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

source code of linux/kernel/locking/qspinlock_paravirt.h