1 // SPDX-License-Identifier: GPL-2.0-or-later
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
14 #ifndef _GEN_PV_LOCK_SLOWPATH
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 #include <trace/events/lock.h>
28 * Include queued spinlock statistics code
30 #include "qspinlock_stat.h"
33 * The basic principle of a queue-based spinlock can best be understood
34 * by studying a classic queue-based spinlock implementation called the
35 * MCS lock. A copy of the original MCS lock paper ("Algorithms for Scalable
36 * Synchronization on Shared-Memory Multiprocessors by Mellor-Crummey and
37 * Scott") is available at
39 * https://bugzilla.kernel.org/show_bug.cgi?id=206115
41 * This queued spinlock implementation is based on the MCS lock, however to
42 * make it fit the 4 bytes we assume spinlock_t to be, and preserve its
43 * existing API, we must modify it somehow.
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.
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.
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
64 * N.B. The current implementation only supports architectures that allow
65 * atomic operations on smaller 8-bit and 16-bit data types.
69 #include "mcs_spinlock.h"
73 * On 64-bit architectures, the mcs_spinlock structure will be 16 bytes in
74 * size and four of them will fit nicely in one 64-byte cacheline. For
75 * pvqspinlock, however, we need more space for extra data. To accommodate
76 * that, we insert two more long words to pad it up to 32 bytes. IOW, only
77 * two of them can fit in a cacheline in this case. That is OK as it is rare
78 * to have more than 2 levels of slowpath nesting in actual use. We don't
79 * want to penalize pvqspinlocks to optimize for a rare case in native
83 struct mcs_spinlock mcs;
84 #ifdef CONFIG_PARAVIRT_SPINLOCKS
90 * The pending bit spinning loop count.
91 * This heuristic is used to limit the number of lockword accesses
92 * made by atomic_cond_read_relaxed when waiting for the lock to
93 * transition out of the "== _Q_PENDING_VAL" state. We don't spin
94 * indefinitely because there's no guarantee that we'll make forward
97 #ifndef _Q_PENDING_LOOPS
98 #define _Q_PENDING_LOOPS 1
102 * Per-CPU queue node structures; we can never have more than 4 nested
103 * contexts: task, softirq, hardirq, nmi.
105 * Exactly fits one 64-byte cacheline on a 64-bit architecture.
107 * PV doubles the storage and uses the second cacheline for PV state.
109 static DEFINE_PER_CPU_ALIGNED(struct qnode, qnodes[MAX_NODES]);
112 * We must be able to distinguish between no-tail and the tail at 0:0,
113 * therefore increment the cpu number by one.
116 static inline __pure u32 encode_tail(int cpu, int idx)
120 tail = (cpu + 1) << _Q_TAIL_CPU_OFFSET;
121 tail |= idx << _Q_TAIL_IDX_OFFSET; /* assume < 4 */
126 static inline __pure struct mcs_spinlock *decode_tail(u32 tail)
128 int cpu = (tail >> _Q_TAIL_CPU_OFFSET) - 1;
129 int idx = (tail & _Q_TAIL_IDX_MASK) >> _Q_TAIL_IDX_OFFSET;
131 return per_cpu_ptr(&qnodes[idx].mcs, cpu);
135 struct mcs_spinlock *grab_mcs_node(struct mcs_spinlock *base, int idx)
137 return &((struct qnode *)base + idx)->mcs;
140 #define _Q_LOCKED_PENDING_MASK (_Q_LOCKED_MASK | _Q_PENDING_MASK)
142 #if _Q_PENDING_BITS == 8
144 * clear_pending - clear the pending bit.
145 * @lock: Pointer to queued spinlock structure
149 static __always_inline void clear_pending(struct qspinlock *lock)
151 WRITE_ONCE(lock->pending, 0);
155 * clear_pending_set_locked - take ownership and clear the pending bit.
156 * @lock: Pointer to queued spinlock structure
160 * Lock stealing is not allowed if this function is used.
162 static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
164 WRITE_ONCE(lock->locked_pending, _Q_LOCKED_VAL);
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
173 * xchg(lock, tail), which heads an address dependency
175 * p,*,* -> n,*,* ; prev = xchg(lock, node)
177 static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
180 * We can use relaxed semantics since the caller ensures that the
181 * MCS node is properly initialized before updating the tail.
183 return (u32)xchg_relaxed(&lock->tail,
184 tail >> _Q_TAIL_OFFSET) << _Q_TAIL_OFFSET;
187 #else /* _Q_PENDING_BITS == 8 */
190 * clear_pending - clear the pending bit.
191 * @lock: Pointer to queued spinlock structure
195 static __always_inline void clear_pending(struct qspinlock *lock)
197 atomic_andnot(_Q_PENDING_VAL, &lock->val);
201 * clear_pending_set_locked - take ownership and clear the pending bit.
202 * @lock: Pointer to queued spinlock structure
206 static __always_inline void clear_pending_set_locked(struct qspinlock *lock)
208 atomic_add(-_Q_PENDING_VAL + _Q_LOCKED_VAL, &lock->val);
212 * xchg_tail - Put in the new queue tail code word & retrieve previous one
213 * @lock : Pointer to queued spinlock structure
214 * @tail : The new queue tail code word
215 * Return: The previous queue tail code word
219 * p,*,* -> n,*,* ; prev = xchg(lock, node)
221 static __always_inline u32 xchg_tail(struct qspinlock *lock, u32 tail)
225 old = atomic_read(&lock->val);
227 new = (old & _Q_LOCKED_PENDING_MASK) | tail;
229 * We can use relaxed semantics since the caller ensures that
230 * the MCS node is properly initialized before updating the
233 } while (!atomic_try_cmpxchg_relaxed(&lock->val, &old, new));
237 #endif /* _Q_PENDING_BITS == 8 */
240 * queued_fetch_set_pending_acquire - fetch the whole lock value and set pending
241 * @lock : Pointer to queued spinlock structure
242 * Return: The previous lock value
246 #ifndef queued_fetch_set_pending_acquire
247 static __always_inline u32 queued_fetch_set_pending_acquire(struct qspinlock *lock)
249 return atomic_fetch_or_acquire(_Q_PENDING_VAL, &lock->val);
254 * set_locked - Set the lock bit and own the lock
255 * @lock: Pointer to queued spinlock structure
259 static __always_inline void set_locked(struct qspinlock *lock)
261 WRITE_ONCE(lock->locked, _Q_LOCKED_VAL);
266 * Generate the native code for queued_spin_unlock_slowpath(); provide NOPs for
267 * all the PV callbacks.
270 static __always_inline void __pv_init_node(struct mcs_spinlock *node) { }
271 static __always_inline void __pv_wait_node(struct mcs_spinlock *node,
272 struct mcs_spinlock *prev) { }
273 static __always_inline void __pv_kick_node(struct qspinlock *lock,
274 struct mcs_spinlock *node) { }
275 static __always_inline u32 __pv_wait_head_or_lock(struct qspinlock *lock,
276 struct mcs_spinlock *node)
279 #define pv_enabled() false
281 #define pv_init_node __pv_init_node
282 #define pv_wait_node __pv_wait_node
283 #define pv_kick_node __pv_kick_node
284 #define pv_wait_head_or_lock __pv_wait_head_or_lock
286 #ifdef CONFIG_PARAVIRT_SPINLOCKS
287 #define queued_spin_lock_slowpath native_queued_spin_lock_slowpath
290 #endif /* _GEN_PV_LOCK_SLOWPATH */
293 * queued_spin_lock_slowpath - acquire the queued spinlock
294 * @lock: Pointer to queued spinlock structure
295 * @val: Current value of the queued spinlock 32-bit word
297 * (queue tail, pending bit, lock value)
299 * fast : slow : unlock
301 * uncontended (0,0,0) -:--> (0,0,1) ------------------------------:--> (*,*,0)
302 * : | ^--------.------. / :
304 * pending : (0,1,1) +--> (0,1,0) \ | :
307 * uncontended : (n,x,y) +--> (n,0,0) --' | :
310 * contended : (*,x,y) +--> (*,0,0) ---> (*,0,1) -' :
313 void __lockfunc queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
315 struct mcs_spinlock *prev, *next, *node;
319 BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
324 if (virt_spin_lock(lock))
328 * Wait for in-progress pending->locked hand-overs with a bounded
329 * number of spins so that we guarantee forward progress.
333 if (val == _Q_PENDING_VAL) {
334 int cnt = _Q_PENDING_LOOPS;
335 val = atomic_cond_read_relaxed(&lock->val,
336 (VAL != _Q_PENDING_VAL) || !cnt--);
340 * If we observe any contention; queue.
342 if (val & ~_Q_LOCKED_MASK)
348 * 0,0,* -> 0,1,* -> 0,0,1 pending, trylock
350 val = queued_fetch_set_pending_acquire(lock);
353 * If we observe contention, there is a concurrent locker.
355 * Undo and queue; our setting of PENDING might have made the
356 * n,0,0 -> 0,0,0 transition fail and it will now be waiting
357 * on @next to become !NULL.
359 if (unlikely(val & ~_Q_LOCKED_MASK)) {
361 /* Undo PENDING if we set it. */
362 if (!(val & _Q_PENDING_MASK))
369 * We're pending, wait for the owner to go away.
373 * this wait loop must be a load-acquire such that we match the
374 * store-release that clears the locked bit and create lock
375 * sequentiality; this is because not all
376 * clear_pending_set_locked() implementations imply full
379 if (val & _Q_LOCKED_MASK)
380 smp_cond_load_acquire(&lock->locked, !VAL);
383 * take ownership and clear the pending bit.
387 clear_pending_set_locked(lock);
388 lockevent_inc(lock_pending);
392 * End of pending bit optimistic spinning and beginning of MCS
396 lockevent_inc(lock_slowpath);
398 node = this_cpu_ptr(&qnodes[0].mcs);
400 tail = encode_tail(smp_processor_id(), idx);
402 trace_contention_begin(lock, LCB_F_SPIN);
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
413 if (unlikely(idx >= MAX_NODES)) {
414 lockevent_inc(lock_no_node);
415 while (!queued_spin_trylock(lock))
420 node = grab_mcs_node(node, idx);
423 * Keep counts of non-zero index values:
425 lockevent_cond_inc(lock_use_node2 + idx - 1, idx);
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.
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
443 if (queued_spin_trylock(lock))
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.
454 * Publish the updated tail.
455 * We have already touched the queueing cacheline; don't bother with
460 old = xchg_tail(lock, tail);
464 * if there was a previous node; link it and wait until reaching the
465 * head of the waitqueue.
467 if (old & _Q_TAIL_MASK) {
468 prev = decode_tail(old);
470 /* Link @node into the waitqueue. */
471 WRITE_ONCE(prev->next, node);
473 pv_wait_node(node, prev);
474 arch_mcs_spin_lock_contended(&node->locked);
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.
482 next = READ_ONCE(node->next);
488 * we're at the head of the waitqueue, wait for the owner & pending to
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.
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.
505 * If PV isn't active, 0 will be returned instead.
508 if ((val = pv_wait_head_or_lock(lock, node)))
511 val = atomic_cond_read_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK));
517 * n,0,0 -> 0,0,1 : lock, uncontended
518 * *,*,0 -> *,*,1 : lock, contended
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.
526 * In the PV case we might already have _Q_LOCKED_VAL set, because
527 * of lock stealing; therefore we must also allow:
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.
535 if ((val & _Q_TAIL_MASK) == tail) {
536 if (atomic_try_cmpxchg_relaxed(&lock->val, &val, _Q_LOCKED_VAL))
537 goto release; /* No contention */
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.
548 * contended path; wait for next if not observed yet, release.
551 next = smp_cond_load_relaxed(&node->next, (VAL));
553 arch_mcs_spin_unlock_contended(&next->locked);
554 pv_kick_node(lock, next);
557 trace_contention_end(lock, 0);
562 __this_cpu_dec(qnodes[0].mcs.count);
564 EXPORT_SYMBOL(queued_spin_lock_slowpath);
567 * Generate the paravirt code for queued_spin_unlock_slowpath().
569 #if !defined(_GEN_PV_LOCK_SLOWPATH) && defined(CONFIG_PARAVIRT_SPINLOCKS)
570 #define _GEN_PV_LOCK_SLOWPATH
573 #define pv_enabled() true
578 #undef pv_wait_head_or_lock
580 #undef queued_spin_lock_slowpath
581 #define queued_spin_lock_slowpath __pv_queued_spin_lock_slowpath
583 #include "qspinlock_paravirt.h"
584 #include "qspinlock.c"
587 static __init int parse_nopvspin(char *arg)
592 early_param("nopvspin", parse_nopvspin);