1 // SPDX-License-Identifier: GPL-2.0-or-later
3 #include <linux/compiler.h>
4 #include <linux/export.h>
5 #include <linux/percpu.h>
6 #include <linux/processor.h>
8 #include <linux/topology.h>
9 #include <linux/sched/clock.h>
10 #include <asm/qspinlock.h>
11 #include <asm/paravirt.h>
17 struct qspinlock *lock;
19 u8 sleepy; /* 1 if the previous vCPU was preempted or
20 * if the previous node was sleepy */
21 u8 locked; /* 1 if lock acquired */
26 struct qnode nodes[MAX_NODES];
29 /* Tuning parameters */
30 static int steal_spins __read_mostly = (1 << 5);
31 static int remote_steal_spins __read_mostly = (1 << 2);
32 #if _Q_SPIN_TRY_LOCK_STEAL == 1
33 static const bool maybe_stealers = true;
35 static bool maybe_stealers __read_mostly = true;
37 static int head_spins __read_mostly = (1 << 8);
39 static bool pv_yield_owner __read_mostly = true;
40 static bool pv_yield_allow_steal __read_mostly = false;
41 static bool pv_spin_on_preempted_owner __read_mostly = false;
42 static bool pv_sleepy_lock __read_mostly = true;
43 static bool pv_sleepy_lock_sticky __read_mostly = false;
44 static u64 pv_sleepy_lock_interval_ns __read_mostly = 0;
45 static int pv_sleepy_lock_factor __read_mostly = 256;
46 static bool pv_yield_prev __read_mostly = true;
47 static bool pv_yield_sleepy_owner __read_mostly = true;
48 static bool pv_prod_head __read_mostly = false;
50 static DEFINE_PER_CPU_ALIGNED(struct qnodes, qnodes);
51 static DEFINE_PER_CPU_ALIGNED(u64, sleepy_lock_seen_clock);
53 #if _Q_SPIN_SPEC_BARRIER == 1
54 #define spec_barrier() do { asm volatile("ori 31,31,0" ::: "memory"); } while (0)
56 #define spec_barrier() do { } while (0)
59 static __always_inline bool recently_sleepy(void)
61 /* pv_sleepy_lock is true when this is called */
62 if (pv_sleepy_lock_interval_ns) {
63 u64 seen = this_cpu_read(sleepy_lock_seen_clock);
66 u64 delta = sched_clock() - seen;
67 if (delta < pv_sleepy_lock_interval_ns)
69 this_cpu_write(sleepy_lock_seen_clock, 0);
76 static __always_inline int get_steal_spins(bool paravirt, bool sleepy)
78 if (paravirt && sleepy)
79 return steal_spins * pv_sleepy_lock_factor;
84 static __always_inline int get_remote_steal_spins(bool paravirt, bool sleepy)
86 if (paravirt && sleepy)
87 return remote_steal_spins * pv_sleepy_lock_factor;
89 return remote_steal_spins;
92 static __always_inline int get_head_spins(bool paravirt, bool sleepy)
94 if (paravirt && sleepy)
95 return head_spins * pv_sleepy_lock_factor;
100 static inline u32 encode_tail_cpu(int cpu)
102 return (cpu + 1) << _Q_TAIL_CPU_OFFSET;
105 static inline int decode_tail_cpu(u32 val)
107 return (val >> _Q_TAIL_CPU_OFFSET) - 1;
110 static inline int get_owner_cpu(u32 val)
112 return (val & _Q_OWNER_CPU_MASK) >> _Q_OWNER_CPU_OFFSET;
116 * Try to acquire the lock if it was not already locked. If the tail matches
117 * mytail then clear it, otherwise leave it unchnaged. Return previous value.
119 * This is used by the head of the queue to acquire the lock and clean up
120 * its tail if it was the last one queued.
122 static __always_inline u32 trylock_clean_tail(struct qspinlock *lock, u32 tail)
124 u32 newval = queued_spin_encode_locked_val();
128 "1: lwarx %0,0,%2,%7 # trylock_clean_tail \n"
129 /* This test is necessary if there could be stealers */
132 /* Test whether the lock tail == mytail */
135 /* Merge the new locked value */
138 /* If the lock tail matched, then clear it, otherwise leave it. */
140 "2: stwcx. %1,0,%2 \n"
142 "\t" PPC_ACQUIRE_BARRIER " \n"
144 : "=&r" (prev), "=&r" (tmp)
145 : "r" (&lock->val), "r"(tail), "r" (newval),
147 "r" (_Q_TAIL_CPU_MASK),
148 "i" (_Q_SPIN_EH_HINT)
155 * Publish our tail, replacing previous tail. Return previous value.
157 * This provides a release barrier for publishing node, this pairs with the
158 * acquire barrier in get_tail_qnode() when the next CPU finds this tail
161 static __always_inline u32 publish_tail_cpu(struct qspinlock *lock, u32 tail)
168 "\t" PPC_RELEASE_BARRIER " \n"
169 "1: lwarx %0,0,%2 # publish_tail_cpu \n"
174 : "=&r" (prev), "=&r"(tmp)
175 : "r" (&lock->val), "r" (tail), "r"(_Q_TAIL_CPU_MASK)
181 static __always_inline u32 set_mustq(struct qspinlock *lock)
186 "1: lwarx %0,0,%1 # set_mustq \n"
191 : "r" (&lock->val), "r" (_Q_MUST_Q_VAL)
197 static __always_inline u32 clear_mustq(struct qspinlock *lock)
202 "1: lwarx %0,0,%1 # clear_mustq \n"
207 : "r" (&lock->val), "r" (_Q_MUST_Q_VAL)
213 static __always_inline bool try_set_sleepy(struct qspinlock *lock, u32 old)
216 u32 new = old | _Q_SLEEPY_VAL;
218 BUG_ON(!(old & _Q_LOCKED_VAL));
219 BUG_ON(old & _Q_SLEEPY_VAL);
222 "1: lwarx %0,0,%1 # try_set_sleepy \n"
229 : "r" (&lock->val), "r"(old), "r" (new)
232 return likely(prev == old);
235 static __always_inline void seen_sleepy_owner(struct qspinlock *lock, u32 val)
237 if (pv_sleepy_lock) {
238 if (pv_sleepy_lock_interval_ns)
239 this_cpu_write(sleepy_lock_seen_clock, sched_clock());
240 if (!(val & _Q_SLEEPY_VAL))
241 try_set_sleepy(lock, val);
245 static __always_inline void seen_sleepy_lock(void)
247 if (pv_sleepy_lock && pv_sleepy_lock_interval_ns)
248 this_cpu_write(sleepy_lock_seen_clock, sched_clock());
251 static __always_inline void seen_sleepy_node(void)
253 if (pv_sleepy_lock) {
254 if (pv_sleepy_lock_interval_ns)
255 this_cpu_write(sleepy_lock_seen_clock, sched_clock());
256 /* Don't set sleepy because we likely have a stale val */
260 static struct qnode *get_tail_qnode(struct qspinlock *lock, int prev_cpu)
262 struct qnodes *qnodesp = per_cpu_ptr(&qnodes, prev_cpu);
266 * After publishing the new tail and finding a previous tail in the
267 * previous val (which is the control dependency), this barrier
268 * orders the release barrier in publish_tail_cpu performed by the
269 * last CPU, with subsequently looking at its qnode structures
272 smp_acquire__after_ctrl_dep();
274 for (idx = 0; idx < MAX_NODES; idx++) {
275 struct qnode *qnode = &qnodesp->nodes[idx];
276 if (qnode->lock == lock)
283 /* Called inside spin_begin(). Returns whether or not the vCPU was preempted. */
284 static __always_inline bool __yield_to_locked_owner(struct qspinlock *lock, u32 val, bool paravirt, bool mustq)
288 bool preempted = false;
290 BUG_ON(!(val & _Q_LOCKED_VAL));
298 owner = get_owner_cpu(val);
299 yield_count = yield_count_of(owner);
301 if ((yield_count & 1) == 0)
302 goto relax; /* owner vcpu is running */
306 seen_sleepy_owner(lock, val);
310 * Read the lock word after sampling the yield count. On the other side
311 * there may a wmb because the yield count update is done by the
312 * hypervisor preemption and the value update by the OS, however this
313 * ordering might reduce the chance of out of order accesses and
314 * improve the heuristic.
318 if (READ_ONCE(lock->val) == val) {
321 yield_to_preempted(owner, yield_count);
326 /* Don't relax if we yielded. Maybe we should? */
336 /* Called inside spin_begin(). Returns whether or not the vCPU was preempted. */
337 static __always_inline bool yield_to_locked_owner(struct qspinlock *lock, u32 val, bool paravirt)
339 return __yield_to_locked_owner(lock, val, paravirt, false);
342 /* Called inside spin_begin(). Returns whether or not the vCPU was preempted. */
343 static __always_inline bool yield_head_to_locked_owner(struct qspinlock *lock, u32 val, bool paravirt)
347 if ((val & _Q_MUST_Q_VAL) && pv_yield_allow_steal)
350 return __yield_to_locked_owner(lock, val, paravirt, mustq);
353 static __always_inline void propagate_sleepy(struct qnode *node, u32 val, bool paravirt)
360 if (!pv_yield_sleepy_owner)
363 next = READ_ONCE(node->next);
370 owner = get_owner_cpu(val);
371 if (vcpu_is_preempted(owner))
375 /* Called inside spin_begin() */
376 static __always_inline bool yield_to_prev(struct qspinlock *lock, struct qnode *node, int prev_cpu, bool paravirt)
379 bool preempted = false;
384 if (!pv_yield_sleepy_owner)
388 * If the previous waiter was preempted it might not be able to
389 * propagate sleepy to us, so check the lock in that case too.
391 if (node->sleepy || vcpu_is_preempted(prev_cpu)) {
392 u32 val = READ_ONCE(lock->val);
394 if (val & _Q_LOCKED_VAL) {
395 if (node->next && !node->next->sleepy) {
397 * Propagate sleepy to next waiter. Only if
398 * owner is preempted, which allows the queue
399 * to become "non-sleepy" if vCPU preemption
400 * ceases to occur, even if the lock remains
403 if (vcpu_is_preempted(get_owner_cpu(val)))
404 node->next->sleepy = 1;
407 preempted = yield_to_locked_owner(lock, val, paravirt);
411 node->sleepy = false;
418 yield_count = yield_count_of(prev_cpu);
419 if ((yield_count & 1) == 0)
420 goto relax; /* owner vcpu is running */
427 smp_rmb(); /* See __yield_to_locked_owner comment */
429 if (!READ_ONCE(node->locked)) {
430 yield_to_preempted(prev_cpu, yield_count);
442 static __always_inline bool steal_break(u32 val, int iters, bool paravirt, bool sleepy)
444 if (iters >= get_steal_spins(paravirt, sleepy))
447 if (IS_ENABLED(CONFIG_NUMA) &&
448 (iters >= get_remote_steal_spins(paravirt, sleepy))) {
449 int cpu = get_owner_cpu(val);
450 if (numa_node_id() != cpu_to_node(cpu))
456 static __always_inline bool try_to_steal_lock(struct qspinlock *lock, bool paravirt)
458 bool seen_preempted = false;
464 /* XXX: should spin_on_preempted_owner do anything here? */
468 /* Attempt to steal the lock */
471 bool preempted = false;
473 val = READ_ONCE(lock->val);
474 if (val & _Q_MUST_Q_VAL)
478 if (unlikely(!(val & _Q_LOCKED_VAL))) {
480 if (__queued_spin_trylock_steal(lock))
484 preempted = yield_to_locked_owner(lock, val, paravirt);
487 if (paravirt && pv_sleepy_lock) {
489 if (val & _Q_SLEEPY_VAL) {
492 } else if (recently_sleepy()) {
496 if (pv_sleepy_lock_sticky && seen_preempted &&
497 !(val & _Q_SLEEPY_VAL)) {
498 if (try_set_sleepy(lock, val))
499 val |= _Q_SLEEPY_VAL;
504 seen_preempted = true;
506 if (!pv_spin_on_preempted_owner)
509 * pv_spin_on_preempted_owner don't increase iters
510 * while the owner is preempted -- we won't interfere
511 * with it by definition. This could introduce some
512 * latency issue if we continually observe preempted
513 * owners, but hopefully that's a rare corner case of
514 * a badly oversubscribed system.
519 } while (!steal_break(val, iters, paravirt, sleepy));
526 static __always_inline void queued_spin_lock_mcs_queue(struct qspinlock *lock, bool paravirt)
528 struct qnodes *qnodesp;
529 struct qnode *next, *node;
531 bool seen_preempted = false;
537 BUILD_BUG_ON(CONFIG_NR_CPUS >= (1U << _Q_TAIL_CPU_BITS));
539 qnodesp = this_cpu_ptr(&qnodes);
540 if (unlikely(qnodesp->count >= MAX_NODES)) {
542 while (!queued_spin_trylock(lock))
547 idx = qnodesp->count++;
549 * Ensure that we increment the head node->count before initialising
550 * the actual node. If the compiler is kind enough to reorder these
551 * stores, then an IRQ could overwrite our assignments.
554 node = &qnodesp->nodes[idx];
557 node->cpu = smp_processor_id();
561 tail = encode_tail_cpu(node->cpu);
564 * Assign all attributes of a node before it can be published.
565 * Issues an lwsync, serving as a release barrier, as well as a
568 old = publish_tail_cpu(lock, tail);
571 * If there was a previous node; link it and wait until reaching the
572 * head of the waitqueue.
574 if (old & _Q_TAIL_CPU_MASK) {
575 int prev_cpu = decode_tail_cpu(old);
576 struct qnode *prev = get_tail_qnode(lock, prev_cpu);
578 /* Link @node into the waitqueue. */
579 WRITE_ONCE(prev->next, node);
581 /* Wait for mcs node lock to be released */
583 while (!READ_ONCE(node->locked)) {
586 if (yield_to_prev(lock, node, prev_cpu, paravirt))
587 seen_preempted = true;
592 smp_rmb(); /* acquire barrier for the mcs lock */
595 * Generic qspinlocks have this prefetch here, but it seems
596 * like it could cause additional line transitions because
597 * the waiter will keep loading from it.
599 if (_Q_SPIN_PREFETCH_NEXT) {
600 next = READ_ONCE(node->next);
606 /* We're at the head of the waitqueue, wait for the lock. */
612 val = READ_ONCE(lock->val);
613 if (!(val & _Q_LOCKED_VAL))
617 if (paravirt && pv_sleepy_lock && maybe_stealers) {
619 if (val & _Q_SLEEPY_VAL) {
622 } else if (recently_sleepy()) {
626 if (pv_sleepy_lock_sticky && seen_preempted &&
627 !(val & _Q_SLEEPY_VAL)) {
628 if (try_set_sleepy(lock, val))
629 val |= _Q_SLEEPY_VAL;
633 propagate_sleepy(node, val, paravirt);
634 preempted = yield_head_to_locked_owner(lock, val, paravirt);
639 seen_preempted = true;
641 if (paravirt && preempted) {
644 if (!pv_spin_on_preempted_owner)
650 if (!mustq && iters >= get_head_spins(paravirt, sleepy)) {
653 val |= _Q_MUST_Q_VAL;
659 /* If we're the last queued, must clean up the tail. */
660 old = trylock_clean_tail(lock, tail);
661 if (unlikely(old & _Q_LOCKED_VAL)) {
662 BUG_ON(!maybe_stealers);
663 goto again; /* Can only be true if maybe_stealers. */
666 if ((old & _Q_TAIL_CPU_MASK) == tail)
667 goto release; /* We were the tail, no next. */
669 /* There is a next, must wait for node->next != NULL (MCS protocol) */
670 next = READ_ONCE(node->next);
673 while (!(next = READ_ONCE(node->next)))
680 * Unlock the next mcs waiter node. Release barrier is not required
681 * here because the acquirer is only accessing the lock word, and
682 * the acquire barrier we took the lock with orders that update vs
683 * this store to locked. The corresponding barrier is the smp_rmb()
684 * acquire barrier for mcs lock, above.
686 if (paravirt && pv_prod_head) {
687 int next_cpu = next->cpu;
688 WRITE_ONCE(next->locked, 1);
690 asm volatile("miso" ::: "memory");
691 if (vcpu_is_preempted(next_cpu))
694 WRITE_ONCE(next->locked, 1);
696 asm volatile("miso" ::: "memory");
700 qnodesp->count--; /* release the node */
703 void queued_spin_lock_slowpath(struct qspinlock *lock)
706 * This looks funny, but it induces the compiler to inline both
707 * sides of the branch rather than share code as when the condition
708 * is passed as the paravirt argument to the functions.
710 if (IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS) && is_shared_processor()) {
711 if (try_to_steal_lock(lock, true)) {
715 queued_spin_lock_mcs_queue(lock, true);
717 if (try_to_steal_lock(lock, false)) {
721 queued_spin_lock_mcs_queue(lock, false);
724 EXPORT_SYMBOL(queued_spin_lock_slowpath);
726 #ifdef CONFIG_PARAVIRT_SPINLOCKS
727 void pv_spinlocks_init(void)
732 #include <linux/debugfs.h>
733 static int steal_spins_set(void *data, u64 val)
735 #if _Q_SPIN_TRY_LOCK_STEAL == 1
736 /* MAYBE_STEAL remains true */
739 static DEFINE_MUTEX(lock);
742 * The lock slow path has a !maybe_stealers case that can assume
743 * the head of queue will not see concurrent waiters. That waiter
744 * is unsafe in the presence of stealers, so must keep them away
749 if (val && !steal_spins) {
750 maybe_stealers = true;
751 /* wait for queue head waiter to go away */
754 } else if (!val && steal_spins) {
756 /* wait for all possible stealers to go away */
758 maybe_stealers = false;
768 static int steal_spins_get(void *data, u64 *val)
775 DEFINE_SIMPLE_ATTRIBUTE(fops_steal_spins, steal_spins_get, steal_spins_set, "%llu\n");
777 static int remote_steal_spins_set(void *data, u64 val)
779 remote_steal_spins = val;
784 static int remote_steal_spins_get(void *data, u64 *val)
786 *val = remote_steal_spins;
791 DEFINE_SIMPLE_ATTRIBUTE(fops_remote_steal_spins, remote_steal_spins_get, remote_steal_spins_set, "%llu\n");
793 static int head_spins_set(void *data, u64 val)
800 static int head_spins_get(void *data, u64 *val)
807 DEFINE_SIMPLE_ATTRIBUTE(fops_head_spins, head_spins_get, head_spins_set, "%llu\n");
809 static int pv_yield_owner_set(void *data, u64 val)
811 pv_yield_owner = !!val;
816 static int pv_yield_owner_get(void *data, u64 *val)
818 *val = pv_yield_owner;
823 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_yield_owner, pv_yield_owner_get, pv_yield_owner_set, "%llu\n");
825 static int pv_yield_allow_steal_set(void *data, u64 val)
827 pv_yield_allow_steal = !!val;
832 static int pv_yield_allow_steal_get(void *data, u64 *val)
834 *val = pv_yield_allow_steal;
839 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_yield_allow_steal, pv_yield_allow_steal_get, pv_yield_allow_steal_set, "%llu\n");
841 static int pv_spin_on_preempted_owner_set(void *data, u64 val)
843 pv_spin_on_preempted_owner = !!val;
848 static int pv_spin_on_preempted_owner_get(void *data, u64 *val)
850 *val = pv_spin_on_preempted_owner;
855 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_spin_on_preempted_owner, pv_spin_on_preempted_owner_get, pv_spin_on_preempted_owner_set, "%llu\n");
857 static int pv_sleepy_lock_set(void *data, u64 val)
859 pv_sleepy_lock = !!val;
864 static int pv_sleepy_lock_get(void *data, u64 *val)
866 *val = pv_sleepy_lock;
871 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_sleepy_lock, pv_sleepy_lock_get, pv_sleepy_lock_set, "%llu\n");
873 static int pv_sleepy_lock_sticky_set(void *data, u64 val)
875 pv_sleepy_lock_sticky = !!val;
880 static int pv_sleepy_lock_sticky_get(void *data, u64 *val)
882 *val = pv_sleepy_lock_sticky;
887 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_sleepy_lock_sticky, pv_sleepy_lock_sticky_get, pv_sleepy_lock_sticky_set, "%llu\n");
889 static int pv_sleepy_lock_interval_ns_set(void *data, u64 val)
891 pv_sleepy_lock_interval_ns = val;
896 static int pv_sleepy_lock_interval_ns_get(void *data, u64 *val)
898 *val = pv_sleepy_lock_interval_ns;
903 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_sleepy_lock_interval_ns, pv_sleepy_lock_interval_ns_get, pv_sleepy_lock_interval_ns_set, "%llu\n");
905 static int pv_sleepy_lock_factor_set(void *data, u64 val)
907 pv_sleepy_lock_factor = val;
912 static int pv_sleepy_lock_factor_get(void *data, u64 *val)
914 *val = pv_sleepy_lock_factor;
919 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_sleepy_lock_factor, pv_sleepy_lock_factor_get, pv_sleepy_lock_factor_set, "%llu\n");
921 static int pv_yield_prev_set(void *data, u64 val)
923 pv_yield_prev = !!val;
928 static int pv_yield_prev_get(void *data, u64 *val)
930 *val = pv_yield_prev;
935 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_yield_prev, pv_yield_prev_get, pv_yield_prev_set, "%llu\n");
937 static int pv_yield_sleepy_owner_set(void *data, u64 val)
939 pv_yield_sleepy_owner = !!val;
944 static int pv_yield_sleepy_owner_get(void *data, u64 *val)
946 *val = pv_yield_sleepy_owner;
951 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_yield_sleepy_owner, pv_yield_sleepy_owner_get, pv_yield_sleepy_owner_set, "%llu\n");
953 static int pv_prod_head_set(void *data, u64 val)
955 pv_prod_head = !!val;
960 static int pv_prod_head_get(void *data, u64 *val)
967 DEFINE_SIMPLE_ATTRIBUTE(fops_pv_prod_head, pv_prod_head_get, pv_prod_head_set, "%llu\n");
969 static __init int spinlock_debugfs_init(void)
971 debugfs_create_file("qspl_steal_spins", 0600, arch_debugfs_dir, NULL, &fops_steal_spins);
972 debugfs_create_file("qspl_remote_steal_spins", 0600, arch_debugfs_dir, NULL, &fops_remote_steal_spins);
973 debugfs_create_file("qspl_head_spins", 0600, arch_debugfs_dir, NULL, &fops_head_spins);
974 if (is_shared_processor()) {
975 debugfs_create_file("qspl_pv_yield_owner", 0600, arch_debugfs_dir, NULL, &fops_pv_yield_owner);
976 debugfs_create_file("qspl_pv_yield_allow_steal", 0600, arch_debugfs_dir, NULL, &fops_pv_yield_allow_steal);
977 debugfs_create_file("qspl_pv_spin_on_preempted_owner", 0600, arch_debugfs_dir, NULL, &fops_pv_spin_on_preempted_owner);
978 debugfs_create_file("qspl_pv_sleepy_lock", 0600, arch_debugfs_dir, NULL, &fops_pv_sleepy_lock);
979 debugfs_create_file("qspl_pv_sleepy_lock_sticky", 0600, arch_debugfs_dir, NULL, &fops_pv_sleepy_lock_sticky);
980 debugfs_create_file("qspl_pv_sleepy_lock_interval_ns", 0600, arch_debugfs_dir, NULL, &fops_pv_sleepy_lock_interval_ns);
981 debugfs_create_file("qspl_pv_sleepy_lock_factor", 0600, arch_debugfs_dir, NULL, &fops_pv_sleepy_lock_factor);
982 debugfs_create_file("qspl_pv_yield_prev", 0600, arch_debugfs_dir, NULL, &fops_pv_yield_prev);
983 debugfs_create_file("qspl_pv_yield_sleepy_owner", 0600, arch_debugfs_dir, NULL, &fops_pv_yield_sleepy_owner);
984 debugfs_create_file("qspl_pv_prod_head", 0600, arch_debugfs_dir, NULL, &fops_pv_prod_head);
989 device_initcall(spinlock_debugfs_init);