next);
548		if (next)
549			prefetchw(next);
550	}
551
552	/*
553	 * we're at the head of the waitqueue, wait for the owner & pending to
554	 * go away.
555	 *
556	 * *,x,y -> *,0,0
557	 *
558	 * this wait loop must use a load-acquire such that we match the
559	 * store-release that clears the locked bit and create lock
560	 * sequentiality; this is because the set_locked() function below
561	 * does not imply a full barrier.
562	 *
563	 * The PV pv_wait_head_or_lock function, if active, will acquire
564	 * the lock and return a non-zero value. So we have to skip the
565	 * smp_cond_load_acquire() call. As the next PV queue head hasn't been
566	 * designated yet, there is no way for the locked value to become
567	 * _Q_SLOW_VAL. So both the set_locked() and the
568	 * atomic_cmpxchg_relaxed() calls will be safe.
569	 *
570	 * If PV isn't active, 0 will be returned instead.
571	 *
572	 */
573	if ((val = pv_wait_head_or_lock(lock, node)))
574		goto locked;
575
576	val = smp_cond_load_acquire(&lock->val.counter, !(VAL & _Q_LOCKED_PENDING_MASK));
577
578locked:
579	/*
580	 * claim the lock:
581	 *
582	 * n,0,0 -> 0,0,1 : lock, uncontended
583	 * *,0,0 -> *,0,1 : lock, contended
584	 *
585	 * If the queue head is the only one in the queue (lock value == tail),
586	 * clear the tail code and grab the lock. Otherwise, we only need
587	 * to grab the lock.
588	 */
589	for (;;) {
590		/* In the PV case we might already have _Q_LOCKED_VAL set */
591		if ((val & _Q_TAIL_MASK) != tail) {
592			set_locked(lock);
593			break;
594		}
595		/*
596		 * The smp_cond_load_acquire() call above has provided the
597		 * necessary acquire semantics required for locking. At most
598		 * two iterations of this loop may be ran.
599		 */
600		old = atomic_cmpxchg_relaxed(&lock->val, val, _Q_LOCKED_VAL);
601		if (old == val)
602			goto release;	/* No contention */
603
604		val = old;
 
 
 
 
 
 
 
 
 
 
 
 
605	}
606
607	/*
 
 
 
 
 
 
 
608	 * contended path; wait for next if not observed yet, release.
609	 */
610	if (!next) {
611		while (!(next = READ_ONCE(node->next)))
612			cpu_relax();
613	}
614
615	arch_mcs_spin_unlock_contended(&next->locked);
616	pv_kick_node(lock, next);
617
618release:
619	/*
620	 * release the node
621	 */
622	__this_cpu_dec(mcs_nodes[0].count);
623}
624EXPORT_SYMBOL(queued_spin_lock_slowpath);
625
626/*
627 * Generate the paravirt code for queued_spin_unlock_slowpath().
628 */
629#if !defined(_GEN_PV_LOCK_SLOWPATH) && defined(CONFIG_PARAVIRT_SPINLOCKS)
630#define _GEN_PV_LOCK_SLOWPATH
631
632#undef  pv_enabled
633#define pv_enabled()	true
634
635#undef pv_init_node
636#undef pv_wait_node
637#undef pv_kick_node
638#undef pv_wait_head_or_lock
639
640#undef  queued_spin_lock_slowpath
641#define queued_spin_lock_slowpath	__pv_queued_spin_lock_slowpath
642
643#include "qspinlock_paravirt.h"
644#include "qspinlock.c"
645
 
 
 
 
 
 
 
646#endif
v5.9
  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 * Queued spinlock
  4 *
 
 
 
 
 
 
 
 
 
 
  5 * (C) Copyright 2013-2015 Hewlett-Packard Development Company, L.P.
  6 * (C) Copyright 2013-2014,2018 Red Hat, Inc.
  7 * (C) Copyright 2015 Intel Corp.
  8 * (C) Copyright 2015 Hewlett-Packard Enterprise Development LP
  9 *
 10 * Authors: Waiman Long <longman@redhat.com>
 11 *          Peter Zijlstra <peterz@infradead.org>
 12 */
 13
 14#ifndef _GEN_PV_LOCK_SLOWPATH
 15
 16#include <linux/smp.h>
 17#include <linux/bug.h>
 18#include <linux/cpumask.h>
 19#include <linux/percpu.h>
 20#include <linux/hardirq.h>
 21#include <linux/mutex.h>
 22#include <linux/prefetch.h>
 23#include <asm/byteorder.h>
 24#include <asm/qspinlock.h>
 25
 26/*
 27 * Include queued spinlock statistics code
 28 */
 29#include "qspinlock_stat.h"
 30
 31/*
 32 * The basic principle of a queue-based spinlock can best be understood
 33 * by studying a classic queue-based spinlock implementation called the
 34 * MCS lock. A copy of the original MCS lock paper ("Algorithms for Scalable
 35 * Synchronization on Shared-Memory Multiprocessors by Mellor-Crummey and
 36 * Scott") is available at
 37 *
 38 * https://bugzilla.kernel.org/show_bug.cgi?id=206115
 39 *
 40 * This queued spinlock implementation is based on the MCS lock, however to
 41 * make it fit the 4 bytes we assume spinlock_t to be, and preserve its
 42 * existing API, we must modify it somehow.
 43 *
 44 * In particular; where the traditional MCS lock consists of a tail pointer
 45 * (8 bytes) and needs the next pointer (another 8 bytes) of its own node to
 46 * unlock the next pending (next->locked), we compress both these: {tail,
 47 * next->locked} into a single u32 value.
 48 *
 49 * Since a spinlock disables recursion of its own context and there is a limit
 50 * to the contexts that can nest; namely: task, softirq, hardirq, nmi. As there
 51 * are at most 4 nesting levels, it can be encoded by a 2-bit number. Now
 52 * we can encode the tail by combining the 2-bit nesting level with the cpu
 53 * number. With one byte for the lock value and 3 bytes for the tail, only a
 54 * 32-bit word is now needed. Even though we only need 1 bit for the lock,
 55 * we extend it to a full byte to achieve better performance for architectures
 56 * that support atomic byte write.
 57 *
 58 * We also change the first spinner to spin on the lock bit instead of its
 59 * node; whereby avoiding the need to carry a node from lock to unlock, and
 60 * preserving existing lock API. This also makes the unlock code simpler and
 61 * faster.
 62 *
 63 * N.B. The current implementation only supports architectures that allow
 64 *      atomic operations on smaller 8-bit and 16-bit data types.
 65 *
 66 */
 67
 68#include "mcs_spinlock.h"
 69#define MAX_NODES	4
 70
 71/*
 72 * On 64-bit architectures, the mcs_spinlock structure will be 16 bytes in
 73 * size and four of them will fit nicely in one 64-byte cacheline. For
 74 * pvqspinlock, however, we need more space for extra data. To accommodate
 75 * that, we insert two more long words to pad it up to 32 bytes. IOW, only
 76 * two of them can fit in a cacheline in this case. That is OK as it is rare
 77 * to have more than 2 levels of slowpath nesting in actual use. We don't
 78 * want to penalize pvqspinlocks to optimize for a rare case in native
 79 * qspinlocks.
 80 */
 81struct qnode {
 82	struct mcs_spinlock mcs;
 83#ifdef CONFIG_PARAVIRT_SPINLOCKS
 84	long reserved[2];
 85#endif
 86};
 87
 88/*
 89 * The pending bit spinning loop count.
 90 * This heuristic is used to limit the number of lockword accesses
 91 * made by atomic_cond_read_relaxed when waiting for the lock to
 92 * transition out of the "== _Q_PENDING_VAL" state. We don't spin
 93 * indefinitely because there's no guarantee that we'll make forward
 94 * progress.
 95 */
 96#ifndef _Q_PENDING_LOOPS
 97#define _Q_PENDING_LOOPS	1
 98#endif
 99
100/*
101 * Per-CPU queue node structures; we can never have more than 4 nested
102 * contexts: task, softirq, hardirq, nmi.
103 *
104 * Exactly fits one 64-byte cacheline on a 64-bit architecture.
105 *
106 * PV doubles the storage and uses the second cacheline for PV state.
107 */
108static DEFINE_PER_CPU_ALIGNED(struct qnode, qnodes[MAX_NODES]);
109
110/*
111 * We must be able to distinguish between no-tail and the tail at 0:0,
112 * therefore increment the cpu number by one.
113 */
114
115static inline __pure u32 encode_tail(int cpu, int idx)
116{
117	u32 tail;
118
 
 
 
119	tail  = (cpu + 1) << _Q_TAIL_CPU_OFFSET;
120	tail |= idx << _Q_TAIL_IDX_OFFSET; /* assume < 4 */
121
122	return tail;
123}
124
125static inline __pure struct mcs_spinlock *decode_tail(u32 tail)
126{
127	int cpu = (tail >> _Q_TAIL_CPU_OFFSET) - 1;
128	int idx = (tail &  _Q_TAIL_IDX_MASK) >> _Q_TAIL_IDX_OFFSET;
129
130	return per_cpu_ptr(&qnodes[idx].mcs, cpu);
131}
132
133static inline __pure
134struct mcs_spinlock *grab_mcs_node(struct mcs_spinlock *base, int idx)
135{
136	return &((struct qnode *)base + idx)->mcs;
137}
138
139#define _Q_LOCKED_PENDING_MASK (_Q_LOCKED_MASK | _Q_PENDING_MASK)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
140
141#if _Q_PENDING_BITS == 8
142/**
143 * clear_pending - clear the pending bit.
144 * @lock: Pointer to queued spinlock structure
145 *
146 * *,1,* -> *,0,*
147 */
148static __always_inline void clear_pending(struct qspinlock *lock)
149{
150	WRITE_ONCE(lock->pending, 0);
151}
152
153/**
154 * clear_pending_set_locked - take ownership and clear the pending bit.
155 * @lock: Pointer to queued spinlock structure
156 *
157 * *,1,0 -> *,0,1
158 *
159 * Lock stealing is not allowed if this function is used.
160 */
161static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
162{
163	WRITE_ONCE(lock->locked_pending, _Q_LOCKED_VAL);
 
 
164}
165
166/*
167 * xchg_tail - Put in the new queue tail code word & retrieve previous one
168 * @lock : Pointer to queued spinlock structure
169 * @tail : The new queue tail code word
170 * Return: The previous queue tail code word
171 *
172 * xchg(lock, tail), which heads an address dependency
173 *
174 * p,*,* -> n,*,* ; prev = xchg(lock, node)
175 */
176static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
177{
 
 
178	/*
179	 * We can use relaxed semantics since the caller ensures that the
180	 * MCS node is properly initialized before updating the tail.
181	 */
182	return (u32)xchg_relaxed(&lock->tail,
183				 tail >> _Q_TAIL_OFFSET) << _Q_TAIL_OFFSET;
184}
185
186#else /* _Q_PENDING_BITS == 8 */
187
188/**
189 * clear_pending - clear the pending bit.
190 * @lock: Pointer to queued spinlock structure
191 *
192 * *,1,* -> *,0,*
193 */
194static __always_inline void clear_pending(struct qspinlock *lock)
195{
196	atomic_andnot(_Q_PENDING_VAL, &lock->val);
197}
198
199/**
200 * clear_pending_set_locked - take ownership and clear the pending bit.
201 * @lock: Pointer to queued spinlock structure
202 *
203 * *,1,0 -> *,0,1
204 */
205static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
206{
207	atomic_add(-_Q_PENDING_VAL + _Q_LOCKED_VAL, &lock->val);
208}
209
210/**
211 * xchg_tail - Put in the new queue tail code word & retrieve previous one
212 * @lock : Pointer to queued spinlock structure
213 * @tail : The new queue tail code word
214 * Return: The previous queue tail code word
215 *
216 * xchg(lock, tail)
217 *
218 * p,*,* -> n,*,* ; prev = xchg(lock, node)
219 */
220static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
221{
222	u32 old, new, val = atomic_read(&lock->val);
223
224	for (;;) {
225		new = (val & _Q_LOCKED_PENDING_MASK) | tail;
226		/*
227		 * We can use relaxed semantics since the caller ensures that
228		 * the MCS node is properly initialized before updating the
229		 * tail.
230		 */
231		old = atomic_cmpxchg_relaxed(&lock->val, val, new);
232		if (old == val)
233			break;
234
235		val = old;
236	}
237	return old;
238}
239#endif /* _Q_PENDING_BITS == 8 */
240
241/**
242 * queued_fetch_set_pending_acquire - fetch the whole lock value and set pending
243 * @lock : Pointer to queued spinlock structure
244 * Return: The previous lock value
245 *
246 * *,*,* -> *,1,*
247 */
248#ifndef queued_fetch_set_pending_acquire
249static __always_inline u32 queued_fetch_set_pending_acquire(struct qspinlock *lock)
250{
251	return atomic_fetch_or_acquire(_Q_PENDING_VAL, &lock->val);
252}
253#endif
254
255/**
256 * set_locked - Set the lock bit and own the lock
257 * @lock: Pointer to queued spinlock structure
258 *
259 * *,*,0 -> *,0,1
260 */
261static __always_inline void set_locked(struct qspinlock *lock)
262{
263	WRITE_ONCE(lock->locked, _Q_LOCKED_VAL);
 
 
264}
265
266
267/*
268 * Generate the native code for queued_spin_unlock_slowpath(); provide NOPs for
269 * all the PV callbacks.
270 */
271
272static __always_inline void __pv_init_node(struct mcs_spinlock *node) { }
273static __always_inline void __pv_wait_node(struct mcs_spinlock *node,
274					   struct mcs_spinlock *prev) { }
275static __always_inline void __pv_kick_node(struct qspinlock *lock,
276					   struct mcs_spinlock *node) { }
277static __always_inline u32  __pv_wait_head_or_lock(struct qspinlock *lock,
278						   struct mcs_spinlock *node)
279						   { return 0; }
280
281#define pv_enabled()		false
282
283#define pv_init_node		__pv_init_node
284#define pv_wait_node		__pv_wait_node
285#define pv_kick_node		__pv_kick_node
286#define pv_wait_head_or_lock	__pv_wait_head_or_lock
287
288#ifdef CONFIG_PARAVIRT_SPINLOCKS
289#define queued_spin_lock_slowpath	native_queued_spin_lock_slowpath
290#endif
291
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
292#endif /* _GEN_PV_LOCK_SLOWPATH */
293
294/**
295 * queued_spin_lock_slowpath - acquire the queued spinlock
296 * @lock: Pointer to queued spinlock structure
297 * @val: Current value of the queued spinlock 32-bit word
298 *
299 * (queue tail, pending bit, lock value)
300 *
301 *              fast     :    slow                                  :    unlock
302 *                       :                                          :
303 * uncontended  (0,0,0) -:--> (0,0,1) ------------------------------:--> (*,*,0)
304 *                       :       | ^--------.------.             /  :
305 *                       :       v           \      \            |  :
306 * pending               :    (0,1,1) +--> (0,1,0)   \           |  :
307 *                       :       | ^--'              |           |  :
308 *                       :       v                   |           |  :
309 * uncontended           :    (n,x,y) +--> (n,0,0) --'           |  :
310 *   queue               :       | ^--'                          |  :
311 *                       :       v                               |  :
312 * contended             :    (*,x,y) +--> (*,0,0) ---> (*,0,1) -'  :
313 *   queue               :         ^--'                             :
314 */
315void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
316{
317	struct mcs_spinlock *prev, *next, *node;
318	u32 old, tail;
319	int idx;
320
321	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
322
323	if (pv_enabled())
324		goto pv_queue;
325
326	if (virt_spin_lock(lock))
327		return;
328
329	/*
330	 * Wait for in-progress pending->locked hand-overs with a bounded
331	 * number of spins so that we guarantee forward progress.
332	 *
333	 * 0,1,0 -> 0,0,1
334	 */
335	if (val == _Q_PENDING_VAL) {
336		int cnt = _Q_PENDING_LOOPS;
337		val = atomic_cond_read_relaxed(&lock->val,
338					       (VAL != _Q_PENDING_VAL) || !cnt--);
339	}
340
341	/*
342	 * If we observe any contention; queue.
343	 */
344	if (val & ~_Q_LOCKED_MASK)
345		goto queue;
346
347	/*
348	 * trylock || pending
349	 *
350	 * 0,0,* -> 0,1,* -> 0,0,1 pending, trylock
 
351	 */
352	val = queued_fetch_set_pending_acquire(lock);
 
 
 
 
 
353
354	/*
355	 * If we observe contention, there is a concurrent locker.
356	 *
357	 * Undo and queue; our setting of PENDING might have made the
358	 * n,0,0 -> 0,0,0 transition fail and it will now be waiting
359	 * on @next to become !NULL.
360	 */
361	if (unlikely(val & ~_Q_LOCKED_MASK)) {
362
363		/* Undo PENDING if we set it. */
364		if (!(val & _Q_PENDING_MASK))
365			clear_pending(lock);
 
 
 
 
366
367		goto queue;
368	}
369
370	/*
371	 * We're pending, wait for the owner to go away.
 
 
 
 
 
 
372	 *
373	 * 0,1,1 -> 0,1,0
374	 *
375	 * this wait loop must be a load-acquire such that we match the
376	 * store-release that clears the locked bit and create lock
377	 * sequentiality; this is because not all
378	 * clear_pending_set_locked() implementations imply full
379	 * barriers.
380	 */
381	if (val & _Q_LOCKED_MASK)
382		atomic_cond_read_acquire(&lock->val, !(VAL & _Q_LOCKED_MASK));
383
384	/*
385	 * take ownership and clear the pending bit.
386	 *
387	 * 0,1,0 -> 0,0,1
388	 */
389	clear_pending_set_locked(lock);
390	lockevent_inc(lock_pending);
391	return;
392
393	/*
394	 * End of pending bit optimistic spinning and beginning of MCS
395	 * queuing.
396	 */
397queue:
398	lockevent_inc(lock_slowpath);
399pv_queue:
400	node = this_cpu_ptr(&qnodes[0].mcs);
401	idx = node->count++;
402	tail = encode_tail(smp_processor_id(), idx);
403
404	/*
405	 * 4 nodes are allocated based on the assumption that there will
406	 * not be nested NMIs taking spinlocks. That may not be true in
407	 * some architectures even though the chance of needing more than
408	 * 4 nodes will still be extremely unlikely. When that happens,
409	 * we fall back to spinning on the lock directly without using
410	 * any MCS node. This is not the most elegant solution, but is
411	 * simple enough.
412	 */
413	if (unlikely(idx >= MAX_NODES)) {
414		lockevent_inc(lock_no_node);
415		while (!queued_spin_trylock(lock))
416			cpu_relax();
417		goto release;
418	}
419
420	node = grab_mcs_node(node, idx);
421
422	/*
423	 * Keep counts of non-zero index values:
424	 */
425	lockevent_cond_inc(lock_use_node2 + idx - 1, idx);
426
427	/*
428	 * Ensure that we increment the head node->count before initialising
429	 * the actual node. If the compiler is kind enough to reorder these
430	 * stores, then an IRQ could overwrite our assignments.
431	 */
432	barrier();
433
434	node->locked = 0;
435	node->next = NULL;
436	pv_init_node(node);
437
438	/*
439	 * We touched a (possibly) cold cacheline in the per-cpu queue node;
440	 * attempt the trylock once more in the hope someone let go while we
441	 * weren't watching.
442	 */
443	if (queued_spin_trylock(lock))
444		goto release;
445
446	/*
447	 * Ensure that the initialisation of @node is complete before we
448	 * publish the updated tail via xchg_tail() and potentially link
449	 * @node into the waitqueue via WRITE_ONCE(prev->next, node) below.
450	 */
451	smp_wmb();
452
453	/*
454	 * Publish the updated tail.
455	 * We have already touched the queueing cacheline; don't bother with
456	 * pending stuff.
457	 *
458	 * p,*,* -> n,*,*
 
 
459	 */
460	old = xchg_tail(lock, tail);
461	next = NULL;
462
463	/*
464	 * if there was a previous node; link it and wait until reaching the
465	 * head of the waitqueue.
466	 */
467	if (old & _Q_TAIL_MASK) {
468		prev = decode_tail(old);
 
 
 
 
 
 
 
 
469
470		/* Link @node into the waitqueue. */
471		WRITE_ONCE(prev->next, node);
472
473		pv_wait_node(node, prev);
474		arch_mcs_spin_lock_contended(&node->locked);
475
476		/*
477		 * While waiting for the MCS lock, the next pointer may have
478		 * been set by another lock waiter. We optimistically load
479		 * the next pointer & prefetch the cacheline for writing
480		 * to reduce latency in the upcoming MCS unlock operation.
481		 */
482		next = READ_ONCE(node->next);
483		if (next)
484			prefetchw(next);
485	}
486
487	/*
488	 * we're at the head of the waitqueue, wait for the owner & pending to
489	 * go away.
490	 *
491	 * *,x,y -> *,0,0
492	 *
493	 * this wait loop must use a load-acquire such that we match the
494	 * store-release that clears the locked bit and create lock
495	 * sequentiality; this is because the set_locked() function below
496	 * does not imply a full barrier.
497	 *
498	 * The PV pv_wait_head_or_lock function, if active, will acquire
499	 * the lock and return a non-zero value. So we have to skip the
500	 * atomic_cond_read_acquire() call. As the next PV queue head hasn't
501	 * been designated yet, there is no way for the locked value to become
502	 * _Q_SLOW_VAL. So both the set_locked() and the
503	 * atomic_cmpxchg_relaxed() calls will be safe.
504	 *
505	 * If PV isn't active, 0 will be returned instead.
506	 *
507	 */
508	if ((val = pv_wait_head_or_lock(lock, node)))
509		goto locked;
510
511	val = atomic_cond_read_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK));
512
513locked:
514	/*
515	 * claim the lock:
516	 *
517	 * n,0,0 -> 0,0,1 : lock, uncontended
518	 * *,*,0 -> *,*,1 : lock, contended
519	 *
520	 * If the queue head is the only one in the queue (lock value == tail)
521	 * and nobody is pending, clear the tail code and grab the lock.
522	 * Otherwise, we only need to grab the lock.
523	 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
524
525	/*
526	 * In the PV case we might already have _Q_LOCKED_VAL set, because
527	 * of lock stealing; therefore we must also allow:
528	 *
529	 * n,0,1 -> 0,0,1
530	 *
531	 * Note: at this point: (val & _Q_PENDING_MASK) == 0, because of the
532	 *       above wait condition, therefore any concurrent setting of
533	 *       PENDING will make the uncontended transition fail.
534	 */
535	if ((val & _Q_TAIL_MASK) == tail) {
536		if (atomic_try_cmpxchg_relaxed(&lock->val, &val, _Q_LOCKED_VAL))
537			goto release; /* No contention */
538	}
539
540	/*
541	 * Either somebody is queued behind us or _Q_PENDING_VAL got set
542	 * which will then detect the remaining tail and queue behind us
543	 * ensuring we'll see a @next.
544	 */
545	set_locked(lock);
546
547	/*
548	 * contended path; wait for next if not observed yet, release.
549	 */
550	if (!next)
551		next = smp_cond_load_relaxed(&node->next, (VAL));
 
 
552
553	arch_mcs_spin_unlock_contended(&next->locked);
554	pv_kick_node(lock, next);
555
556release:
557	/*
558	 * release the node
559	 */
560	__this_cpu_dec(qnodes[0].mcs.count);
561}
562EXPORT_SYMBOL(queued_spin_lock_slowpath);
563
564/*
565 * Generate the paravirt code for queued_spin_unlock_slowpath().
566 */
567#if !defined(_GEN_PV_LOCK_SLOWPATH) && defined(CONFIG_PARAVIRT_SPINLOCKS)
568#define _GEN_PV_LOCK_SLOWPATH
569
570#undef  pv_enabled
571#define pv_enabled()	true
572
573#undef pv_init_node
574#undef pv_wait_node
575#undef pv_kick_node
576#undef pv_wait_head_or_lock
577
578#undef  queued_spin_lock_slowpath
579#define queued_spin_lock_slowpath	__pv_queued_spin_lock_slowpath
580
581#include "qspinlock_paravirt.h"
582#include "qspinlock.c"
583
584bool nopvspin __initdata;
585static __init int parse_nopvspin(char *arg)
586{
587	nopvspin = true;
588	return 0;
589}
590early_param("nopvspin", parse_nopvspin);
591#endif