xref: /openbmc/linux/kernel/locking/qspinlock.c (revision 2c684d89)
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  *
18  * Authors: Waiman Long <waiman.long@hp.com>
19  *          Peter Zijlstra <peterz@infradead.org>
20  */
21 
22 #ifndef _GEN_PV_LOCK_SLOWPATH
23 
24 #include <linux/smp.h>
25 #include <linux/bug.h>
26 #include <linux/cpumask.h>
27 #include <linux/percpu.h>
28 #include <linux/hardirq.h>
29 #include <linux/mutex.h>
30 #include <asm/byteorder.h>
31 #include <asm/qspinlock.h>
32 
33 /*
34  * The basic principle of a queue-based spinlock can best be understood
35  * by studying a classic queue-based spinlock implementation called the
36  * MCS lock. The paper below provides a good description for this kind
37  * of lock.
38  *
39  * http://www.cise.ufl.edu/tr/DOC/REP-1992-71.pdf
40  *
41  * This queued spinlock implementation is based on the MCS lock, however to make
42  * it fit the 4 bytes we assume spinlock_t to be, and preserve its existing
43  * API, we must modify it somehow.
44  *
45  * In particular; where the traditional MCS lock consists of a tail pointer
46  * (8 bytes) and needs the next pointer (another 8 bytes) of its own node to
47  * unlock the next pending (next->locked), we compress both these: {tail,
48  * next->locked} into a single u32 value.
49  *
50  * Since a spinlock disables recursion of its own context and there is a limit
51  * to the contexts that can nest; namely: task, softirq, hardirq, nmi. As there
52  * are at most 4 nesting levels, it can be encoded by a 2-bit number. Now
53  * we can encode the tail by combining the 2-bit nesting level with the cpu
54  * number. With one byte for the lock value and 3 bytes for the tail, only a
55  * 32-bit word is now needed. Even though we only need 1 bit for the lock,
56  * we extend it to a full byte to achieve better performance for architectures
57  * that support atomic byte write.
58  *
59  * We also change the first spinner to spin on the lock bit instead of its
60  * node; whereby avoiding the need to carry a node from lock to unlock, and
61  * preserving existing lock API. This also makes the unlock code simpler and
62  * faster.
63  *
64  * N.B. The current implementation only supports architectures that allow
65  *      atomic operations on smaller 8-bit and 16-bit data types.
66  *
67  */
68 
69 #include "mcs_spinlock.h"
70 
71 #ifdef CONFIG_PARAVIRT_SPINLOCKS
72 #define MAX_NODES	8
73 #else
74 #define MAX_NODES	4
75 #endif
76 
77 /*
78  * Per-CPU queue node structures; we can never have more than 4 nested
79  * contexts: task, softirq, hardirq, nmi.
80  *
81  * Exactly fits one 64-byte cacheline on a 64-bit architecture.
82  *
83  * PV doubles the storage and uses the second cacheline for PV state.
84  */
85 static DEFINE_PER_CPU_ALIGNED(struct mcs_spinlock, mcs_nodes[MAX_NODES]);
86 
87 /*
88  * We must be able to distinguish between no-tail and the tail at 0:0,
89  * therefore increment the cpu number by one.
90  */
91 
92 static inline u32 encode_tail(int cpu, int idx)
93 {
94 	u32 tail;
95 
96 #ifdef CONFIG_DEBUG_SPINLOCK
97 	BUG_ON(idx > 3);
98 #endif
99 	tail  = (cpu + 1) << _Q_TAIL_CPU_OFFSET;
100 	tail |= idx << _Q_TAIL_IDX_OFFSET; /* assume < 4 */
101 
102 	return tail;
103 }
104 
105 static inline struct mcs_spinlock *decode_tail(u32 tail)
106 {
107 	int cpu = (tail >> _Q_TAIL_CPU_OFFSET) - 1;
108 	int idx = (tail &  _Q_TAIL_IDX_MASK) >> _Q_TAIL_IDX_OFFSET;
109 
110 	return per_cpu_ptr(&mcs_nodes[idx], cpu);
111 }
112 
113 #define _Q_LOCKED_PENDING_MASK (_Q_LOCKED_MASK | _Q_PENDING_MASK)
114 
115 /*
116  * By using the whole 2nd least significant byte for the pending bit, we
117  * can allow better optimization of the lock acquisition for the pending
118  * bit holder.
119  *
120  * This internal structure is also used by the set_locked function which
121  * is not restricted to _Q_PENDING_BITS == 8.
122  */
123 struct __qspinlock {
124 	union {
125 		atomic_t val;
126 #ifdef __LITTLE_ENDIAN
127 		struct {
128 			u8	locked;
129 			u8	pending;
130 		};
131 		struct {
132 			u16	locked_pending;
133 			u16	tail;
134 		};
135 #else
136 		struct {
137 			u16	tail;
138 			u16	locked_pending;
139 		};
140 		struct {
141 			u8	reserved[2];
142 			u8	pending;
143 			u8	locked;
144 		};
145 #endif
146 	};
147 };
148 
149 #if _Q_PENDING_BITS == 8
150 /**
151  * clear_pending_set_locked - take ownership and clear the pending bit.
152  * @lock: Pointer to queued spinlock structure
153  *
154  * *,1,0 -> *,0,1
155  *
156  * Lock stealing is not allowed if this function is used.
157  */
158 static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
159 {
160 	struct __qspinlock *l = (void *)lock;
161 
162 	WRITE_ONCE(l->locked_pending, _Q_LOCKED_VAL);
163 }
164 
165 /*
166  * xchg_tail - Put in the new queue tail code word & retrieve previous one
167  * @lock : Pointer to queued spinlock structure
168  * @tail : The new queue tail code word
169  * Return: The previous queue tail code word
170  *
171  * xchg(lock, tail)
172  *
173  * p,*,* -> n,*,* ; prev = xchg(lock, node)
174  */
175 static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
176 {
177 	struct __qspinlock *l = (void *)lock;
178 
179 	return (u32)xchg(&l->tail, tail >> _Q_TAIL_OFFSET) << _Q_TAIL_OFFSET;
180 }
181 
182 #else /* _Q_PENDING_BITS == 8 */
183 
184 /**
185  * clear_pending_set_locked - take ownership and clear the pending bit.
186  * @lock: Pointer to queued spinlock structure
187  *
188  * *,1,0 -> *,0,1
189  */
190 static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
191 {
192 	atomic_add(-_Q_PENDING_VAL + _Q_LOCKED_VAL, &lock->val);
193 }
194 
195 /**
196  * xchg_tail - Put in the new queue tail code word & retrieve previous one
197  * @lock : Pointer to queued spinlock structure
198  * @tail : The new queue tail code word
199  * Return: The previous queue tail code word
200  *
201  * xchg(lock, tail)
202  *
203  * p,*,* -> n,*,* ; prev = xchg(lock, node)
204  */
205 static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
206 {
207 	u32 old, new, val = atomic_read(&lock->val);
208 
209 	for (;;) {
210 		new = (val & _Q_LOCKED_PENDING_MASK) | tail;
211 		old = atomic_cmpxchg(&lock->val, val, new);
212 		if (old == val)
213 			break;
214 
215 		val = old;
216 	}
217 	return old;
218 }
219 #endif /* _Q_PENDING_BITS == 8 */
220 
221 /**
222  * set_locked - Set the lock bit and own the lock
223  * @lock: Pointer to queued spinlock structure
224  *
225  * *,*,0 -> *,0,1
226  */
227 static __always_inline void set_locked(struct qspinlock *lock)
228 {
229 	struct __qspinlock *l = (void *)lock;
230 
231 	WRITE_ONCE(l->locked, _Q_LOCKED_VAL);
232 }
233 
234 
235 /*
236  * Generate the native code for queued_spin_unlock_slowpath(); provide NOPs for
237  * all the PV callbacks.
238  */
239 
240 static __always_inline void __pv_init_node(struct mcs_spinlock *node) { }
241 static __always_inline void __pv_wait_node(struct mcs_spinlock *node) { }
242 static __always_inline void __pv_kick_node(struct qspinlock *lock,
243 					   struct mcs_spinlock *node) { }
244 static __always_inline void __pv_wait_head(struct qspinlock *lock,
245 					   struct mcs_spinlock *node) { }
246 
247 #define pv_enabled()		false
248 
249 #define pv_init_node		__pv_init_node
250 #define pv_wait_node		__pv_wait_node
251 #define pv_kick_node		__pv_kick_node
252 #define pv_wait_head		__pv_wait_head
253 
254 #ifdef CONFIG_PARAVIRT_SPINLOCKS
255 #define queued_spin_lock_slowpath	native_queued_spin_lock_slowpath
256 #endif
257 
258 #endif /* _GEN_PV_LOCK_SLOWPATH */
259 
260 /**
261  * queued_spin_lock_slowpath - acquire the queued spinlock
262  * @lock: Pointer to queued spinlock structure
263  * @val: Current value of the queued spinlock 32-bit word
264  *
265  * (queue tail, pending bit, lock value)
266  *
267  *              fast     :    slow                                  :    unlock
268  *                       :                                          :
269  * uncontended  (0,0,0) -:--> (0,0,1) ------------------------------:--> (*,*,0)
270  *                       :       | ^--------.------.             /  :
271  *                       :       v           \      \            |  :
272  * pending               :    (0,1,1) +--> (0,1,0)   \           |  :
273  *                       :       | ^--'              |           |  :
274  *                       :       v                   |           |  :
275  * uncontended           :    (n,x,y) +--> (n,0,0) --'           |  :
276  *   queue               :       | ^--'                          |  :
277  *                       :       v                               |  :
278  * contended             :    (*,x,y) +--> (*,0,0) ---> (*,0,1) -'  :
279  *   queue               :         ^--'                             :
280  */
281 void queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
282 {
283 	struct mcs_spinlock *prev, *next, *node;
284 	u32 new, old, tail;
285 	int idx;
286 
287 	BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
288 
289 	if (pv_enabled())
290 		goto queue;
291 
292 	if (virt_spin_lock(lock))
293 		return;
294 
295 	/*
296 	 * wait for in-progress pending->locked hand-overs
297 	 *
298 	 * 0,1,0 -> 0,0,1
299 	 */
300 	if (val == _Q_PENDING_VAL) {
301 		while ((val = atomic_read(&lock->val)) == _Q_PENDING_VAL)
302 			cpu_relax();
303 	}
304 
305 	/*
306 	 * trylock || pending
307 	 *
308 	 * 0,0,0 -> 0,0,1 ; trylock
309 	 * 0,0,1 -> 0,1,1 ; pending
310 	 */
311 	for (;;) {
312 		/*
313 		 * If we observe any contention; queue.
314 		 */
315 		if (val & ~_Q_LOCKED_MASK)
316 			goto queue;
317 
318 		new = _Q_LOCKED_VAL;
319 		if (val == new)
320 			new |= _Q_PENDING_VAL;
321 
322 		old = atomic_cmpxchg(&lock->val, val, new);
323 		if (old == val)
324 			break;
325 
326 		val = old;
327 	}
328 
329 	/*
330 	 * we won the trylock
331 	 */
332 	if (new == _Q_LOCKED_VAL)
333 		return;
334 
335 	/*
336 	 * we're pending, wait for the owner to go away.
337 	 *
338 	 * *,1,1 -> *,1,0
339 	 *
340 	 * this wait loop must be a load-acquire such that we match the
341 	 * store-release that clears the locked bit and create lock
342 	 * sequentiality; this is because not all clear_pending_set_locked()
343 	 * implementations imply full barriers.
344 	 */
345 	while ((val = smp_load_acquire(&lock->val.counter)) & _Q_LOCKED_MASK)
346 		cpu_relax();
347 
348 	/*
349 	 * take ownership and clear the pending bit.
350 	 *
351 	 * *,1,0 -> *,0,1
352 	 */
353 	clear_pending_set_locked(lock);
354 	return;
355 
356 	/*
357 	 * End of pending bit optimistic spinning and beginning of MCS
358 	 * queuing.
359 	 */
360 queue:
361 	node = this_cpu_ptr(&mcs_nodes[0]);
362 	idx = node->count++;
363 	tail = encode_tail(smp_processor_id(), idx);
364 
365 	node += idx;
366 	node->locked = 0;
367 	node->next = NULL;
368 	pv_init_node(node);
369 
370 	/*
371 	 * We touched a (possibly) cold cacheline in the per-cpu queue node;
372 	 * attempt the trylock once more in the hope someone let go while we
373 	 * weren't watching.
374 	 */
375 	if (queued_spin_trylock(lock))
376 		goto release;
377 
378 	/*
379 	 * We have already touched the queueing cacheline; don't bother with
380 	 * pending stuff.
381 	 *
382 	 * p,*,* -> n,*,*
383 	 */
384 	old = xchg_tail(lock, tail);
385 
386 	/*
387 	 * if there was a previous node; link it and wait until reaching the
388 	 * head of the waitqueue.
389 	 */
390 	if (old & _Q_TAIL_MASK) {
391 		prev = decode_tail(old);
392 		WRITE_ONCE(prev->next, node);
393 
394 		pv_wait_node(node);
395 		arch_mcs_spin_lock_contended(&node->locked);
396 	}
397 
398 	/*
399 	 * we're at the head of the waitqueue, wait for the owner & pending to
400 	 * go away.
401 	 *
402 	 * *,x,y -> *,0,0
403 	 *
404 	 * this wait loop must use a load-acquire such that we match the
405 	 * store-release that clears the locked bit and create lock
406 	 * sequentiality; this is because the set_locked() function below
407 	 * does not imply a full barrier.
408 	 *
409 	 */
410 	pv_wait_head(lock, node);
411 	while ((val = smp_load_acquire(&lock->val.counter)) & _Q_LOCKED_PENDING_MASK)
412 		cpu_relax();
413 
414 	/*
415 	 * claim the lock:
416 	 *
417 	 * n,0,0 -> 0,0,1 : lock, uncontended
418 	 * *,0,0 -> *,0,1 : lock, contended
419 	 *
420 	 * If the queue head is the only one in the queue (lock value == tail),
421 	 * clear the tail code and grab the lock. Otherwise, we only need
422 	 * to grab the lock.
423 	 */
424 	for (;;) {
425 		if (val != tail) {
426 			set_locked(lock);
427 			break;
428 		}
429 		old = atomic_cmpxchg(&lock->val, val, _Q_LOCKED_VAL);
430 		if (old == val)
431 			goto release;	/* No contention */
432 
433 		val = old;
434 	}
435 
436 	/*
437 	 * contended path; wait for next, release.
438 	 */
439 	while (!(next = READ_ONCE(node->next)))
440 		cpu_relax();
441 
442 	arch_mcs_spin_unlock_contended(&next->locked);
443 	pv_kick_node(lock, next);
444 
445 release:
446 	/*
447 	 * release the node
448 	 */
449 	this_cpu_dec(mcs_nodes[0].count);
450 }
451 EXPORT_SYMBOL(queued_spin_lock_slowpath);
452 
453 /*
454  * Generate the paravirt code for queued_spin_unlock_slowpath().
455  */
456 #if !defined(_GEN_PV_LOCK_SLOWPATH) && defined(CONFIG_PARAVIRT_SPINLOCKS)
457 #define _GEN_PV_LOCK_SLOWPATH
458 
459 #undef  pv_enabled
460 #define pv_enabled()	true
461 
462 #undef pv_init_node
463 #undef pv_wait_node
464 #undef pv_kick_node
465 #undef pv_wait_head
466 
467 #undef  queued_spin_lock_slowpath
468 #define queued_spin_lock_slowpath	__pv_queued_spin_lock_slowpath
469 
470 #include "qspinlock_paravirt.h"
471 #include "qspinlock.c"
472 
473 #endif
474