xref: /linux/kernel/locking/qspinlock_paravirt.h (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _GEN_PV_LOCK_SLOWPATH
3 #error "do not include this file"
4 #endif
5 
6 #include <linux/hash.h>
7 #include <linux/bootmem.h>
8 #include <linux/debug_locks.h>
9 
10 /*
11  * Implement paravirt qspinlocks; the general idea is to halt the vcpus instead
12  * of spinning them.
13  *
14  * This relies on the architecture to provide two paravirt hypercalls:
15  *
16  *   pv_wait(u8 *ptr, u8 val) -- suspends the vcpu if *ptr == val
17  *   pv_kick(cpu)             -- wakes a suspended vcpu
18  *
19  * Using these we implement __pv_queued_spin_lock_slowpath() and
20  * __pv_queued_spin_unlock() to replace native_queued_spin_lock_slowpath() and
21  * native_queued_spin_unlock().
22  */
23 
24 #define _Q_SLOW_VAL	(3U << _Q_LOCKED_OFFSET)
25 
26 /*
27  * Queue Node Adaptive Spinning
28  *
29  * A queue node vCPU will stop spinning if the vCPU in the previous node is
30  * not running. The one lock stealing attempt allowed at slowpath entry
31  * mitigates the slight slowdown for non-overcommitted guest with this
32  * aggressive wait-early mechanism.
33  *
34  * The status of the previous node will be checked at fixed interval
35  * controlled by PV_PREV_CHECK_MASK. This is to ensure that we won't
36  * pound on the cacheline of the previous node too heavily.
37  */
38 #define PV_PREV_CHECK_MASK	0xff
39 
40 /*
41  * Queue node uses: vcpu_running & vcpu_halted.
42  * Queue head uses: vcpu_running & vcpu_hashed.
43  */
44 enum vcpu_state {
45 	vcpu_running = 0,
46 	vcpu_halted,		/* Used only in pv_wait_node */
47 	vcpu_hashed,		/* = pv_hash'ed + vcpu_halted */
48 };
49 
50 struct pv_node {
51 	struct mcs_spinlock	mcs;
52 	struct mcs_spinlock	__res[3];
53 
54 	int			cpu;
55 	u8			state;
56 };
57 
58 /*
59  * Hybrid PV queued/unfair lock
60  *
61  * By replacing the regular queued_spin_trylock() with the function below,
62  * it will be called once when a lock waiter enter the PV slowpath before
63  * being queued.
64  *
65  * The pending bit is set by the queue head vCPU of the MCS wait queue in
66  * pv_wait_head_or_lock() to signal that it is ready to spin on the lock.
67  * When that bit becomes visible to the incoming waiters, no lock stealing
68  * is allowed. The function will return immediately to make the waiters
69  * enter the MCS wait queue. So lock starvation shouldn't happen as long
70  * as the queued mode vCPUs are actively running to set the pending bit
71  * and hence disabling lock stealing.
72  *
73  * When the pending bit isn't set, the lock waiters will stay in the unfair
74  * mode spinning on the lock unless the MCS wait queue is empty. In this
75  * case, the lock waiters will enter the queued mode slowpath trying to
76  * become the queue head and set the pending bit.
77  *
78  * This hybrid PV queued/unfair lock combines the best attributes of a
79  * queued lock (no lock starvation) and an unfair lock (good performance
80  * on not heavily contended locks).
81  */
82 #define queued_spin_trylock(l)	pv_hybrid_queued_unfair_trylock(l)
83 static inline bool pv_hybrid_queued_unfair_trylock(struct qspinlock *lock)
84 {
85 	/*
86 	 * Stay in unfair lock mode as long as queued mode waiters are
87 	 * present in the MCS wait queue but the pending bit isn't set.
88 	 */
89 	for (;;) {
90 		int val = atomic_read(&lock->val);
91 
92 		if (!(val & _Q_LOCKED_PENDING_MASK) &&
93 		   (cmpxchg_acquire(&lock->locked, 0, _Q_LOCKED_VAL) == 0)) {
94 			qstat_inc(qstat_pv_lock_stealing, true);
95 			return true;
96 		}
97 		if (!(val & _Q_TAIL_MASK) || (val & _Q_PENDING_MASK))
98 			break;
99 
100 		cpu_relax();
101 	}
102 
103 	return false;
104 }
105 
106 /*
107  * The pending bit is used by the queue head vCPU to indicate that it
108  * is actively spinning on the lock and no lock stealing is allowed.
109  */
110 #if _Q_PENDING_BITS == 8
111 static __always_inline void set_pending(struct qspinlock *lock)
112 {
113 	WRITE_ONCE(lock->pending, 1);
114 }
115 
116 /*
117  * The pending bit check in pv_queued_spin_steal_lock() isn't a memory
118  * barrier. Therefore, an atomic cmpxchg_acquire() is used to acquire the
119  * lock just to be sure that it will get it.
120  */
121 static __always_inline int trylock_clear_pending(struct qspinlock *lock)
122 {
123 	return !READ_ONCE(lock->locked) &&
124 	       (cmpxchg_acquire(&lock->locked_pending, _Q_PENDING_VAL,
125 				_Q_LOCKED_VAL) == _Q_PENDING_VAL);
126 }
127 #else /* _Q_PENDING_BITS == 8 */
128 static __always_inline void set_pending(struct qspinlock *lock)
129 {
130 	atomic_or(_Q_PENDING_VAL, &lock->val);
131 }
132 
133 static __always_inline int trylock_clear_pending(struct qspinlock *lock)
134 {
135 	int val = atomic_read(&lock->val);
136 
137 	for (;;) {
138 		int old, new;
139 
140 		if (val  & _Q_LOCKED_MASK)
141 			break;
142 
143 		/*
144 		 * Try to clear pending bit & set locked bit
145 		 */
146 		old = val;
147 		new = (val & ~_Q_PENDING_MASK) | _Q_LOCKED_VAL;
148 		val = atomic_cmpxchg_acquire(&lock->val, old, new);
149 
150 		if (val == old)
151 			return 1;
152 	}
153 	return 0;
154 }
155 #endif /* _Q_PENDING_BITS == 8 */
156 
157 /*
158  * Lock and MCS node addresses hash table for fast lookup
159  *
160  * Hashing is done on a per-cacheline basis to minimize the need to access
161  * more than one cacheline.
162  *
163  * Dynamically allocate a hash table big enough to hold at least 4X the
164  * number of possible cpus in the system. Allocation is done on page
165  * granularity. So the minimum number of hash buckets should be at least
166  * 256 (64-bit) or 512 (32-bit) to fully utilize a 4k page.
167  *
168  * Since we should not be holding locks from NMI context (very rare indeed) the
169  * max load factor is 0.75, which is around the point where open addressing
170  * breaks down.
171  *
172  */
173 struct pv_hash_entry {
174 	struct qspinlock *lock;
175 	struct pv_node   *node;
176 };
177 
178 #define PV_HE_PER_LINE	(SMP_CACHE_BYTES / sizeof(struct pv_hash_entry))
179 #define PV_HE_MIN	(PAGE_SIZE / sizeof(struct pv_hash_entry))
180 
181 static struct pv_hash_entry *pv_lock_hash;
182 static unsigned int pv_lock_hash_bits __read_mostly;
183 
184 /*
185  * Allocate memory for the PV qspinlock hash buckets
186  *
187  * This function should be called from the paravirt spinlock initialization
188  * routine.
189  */
190 void __init __pv_init_lock_hash(void)
191 {
192 	int pv_hash_size = ALIGN(4 * num_possible_cpus(), PV_HE_PER_LINE);
193 
194 	if (pv_hash_size < PV_HE_MIN)
195 		pv_hash_size = PV_HE_MIN;
196 
197 	/*
198 	 * Allocate space from bootmem which should be page-size aligned
199 	 * and hence cacheline aligned.
200 	 */
201 	pv_lock_hash = alloc_large_system_hash("PV qspinlock",
202 					       sizeof(struct pv_hash_entry),
203 					       pv_hash_size, 0,
204 					       HASH_EARLY | HASH_ZERO,
205 					       &pv_lock_hash_bits, NULL,
206 					       pv_hash_size, pv_hash_size);
207 }
208 
209 #define for_each_hash_entry(he, offset, hash)						\
210 	for (hash &= ~(PV_HE_PER_LINE - 1), he = &pv_lock_hash[hash], offset = 0;	\
211 	     offset < (1 << pv_lock_hash_bits);						\
212 	     offset++, he = &pv_lock_hash[(hash + offset) & ((1 << pv_lock_hash_bits) - 1)])
213 
214 static struct qspinlock **pv_hash(struct qspinlock *lock, struct pv_node *node)
215 {
216 	unsigned long offset, hash = hash_ptr(lock, pv_lock_hash_bits);
217 	struct pv_hash_entry *he;
218 	int hopcnt = 0;
219 
220 	for_each_hash_entry(he, offset, hash) {
221 		hopcnt++;
222 		if (!cmpxchg(&he->lock, NULL, lock)) {
223 			WRITE_ONCE(he->node, node);
224 			qstat_hop(hopcnt);
225 			return &he->lock;
226 		}
227 	}
228 	/*
229 	 * Hard assume there is a free entry for us.
230 	 *
231 	 * This is guaranteed by ensuring every blocked lock only ever consumes
232 	 * a single entry, and since we only have 4 nesting levels per CPU
233 	 * and allocated 4*nr_possible_cpus(), this must be so.
234 	 *
235 	 * The single entry is guaranteed by having the lock owner unhash
236 	 * before it releases.
237 	 */
238 	BUG();
239 }
240 
241 static struct pv_node *pv_unhash(struct qspinlock *lock)
242 {
243 	unsigned long offset, hash = hash_ptr(lock, pv_lock_hash_bits);
244 	struct pv_hash_entry *he;
245 	struct pv_node *node;
246 
247 	for_each_hash_entry(he, offset, hash) {
248 		if (READ_ONCE(he->lock) == lock) {
249 			node = READ_ONCE(he->node);
250 			WRITE_ONCE(he->lock, NULL);
251 			return node;
252 		}
253 	}
254 	/*
255 	 * Hard assume we'll find an entry.
256 	 *
257 	 * This guarantees a limited lookup time and is itself guaranteed by
258 	 * having the lock owner do the unhash -- IFF the unlock sees the
259 	 * SLOW flag, there MUST be a hash entry.
260 	 */
261 	BUG();
262 }
263 
264 /*
265  * Return true if when it is time to check the previous node which is not
266  * in a running state.
267  */
268 static inline bool
269 pv_wait_early(struct pv_node *prev, int loop)
270 {
271 	if ((loop & PV_PREV_CHECK_MASK) != 0)
272 		return false;
273 
274 	return READ_ONCE(prev->state) != vcpu_running || vcpu_is_preempted(prev->cpu);
275 }
276 
277 /*
278  * Initialize the PV part of the mcs_spinlock node.
279  */
280 static void pv_init_node(struct mcs_spinlock *node)
281 {
282 	struct pv_node *pn = (struct pv_node *)node;
283 
284 	BUILD_BUG_ON(sizeof(struct pv_node) > 5*sizeof(struct mcs_spinlock));
285 
286 	pn->cpu = smp_processor_id();
287 	pn->state = vcpu_running;
288 }
289 
290 /*
291  * Wait for node->locked to become true, halt the vcpu after a short spin.
292  * pv_kick_node() is used to set _Q_SLOW_VAL and fill in hash table on its
293  * behalf.
294  */
295 static void pv_wait_node(struct mcs_spinlock *node, struct mcs_spinlock *prev)
296 {
297 	struct pv_node *pn = (struct pv_node *)node;
298 	struct pv_node *pp = (struct pv_node *)prev;
299 	int loop;
300 	bool wait_early;
301 
302 	for (;;) {
303 		for (wait_early = false, loop = SPIN_THRESHOLD; loop; loop--) {
304 			if (READ_ONCE(node->locked))
305 				return;
306 			if (pv_wait_early(pp, loop)) {
307 				wait_early = true;
308 				break;
309 			}
310 			cpu_relax();
311 		}
312 
313 		/*
314 		 * Order pn->state vs pn->locked thusly:
315 		 *
316 		 * [S] pn->state = vcpu_halted	  [S] next->locked = 1
317 		 *     MB			      MB
318 		 * [L] pn->locked		[RmW] pn->state = vcpu_hashed
319 		 *
320 		 * Matches the cmpxchg() from pv_kick_node().
321 		 */
322 		smp_store_mb(pn->state, vcpu_halted);
323 
324 		if (!READ_ONCE(node->locked)) {
325 			qstat_inc(qstat_pv_wait_node, true);
326 			qstat_inc(qstat_pv_wait_early, wait_early);
327 			pv_wait(&pn->state, vcpu_halted);
328 		}
329 
330 		/*
331 		 * If pv_kick_node() changed us to vcpu_hashed, retain that
332 		 * value so that pv_wait_head_or_lock() knows to not also try
333 		 * to hash this lock.
334 		 */
335 		cmpxchg(&pn->state, vcpu_halted, vcpu_running);
336 
337 		/*
338 		 * If the locked flag is still not set after wakeup, it is a
339 		 * spurious wakeup and the vCPU should wait again. However,
340 		 * there is a pretty high overhead for CPU halting and kicking.
341 		 * So it is better to spin for a while in the hope that the
342 		 * MCS lock will be released soon.
343 		 */
344 		qstat_inc(qstat_pv_spurious_wakeup, !READ_ONCE(node->locked));
345 	}
346 
347 	/*
348 	 * By now our node->locked should be 1 and our caller will not actually
349 	 * spin-wait for it. We do however rely on our caller to do a
350 	 * load-acquire for us.
351 	 */
352 }
353 
354 /*
355  * Called after setting next->locked = 1 when we're the lock owner.
356  *
357  * Instead of waking the waiters stuck in pv_wait_node() advance their state
358  * such that they're waiting in pv_wait_head_or_lock(), this avoids a
359  * wake/sleep cycle.
360  */
361 static void pv_kick_node(struct qspinlock *lock, struct mcs_spinlock *node)
362 {
363 	struct pv_node *pn = (struct pv_node *)node;
364 
365 	/*
366 	 * If the vCPU is indeed halted, advance its state to match that of
367 	 * pv_wait_node(). If OTOH this fails, the vCPU was running and will
368 	 * observe its next->locked value and advance itself.
369 	 *
370 	 * Matches with smp_store_mb() and cmpxchg() in pv_wait_node()
371 	 *
372 	 * The write to next->locked in arch_mcs_spin_unlock_contended()
373 	 * must be ordered before the read of pn->state in the cmpxchg()
374 	 * below for the code to work correctly. To guarantee full ordering
375 	 * irrespective of the success or failure of the cmpxchg(),
376 	 * a relaxed version with explicit barrier is used. The control
377 	 * dependency will order the reading of pn->state before any
378 	 * subsequent writes.
379 	 */
380 	smp_mb__before_atomic();
381 	if (cmpxchg_relaxed(&pn->state, vcpu_halted, vcpu_hashed)
382 	    != vcpu_halted)
383 		return;
384 
385 	/*
386 	 * Put the lock into the hash table and set the _Q_SLOW_VAL.
387 	 *
388 	 * As this is the same vCPU that will check the _Q_SLOW_VAL value and
389 	 * the hash table later on at unlock time, no atomic instruction is
390 	 * needed.
391 	 */
392 	WRITE_ONCE(lock->locked, _Q_SLOW_VAL);
393 	(void)pv_hash(lock, pn);
394 }
395 
396 /*
397  * Wait for l->locked to become clear and acquire the lock;
398  * halt the vcpu after a short spin.
399  * __pv_queued_spin_unlock() will wake us.
400  *
401  * The current value of the lock will be returned for additional processing.
402  */
403 static u32
404 pv_wait_head_or_lock(struct qspinlock *lock, struct mcs_spinlock *node)
405 {
406 	struct pv_node *pn = (struct pv_node *)node;
407 	struct qspinlock **lp = NULL;
408 	int waitcnt = 0;
409 	int loop;
410 
411 	/*
412 	 * If pv_kick_node() already advanced our state, we don't need to
413 	 * insert ourselves into the hash table anymore.
414 	 */
415 	if (READ_ONCE(pn->state) == vcpu_hashed)
416 		lp = (struct qspinlock **)1;
417 
418 	/*
419 	 * Tracking # of slowpath locking operations
420 	 */
421 	qstat_inc(qstat_lock_slowpath, true);
422 
423 	for (;; waitcnt++) {
424 		/*
425 		 * Set correct vCPU state to be used by queue node wait-early
426 		 * mechanism.
427 		 */
428 		WRITE_ONCE(pn->state, vcpu_running);
429 
430 		/*
431 		 * Set the pending bit in the active lock spinning loop to
432 		 * disable lock stealing before attempting to acquire the lock.
433 		 */
434 		set_pending(lock);
435 		for (loop = SPIN_THRESHOLD; loop; loop--) {
436 			if (trylock_clear_pending(lock))
437 				goto gotlock;
438 			cpu_relax();
439 		}
440 		clear_pending(lock);
441 
442 
443 		if (!lp) { /* ONCE */
444 			lp = pv_hash(lock, pn);
445 
446 			/*
447 			 * We must hash before setting _Q_SLOW_VAL, such that
448 			 * when we observe _Q_SLOW_VAL in __pv_queued_spin_unlock()
449 			 * we'll be sure to be able to observe our hash entry.
450 			 *
451 			 *   [S] <hash>                 [Rmw] l->locked == _Q_SLOW_VAL
452 			 *       MB                           RMB
453 			 * [RmW] l->locked = _Q_SLOW_VAL  [L] <unhash>
454 			 *
455 			 * Matches the smp_rmb() in __pv_queued_spin_unlock().
456 			 */
457 			if (xchg(&lock->locked, _Q_SLOW_VAL) == 0) {
458 				/*
459 				 * The lock was free and now we own the lock.
460 				 * Change the lock value back to _Q_LOCKED_VAL
461 				 * and unhash the table.
462 				 */
463 				WRITE_ONCE(lock->locked, _Q_LOCKED_VAL);
464 				WRITE_ONCE(*lp, NULL);
465 				goto gotlock;
466 			}
467 		}
468 		WRITE_ONCE(pn->state, vcpu_hashed);
469 		qstat_inc(qstat_pv_wait_head, true);
470 		qstat_inc(qstat_pv_wait_again, waitcnt);
471 		pv_wait(&lock->locked, _Q_SLOW_VAL);
472 
473 		/*
474 		 * Because of lock stealing, the queue head vCPU may not be
475 		 * able to acquire the lock before it has to wait again.
476 		 */
477 	}
478 
479 	/*
480 	 * The cmpxchg() or xchg() call before coming here provides the
481 	 * acquire semantics for locking. The dummy ORing of _Q_LOCKED_VAL
482 	 * here is to indicate to the compiler that the value will always
483 	 * be nozero to enable better code optimization.
484 	 */
485 gotlock:
486 	return (u32)(atomic_read(&lock->val) | _Q_LOCKED_VAL);
487 }
488 
489 /*
490  * PV versions of the unlock fastpath and slowpath functions to be used
491  * instead of queued_spin_unlock().
492  */
493 __visible void
494 __pv_queued_spin_unlock_slowpath(struct qspinlock *lock, u8 locked)
495 {
496 	struct pv_node *node;
497 
498 	if (unlikely(locked != _Q_SLOW_VAL)) {
499 		WARN(!debug_locks_silent,
500 		     "pvqspinlock: lock 0x%lx has corrupted value 0x%x!\n",
501 		     (unsigned long)lock, atomic_read(&lock->val));
502 		return;
503 	}
504 
505 	/*
506 	 * A failed cmpxchg doesn't provide any memory-ordering guarantees,
507 	 * so we need a barrier to order the read of the node data in
508 	 * pv_unhash *after* we've read the lock being _Q_SLOW_VAL.
509 	 *
510 	 * Matches the cmpxchg() in pv_wait_head_or_lock() setting _Q_SLOW_VAL.
511 	 */
512 	smp_rmb();
513 
514 	/*
515 	 * Since the above failed to release, this must be the SLOW path.
516 	 * Therefore start by looking up the blocked node and unhashing it.
517 	 */
518 	node = pv_unhash(lock);
519 
520 	/*
521 	 * Now that we have a reference to the (likely) blocked pv_node,
522 	 * release the lock.
523 	 */
524 	smp_store_release(&lock->locked, 0);
525 
526 	/*
527 	 * At this point the memory pointed at by lock can be freed/reused,
528 	 * however we can still use the pv_node to kick the CPU.
529 	 * The other vCPU may not really be halted, but kicking an active
530 	 * vCPU is harmless other than the additional latency in completing
531 	 * the unlock.
532 	 */
533 	qstat_inc(qstat_pv_kick_unlock, true);
534 	pv_kick(node->cpu);
535 }
536 
537 /*
538  * Include the architecture specific callee-save thunk of the
539  * __pv_queued_spin_unlock(). This thunk is put together with
540  * __pv_queued_spin_unlock() to make the callee-save thunk and the real unlock
541  * function close to each other sharing consecutive instruction cachelines.
542  * Alternatively, architecture specific version of __pv_queued_spin_unlock()
543  * can be defined.
544  */
545 #include <asm/qspinlock_paravirt.h>
546 
547 #ifndef __pv_queued_spin_unlock
548 __visible void __pv_queued_spin_unlock(struct qspinlock *lock)
549 {
550 	u8 locked;
551 
552 	/*
553 	 * We must not unlock if SLOW, because in that case we must first
554 	 * unhash. Otherwise it would be possible to have multiple @lock
555 	 * entries, which would be BAD.
556 	 */
557 	locked = cmpxchg_release(&lock->locked, _Q_LOCKED_VAL, 0);
558 	if (likely(locked == _Q_LOCKED_VAL))
559 		return;
560 
561 	__pv_queued_spin_unlock_slowpath(lock, locked);
562 }
563 #endif /* __pv_queued_spin_unlock */
564