Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Out of line spinlock code.
4 *
5 * Copyright IBM Corp. 2004, 2006
6 * Author(s): Martin Schwidefsky (schwidefsky@de.ibm.com)
7 */
8
9#include <linux/types.h>
10#include <linux/export.h>
11#include <linux/spinlock.h>
12#include <linux/jiffies.h>
13#include <linux/init.h>
14#include <linux/smp.h>
15#include <linux/percpu.h>
16#include <asm/alternative.h>
17#include <asm/io.h>
18
19int spin_retry = -1;
20
21static int __init spin_retry_init(void)
22{
23 if (spin_retry < 0)
24 spin_retry = 1000;
25 return 0;
26}
27early_initcall(spin_retry_init);
28
29/**
30 * spin_retry= parameter
31 */
32static int __init spin_retry_setup(char *str)
33{
34 spin_retry = simple_strtoul(str, &str, 0);
35 return 1;
36}
37__setup("spin_retry=", spin_retry_setup);
38
39struct spin_wait {
40 struct spin_wait *next, *prev;
41 int node_id;
42} __aligned(32);
43
44static DEFINE_PER_CPU_ALIGNED(struct spin_wait, spin_wait[4]);
45
46#define _Q_LOCK_CPU_OFFSET 0
47#define _Q_LOCK_STEAL_OFFSET 16
48#define _Q_TAIL_IDX_OFFSET 18
49#define _Q_TAIL_CPU_OFFSET 20
50
51#define _Q_LOCK_CPU_MASK 0x0000ffff
52#define _Q_LOCK_STEAL_ADD 0x00010000
53#define _Q_LOCK_STEAL_MASK 0x00030000
54#define _Q_TAIL_IDX_MASK 0x000c0000
55#define _Q_TAIL_CPU_MASK 0xfff00000
56
57#define _Q_LOCK_MASK (_Q_LOCK_CPU_MASK | _Q_LOCK_STEAL_MASK)
58#define _Q_TAIL_MASK (_Q_TAIL_IDX_MASK | _Q_TAIL_CPU_MASK)
59
60void arch_spin_lock_setup(int cpu)
61{
62 struct spin_wait *node;
63 int ix;
64
65 node = per_cpu_ptr(&spin_wait[0], cpu);
66 for (ix = 0; ix < 4; ix++, node++) {
67 memset(node, 0, sizeof(*node));
68 node->node_id = ((cpu + 1) << _Q_TAIL_CPU_OFFSET) +
69 (ix << _Q_TAIL_IDX_OFFSET);
70 }
71}
72
73static inline int arch_load_niai4(int *lock)
74{
75 int owner;
76
77 asm volatile(
78 ALTERNATIVE("", ".long 0xb2fa0040", 49) /* NIAI 4 */
79 " l %0,%1\n"
80 : "=d" (owner) : "Q" (*lock) : "memory");
81 return owner;
82}
83
84static inline int arch_cmpxchg_niai8(int *lock, int old, int new)
85{
86 int expected = old;
87
88 asm volatile(
89 ALTERNATIVE("", ".long 0xb2fa0080", 49) /* NIAI 8 */
90 " cs %0,%3,%1\n"
91 : "=d" (old), "=Q" (*lock)
92 : "0" (old), "d" (new), "Q" (*lock)
93 : "cc", "memory");
94 return expected == old;
95}
96
97static inline struct spin_wait *arch_spin_decode_tail(int lock)
98{
99 int ix, cpu;
100
101 ix = (lock & _Q_TAIL_IDX_MASK) >> _Q_TAIL_IDX_OFFSET;
102 cpu = (lock & _Q_TAIL_CPU_MASK) >> _Q_TAIL_CPU_OFFSET;
103 return per_cpu_ptr(&spin_wait[ix], cpu - 1);
104}
105
106static inline int arch_spin_yield_target(int lock, struct spin_wait *node)
107{
108 if (lock & _Q_LOCK_CPU_MASK)
109 return lock & _Q_LOCK_CPU_MASK;
110 if (node == NULL || node->prev == NULL)
111 return 0; /* 0 -> no target cpu */
112 while (node->prev)
113 node = node->prev;
114 return node->node_id >> _Q_TAIL_CPU_OFFSET;
115}
116
117static inline void arch_spin_lock_queued(arch_spinlock_t *lp)
118{
119 struct spin_wait *node, *next;
120 int lockval, ix, node_id, tail_id, old, new, owner, count;
121
122 ix = S390_lowcore.spinlock_index++;
123 barrier();
124 lockval = SPINLOCK_LOCKVAL; /* cpu + 1 */
125 node = this_cpu_ptr(&spin_wait[ix]);
126 node->prev = node->next = NULL;
127 node_id = node->node_id;
128
129 /* Enqueue the node for this CPU in the spinlock wait queue */
130 while (1) {
131 old = READ_ONCE(lp->lock);
132 if ((old & _Q_LOCK_CPU_MASK) == 0 &&
133 (old & _Q_LOCK_STEAL_MASK) != _Q_LOCK_STEAL_MASK) {
134 /*
135 * The lock is free but there may be waiters.
136 * With no waiters simply take the lock, if there
137 * are waiters try to steal the lock. The lock may
138 * be stolen three times before the next queued
139 * waiter will get the lock.
140 */
141 new = (old ? (old + _Q_LOCK_STEAL_ADD) : 0) | lockval;
142 if (__atomic_cmpxchg_bool(&lp->lock, old, new))
143 /* Got the lock */
144 goto out;
145 /* lock passing in progress */
146 continue;
147 }
148 /* Make the node of this CPU the new tail. */
149 new = node_id | (old & _Q_LOCK_MASK);
150 if (__atomic_cmpxchg_bool(&lp->lock, old, new))
151 break;
152 }
153 /* Set the 'next' pointer of the tail node in the queue */
154 tail_id = old & _Q_TAIL_MASK;
155 if (tail_id != 0) {
156 node->prev = arch_spin_decode_tail(tail_id);
157 WRITE_ONCE(node->prev->next, node);
158 }
159
160 /* Pass the virtual CPU to the lock holder if it is not running */
161 owner = arch_spin_yield_target(old, node);
162 if (owner && arch_vcpu_is_preempted(owner - 1))
163 smp_yield_cpu(owner - 1);
164
165 /* Spin on the CPU local node->prev pointer */
166 if (tail_id != 0) {
167 count = spin_retry;
168 while (READ_ONCE(node->prev) != NULL) {
169 if (count-- >= 0)
170 continue;
171 count = spin_retry;
172 /* Query running state of lock holder again. */
173 owner = arch_spin_yield_target(old, node);
174 if (owner && arch_vcpu_is_preempted(owner - 1))
175 smp_yield_cpu(owner - 1);
176 }
177 }
178
179 /* Spin on the lock value in the spinlock_t */
180 count = spin_retry;
181 while (1) {
182 old = READ_ONCE(lp->lock);
183 owner = old & _Q_LOCK_CPU_MASK;
184 if (!owner) {
185 tail_id = old & _Q_TAIL_MASK;
186 new = ((tail_id != node_id) ? tail_id : 0) | lockval;
187 if (__atomic_cmpxchg_bool(&lp->lock, old, new))
188 /* Got the lock */
189 break;
190 continue;
191 }
192 if (count-- >= 0)
193 continue;
194 count = spin_retry;
195 if (!MACHINE_IS_LPAR || arch_vcpu_is_preempted(owner - 1))
196 smp_yield_cpu(owner - 1);
197 }
198
199 /* Pass lock_spin job to next CPU in the queue */
200 if (node_id && tail_id != node_id) {
201 /* Wait until the next CPU has set up the 'next' pointer */
202 while ((next = READ_ONCE(node->next)) == NULL)
203 ;
204 next->prev = NULL;
205 }
206
207 out:
208 S390_lowcore.spinlock_index--;
209}
210
211static inline void arch_spin_lock_classic(arch_spinlock_t *lp)
212{
213 int lockval, old, new, owner, count;
214
215 lockval = SPINLOCK_LOCKVAL; /* cpu + 1 */
216
217 /* Pass the virtual CPU to the lock holder if it is not running */
218 owner = arch_spin_yield_target(READ_ONCE(lp->lock), NULL);
219 if (owner && arch_vcpu_is_preempted(owner - 1))
220 smp_yield_cpu(owner - 1);
221
222 count = spin_retry;
223 while (1) {
224 old = arch_load_niai4(&lp->lock);
225 owner = old & _Q_LOCK_CPU_MASK;
226 /* Try to get the lock if it is free. */
227 if (!owner) {
228 new = (old & _Q_TAIL_MASK) | lockval;
229 if (arch_cmpxchg_niai8(&lp->lock, old, new)) {
230 /* Got the lock */
231 return;
232 }
233 continue;
234 }
235 if (count-- >= 0)
236 continue;
237 count = spin_retry;
238 if (!MACHINE_IS_LPAR || arch_vcpu_is_preempted(owner - 1))
239 smp_yield_cpu(owner - 1);
240 }
241}
242
243void arch_spin_lock_wait(arch_spinlock_t *lp)
244{
245 /* Use classic spinlocks + niai if the steal time is >= 10% */
246 if (test_cpu_flag(CIF_DEDICATED_CPU))
247 arch_spin_lock_queued(lp);
248 else
249 arch_spin_lock_classic(lp);
250}
251EXPORT_SYMBOL(arch_spin_lock_wait);
252
253int arch_spin_trylock_retry(arch_spinlock_t *lp)
254{
255 int cpu = SPINLOCK_LOCKVAL;
256 int owner, count;
257
258 for (count = spin_retry; count > 0; count--) {
259 owner = READ_ONCE(lp->lock);
260 /* Try to get the lock if it is free. */
261 if (!owner) {
262 if (__atomic_cmpxchg_bool(&lp->lock, 0, cpu))
263 return 1;
264 }
265 }
266 return 0;
267}
268EXPORT_SYMBOL(arch_spin_trylock_retry);
269
270void arch_read_lock_wait(arch_rwlock_t *rw)
271{
272 if (unlikely(in_interrupt())) {
273 while (READ_ONCE(rw->cnts) & 0x10000)
274 barrier();
275 return;
276 }
277
278 /* Remove this reader again to allow recursive read locking */
279 __atomic_add_const(-1, &rw->cnts);
280 /* Put the reader into the wait queue */
281 arch_spin_lock(&rw->wait);
282 /* Now add this reader to the count value again */
283 __atomic_add_const(1, &rw->cnts);
284 /* Loop until the writer is done */
285 while (READ_ONCE(rw->cnts) & 0x10000)
286 barrier();
287 arch_spin_unlock(&rw->wait);
288}
289EXPORT_SYMBOL(arch_read_lock_wait);
290
291void arch_write_lock_wait(arch_rwlock_t *rw)
292{
293 int old;
294
295 /* Add this CPU to the write waiters */
296 __atomic_add(0x20000, &rw->cnts);
297
298 /* Put the writer into the wait queue */
299 arch_spin_lock(&rw->wait);
300
301 while (1) {
302 old = READ_ONCE(rw->cnts);
303 if ((old & 0x1ffff) == 0 &&
304 __atomic_cmpxchg_bool(&rw->cnts, old, old | 0x10000))
305 /* Got the lock */
306 break;
307 barrier();
308 }
309
310 arch_spin_unlock(&rw->wait);
311}
312EXPORT_SYMBOL(arch_write_lock_wait);
313
314void arch_spin_relax(arch_spinlock_t *lp)
315{
316 int cpu;
317
318 cpu = READ_ONCE(lp->lock) & _Q_LOCK_CPU_MASK;
319 if (!cpu)
320 return;
321 if (MACHINE_IS_LPAR && !arch_vcpu_is_preempted(cpu - 1))
322 return;
323 smp_yield_cpu(cpu - 1);
324}
325EXPORT_SYMBOL(arch_spin_relax);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Out of line spinlock code.
4 *
5 * Copyright IBM Corp. 2004, 2006
6 * Author(s): Martin Schwidefsky (schwidefsky@de.ibm.com)
7 */
8
9#include <linux/types.h>
10#include <linux/export.h>
11#include <linux/spinlock.h>
12#include <linux/jiffies.h>
13#include <linux/init.h>
14#include <linux/smp.h>
15#include <linux/percpu.h>
16#include <asm/alternative.h>
17#include <asm/io.h>
18
19int spin_retry = -1;
20
21static int __init spin_retry_init(void)
22{
23 if (spin_retry < 0)
24 spin_retry = 1000;
25 return 0;
26}
27early_initcall(spin_retry_init);
28
29/**
30 * spin_retry= parameter
31 */
32static int __init spin_retry_setup(char *str)
33{
34 spin_retry = simple_strtoul(str, &str, 0);
35 return 1;
36}
37__setup("spin_retry=", spin_retry_setup);
38
39struct spin_wait {
40 struct spin_wait *next, *prev;
41 int node_id;
42} __aligned(32);
43
44static DEFINE_PER_CPU_ALIGNED(struct spin_wait, spin_wait[4]);
45
46#define _Q_LOCK_CPU_OFFSET 0
47#define _Q_LOCK_STEAL_OFFSET 16
48#define _Q_TAIL_IDX_OFFSET 18
49#define _Q_TAIL_CPU_OFFSET 20
50
51#define _Q_LOCK_CPU_MASK 0x0000ffff
52#define _Q_LOCK_STEAL_ADD 0x00010000
53#define _Q_LOCK_STEAL_MASK 0x00030000
54#define _Q_TAIL_IDX_MASK 0x000c0000
55#define _Q_TAIL_CPU_MASK 0xfff00000
56
57#define _Q_LOCK_MASK (_Q_LOCK_CPU_MASK | _Q_LOCK_STEAL_MASK)
58#define _Q_TAIL_MASK (_Q_TAIL_IDX_MASK | _Q_TAIL_CPU_MASK)
59
60void arch_spin_lock_setup(int cpu)
61{
62 struct spin_wait *node;
63 int ix;
64
65 node = per_cpu_ptr(&spin_wait[0], cpu);
66 for (ix = 0; ix < 4; ix++, node++) {
67 memset(node, 0, sizeof(*node));
68 node->node_id = ((cpu + 1) << _Q_TAIL_CPU_OFFSET) +
69 (ix << _Q_TAIL_IDX_OFFSET);
70 }
71}
72
73static inline int arch_load_niai4(int *lock)
74{
75 int owner;
76
77 asm_inline volatile(
78 ALTERNATIVE("", ".long 0xb2fa0040", 49) /* NIAI 4 */
79 " l %0,%1\n"
80 : "=d" (owner) : "Q" (*lock) : "memory");
81 return owner;
82}
83
84static inline int arch_cmpxchg_niai8(int *lock, int old, int new)
85{
86 int expected = old;
87
88 asm_inline volatile(
89 ALTERNATIVE("", ".long 0xb2fa0080", 49) /* NIAI 8 */
90 " cs %0,%3,%1\n"
91 : "=d" (old), "=Q" (*lock)
92 : "0" (old), "d" (new), "Q" (*lock)
93 : "cc", "memory");
94 return expected == old;
95}
96
97static inline struct spin_wait *arch_spin_decode_tail(int lock)
98{
99 int ix, cpu;
100
101 ix = (lock & _Q_TAIL_IDX_MASK) >> _Q_TAIL_IDX_OFFSET;
102 cpu = (lock & _Q_TAIL_CPU_MASK) >> _Q_TAIL_CPU_OFFSET;
103 return per_cpu_ptr(&spin_wait[ix], cpu - 1);
104}
105
106static inline int arch_spin_yield_target(int lock, struct spin_wait *node)
107{
108 if (lock & _Q_LOCK_CPU_MASK)
109 return lock & _Q_LOCK_CPU_MASK;
110 if (node == NULL || node->prev == NULL)
111 return 0; /* 0 -> no target cpu */
112 while (node->prev)
113 node = node->prev;
114 return node->node_id >> _Q_TAIL_CPU_OFFSET;
115}
116
117static inline void arch_spin_lock_queued(arch_spinlock_t *lp)
118{
119 struct spin_wait *node, *next;
120 int lockval, ix, node_id, tail_id, old, new, owner, count;
121
122 ix = S390_lowcore.spinlock_index++;
123 barrier();
124 lockval = SPINLOCK_LOCKVAL; /* cpu + 1 */
125 node = this_cpu_ptr(&spin_wait[ix]);
126 node->prev = node->next = NULL;
127 node_id = node->node_id;
128
129 /* Enqueue the node for this CPU in the spinlock wait queue */
130 while (1) {
131 old = READ_ONCE(lp->lock);
132 if ((old & _Q_LOCK_CPU_MASK) == 0 &&
133 (old & _Q_LOCK_STEAL_MASK) != _Q_LOCK_STEAL_MASK) {
134 /*
135 * The lock is free but there may be waiters.
136 * With no waiters simply take the lock, if there
137 * are waiters try to steal the lock. The lock may
138 * be stolen three times before the next queued
139 * waiter will get the lock.
140 */
141 new = (old ? (old + _Q_LOCK_STEAL_ADD) : 0) | lockval;
142 if (__atomic_cmpxchg_bool(&lp->lock, old, new))
143 /* Got the lock */
144 goto out;
145 /* lock passing in progress */
146 continue;
147 }
148 /* Make the node of this CPU the new tail. */
149 new = node_id | (old & _Q_LOCK_MASK);
150 if (__atomic_cmpxchg_bool(&lp->lock, old, new))
151 break;
152 }
153 /* Set the 'next' pointer of the tail node in the queue */
154 tail_id = old & _Q_TAIL_MASK;
155 if (tail_id != 0) {
156 node->prev = arch_spin_decode_tail(tail_id);
157 WRITE_ONCE(node->prev->next, node);
158 }
159
160 /* Pass the virtual CPU to the lock holder if it is not running */
161 owner = arch_spin_yield_target(old, node);
162 if (owner && arch_vcpu_is_preempted(owner - 1))
163 smp_yield_cpu(owner - 1);
164
165 /* Spin on the CPU local node->prev pointer */
166 if (tail_id != 0) {
167 count = spin_retry;
168 while (READ_ONCE(node->prev) != NULL) {
169 if (count-- >= 0)
170 continue;
171 count = spin_retry;
172 /* Query running state of lock holder again. */
173 owner = arch_spin_yield_target(old, node);
174 if (owner && arch_vcpu_is_preempted(owner - 1))
175 smp_yield_cpu(owner - 1);
176 }
177 }
178
179 /* Spin on the lock value in the spinlock_t */
180 count = spin_retry;
181 while (1) {
182 old = READ_ONCE(lp->lock);
183 owner = old & _Q_LOCK_CPU_MASK;
184 if (!owner) {
185 tail_id = old & _Q_TAIL_MASK;
186 new = ((tail_id != node_id) ? tail_id : 0) | lockval;
187 if (__atomic_cmpxchg_bool(&lp->lock, old, new))
188 /* Got the lock */
189 break;
190 continue;
191 }
192 if (count-- >= 0)
193 continue;
194 count = spin_retry;
195 if (!MACHINE_IS_LPAR || arch_vcpu_is_preempted(owner - 1))
196 smp_yield_cpu(owner - 1);
197 }
198
199 /* Pass lock_spin job to next CPU in the queue */
200 if (node_id && tail_id != node_id) {
201 /* Wait until the next CPU has set up the 'next' pointer */
202 while ((next = READ_ONCE(node->next)) == NULL)
203 ;
204 next->prev = NULL;
205 }
206
207 out:
208 S390_lowcore.spinlock_index--;
209}
210
211static inline void arch_spin_lock_classic(arch_spinlock_t *lp)
212{
213 int lockval, old, new, owner, count;
214
215 lockval = SPINLOCK_LOCKVAL; /* cpu + 1 */
216
217 /* Pass the virtual CPU to the lock holder if it is not running */
218 owner = arch_spin_yield_target(READ_ONCE(lp->lock), NULL);
219 if (owner && arch_vcpu_is_preempted(owner - 1))
220 smp_yield_cpu(owner - 1);
221
222 count = spin_retry;
223 while (1) {
224 old = arch_load_niai4(&lp->lock);
225 owner = old & _Q_LOCK_CPU_MASK;
226 /* Try to get the lock if it is free. */
227 if (!owner) {
228 new = (old & _Q_TAIL_MASK) | lockval;
229 if (arch_cmpxchg_niai8(&lp->lock, old, new)) {
230 /* Got the lock */
231 return;
232 }
233 continue;
234 }
235 if (count-- >= 0)
236 continue;
237 count = spin_retry;
238 if (!MACHINE_IS_LPAR || arch_vcpu_is_preempted(owner - 1))
239 smp_yield_cpu(owner - 1);
240 }
241}
242
243void arch_spin_lock_wait(arch_spinlock_t *lp)
244{
245 if (test_cpu_flag(CIF_DEDICATED_CPU))
246 arch_spin_lock_queued(lp);
247 else
248 arch_spin_lock_classic(lp);
249}
250EXPORT_SYMBOL(arch_spin_lock_wait);
251
252int arch_spin_trylock_retry(arch_spinlock_t *lp)
253{
254 int cpu = SPINLOCK_LOCKVAL;
255 int owner, count;
256
257 for (count = spin_retry; count > 0; count--) {
258 owner = READ_ONCE(lp->lock);
259 /* Try to get the lock if it is free. */
260 if (!owner) {
261 if (__atomic_cmpxchg_bool(&lp->lock, 0, cpu))
262 return 1;
263 }
264 }
265 return 0;
266}
267EXPORT_SYMBOL(arch_spin_trylock_retry);
268
269void arch_read_lock_wait(arch_rwlock_t *rw)
270{
271 if (unlikely(in_interrupt())) {
272 while (READ_ONCE(rw->cnts) & 0x10000)
273 barrier();
274 return;
275 }
276
277 /* Remove this reader again to allow recursive read locking */
278 __atomic_add_const(-1, &rw->cnts);
279 /* Put the reader into the wait queue */
280 arch_spin_lock(&rw->wait);
281 /* Now add this reader to the count value again */
282 __atomic_add_const(1, &rw->cnts);
283 /* Loop until the writer is done */
284 while (READ_ONCE(rw->cnts) & 0x10000)
285 barrier();
286 arch_spin_unlock(&rw->wait);
287}
288EXPORT_SYMBOL(arch_read_lock_wait);
289
290void arch_write_lock_wait(arch_rwlock_t *rw)
291{
292 int old;
293
294 /* Add this CPU to the write waiters */
295 __atomic_add(0x20000, &rw->cnts);
296
297 /* Put the writer into the wait queue */
298 arch_spin_lock(&rw->wait);
299
300 while (1) {
301 old = READ_ONCE(rw->cnts);
302 if ((old & 0x1ffff) == 0 &&
303 __atomic_cmpxchg_bool(&rw->cnts, old, old | 0x10000))
304 /* Got the lock */
305 break;
306 barrier();
307 }
308
309 arch_spin_unlock(&rw->wait);
310}
311EXPORT_SYMBOL(arch_write_lock_wait);
312
313void arch_spin_relax(arch_spinlock_t *lp)
314{
315 int cpu;
316
317 cpu = READ_ONCE(lp->lock) & _Q_LOCK_CPU_MASK;
318 if (!cpu)
319 return;
320 if (MACHINE_IS_LPAR && !arch_vcpu_is_preempted(cpu - 1))
321 return;
322 smp_yield_cpu(cpu - 1);
323}
324EXPORT_SYMBOL(arch_spin_relax);