Merge branch 'next' into for-linus
[firefly-linux-kernel-4.4.55.git] / arch / arm / include / asm / spinlock.h
1 #ifndef __ASM_SPINLOCK_H
2 #define __ASM_SPINLOCK_H
3
4 #if __LINUX_ARM_ARCH__ < 6
5 #error SMP not supported on pre-ARMv6 CPUs
6 #endif
7
8 #include <asm/processor.h>
9
10 /*
11  * sev and wfe are ARMv6K extensions.  Uniprocessor ARMv6 may not have the K
12  * extensions, so when running on UP, we have to patch these instructions away.
13  */
14 #define ALT_SMP(smp, up)                                        \
15         "9998:  " smp "\n"                                      \
16         "       .pushsection \".alt.smp.init\", \"a\"\n"        \
17         "       .long   9998b\n"                                \
18         "       " up "\n"                                       \
19         "       .popsection\n"
20
21 #ifdef CONFIG_THUMB2_KERNEL
22 #define SEV             ALT_SMP("sev.w", "nop.w")
23 /*
24  * For Thumb-2, special care is needed to ensure that the conditional WFE
25  * instruction really does assemble to exactly 4 bytes (as required by
26  * the SMP_ON_UP fixup code).   By itself "wfene" might cause the
27  * assembler to insert a extra (16-bit) IT instruction, depending on the
28  * presence or absence of neighbouring conditional instructions.
29  *
30  * To avoid this unpredictableness, an approprite IT is inserted explicitly:
31  * the assembler won't change IT instructions which are explicitly present
32  * in the input.
33  */
34 #define WFE(cond)       ALT_SMP(                \
35         "it " cond "\n\t"                       \
36         "wfe" cond ".n",                        \
37                                                 \
38         "nop.w"                                 \
39 )
40 #else
41 #define SEV             ALT_SMP("sev", "nop")
42 #define WFE(cond)       ALT_SMP("wfe" cond, "nop")
43 #endif
44
45 static inline void dsb_sev(void)
46 {
47 #if __LINUX_ARM_ARCH__ >= 7
48         __asm__ __volatile__ (
49                 "dsb\n"
50                 SEV
51         );
52 #else
53         __asm__ __volatile__ (
54                 "mcr p15, 0, %0, c7, c10, 4\n"
55                 SEV
56                 : : "r" (0)
57         );
58 #endif
59 }
60
61 /*
62  * ARMv6 ticket-based spin-locking.
63  *
64  * A memory barrier is required after we get a lock, and before we
65  * release it, because V6 CPUs are assumed to have weakly ordered
66  * memory.
67  */
68
69 #define arch_spin_unlock_wait(lock) \
70         do { while (arch_spin_is_locked(lock)) cpu_relax(); } while (0)
71
72 #define arch_spin_lock_flags(lock, flags) arch_spin_lock(lock)
73
74 static inline void arch_spin_lock(arch_spinlock_t *lock)
75 {
76         unsigned long tmp;
77         u32 newval;
78         arch_spinlock_t lockval;
79
80         __asm__ __volatile__(
81 "1:     ldrex   %0, [%3]\n"
82 "       add     %1, %0, %4\n"
83 "       strex   %2, %1, [%3]\n"
84 "       teq     %2, #0\n"
85 "       bne     1b"
86         : "=&r" (lockval), "=&r" (newval), "=&r" (tmp)
87         : "r" (&lock->slock), "I" (1 << TICKET_SHIFT)
88         : "cc");
89
90         while (lockval.tickets.next != lockval.tickets.owner) {
91                 wfe();
92                 lockval.tickets.owner = ACCESS_ONCE(lock->tickets.owner);
93         }
94
95         smp_mb();
96 }
97
98 static inline int arch_spin_trylock(arch_spinlock_t *lock)
99 {
100         unsigned long tmp;
101         u32 slock;
102
103         __asm__ __volatile__(
104 "       ldrex   %0, [%2]\n"
105 "       subs    %1, %0, %0, ror #16\n"
106 "       addeq   %0, %0, %3\n"
107 "       strexeq %1, %0, [%2]"
108         : "=&r" (slock), "=&r" (tmp)
109         : "r" (&lock->slock), "I" (1 << TICKET_SHIFT)
110         : "cc");
111
112         if (tmp == 0) {
113                 smp_mb();
114                 return 1;
115         } else {
116                 return 0;
117         }
118 }
119
120 static inline void arch_spin_unlock(arch_spinlock_t *lock)
121 {
122         smp_mb();
123         lock->tickets.owner++;
124         dsb_sev();
125 }
126
127 static inline int arch_spin_is_locked(arch_spinlock_t *lock)
128 {
129         struct __raw_tickets tickets = ACCESS_ONCE(lock->tickets);
130         return tickets.owner != tickets.next;
131 }
132
133 static inline int arch_spin_is_contended(arch_spinlock_t *lock)
134 {
135         struct __raw_tickets tickets = ACCESS_ONCE(lock->tickets);
136         return (tickets.next - tickets.owner) > 1;
137 }
138 #define arch_spin_is_contended  arch_spin_is_contended
139
140 /*
141  * RWLOCKS
142  *
143  *
144  * Write locks are easy - we just set bit 31.  When unlocking, we can
145  * just write zero since the lock is exclusively held.
146  */
147
148 static inline void arch_write_lock(arch_rwlock_t *rw)
149 {
150         unsigned long tmp;
151
152         __asm__ __volatile__(
153 "1:     ldrex   %0, [%1]\n"
154 "       teq     %0, #0\n"
155         WFE("ne")
156 "       strexeq %0, %2, [%1]\n"
157 "       teq     %0, #0\n"
158 "       bne     1b"
159         : "=&r" (tmp)
160         : "r" (&rw->lock), "r" (0x80000000)
161         : "cc");
162
163         smp_mb();
164 }
165
166 static inline int arch_write_trylock(arch_rwlock_t *rw)
167 {
168         unsigned long tmp;
169
170         __asm__ __volatile__(
171 "       ldrex   %0, [%1]\n"
172 "       teq     %0, #0\n"
173 "       strexeq %0, %2, [%1]"
174         : "=&r" (tmp)
175         : "r" (&rw->lock), "r" (0x80000000)
176         : "cc");
177
178         if (tmp == 0) {
179                 smp_mb();
180                 return 1;
181         } else {
182                 return 0;
183         }
184 }
185
186 static inline void arch_write_unlock(arch_rwlock_t *rw)
187 {
188         smp_mb();
189
190         __asm__ __volatile__(
191         "str    %1, [%0]\n"
192         :
193         : "r" (&rw->lock), "r" (0)
194         : "cc");
195
196         dsb_sev();
197 }
198
199 /* write_can_lock - would write_trylock() succeed? */
200 #define arch_write_can_lock(x)          ((x)->lock == 0)
201
202 /*
203  * Read locks are a bit more hairy:
204  *  - Exclusively load the lock value.
205  *  - Increment it.
206  *  - Store new lock value if positive, and we still own this location.
207  *    If the value is negative, we've already failed.
208  *  - If we failed to store the value, we want a negative result.
209  *  - If we failed, try again.
210  * Unlocking is similarly hairy.  We may have multiple read locks
211  * currently active.  However, we know we won't have any write
212  * locks.
213  */
214 static inline void arch_read_lock(arch_rwlock_t *rw)
215 {
216         unsigned long tmp, tmp2;
217
218         __asm__ __volatile__(
219 "1:     ldrex   %0, [%2]\n"
220 "       adds    %0, %0, #1\n"
221 "       strexpl %1, %0, [%2]\n"
222         WFE("mi")
223 "       rsbpls  %0, %1, #0\n"
224 "       bmi     1b"
225         : "=&r" (tmp), "=&r" (tmp2)
226         : "r" (&rw->lock)
227         : "cc");
228
229         smp_mb();
230 }
231
232 static inline void arch_read_unlock(arch_rwlock_t *rw)
233 {
234         unsigned long tmp, tmp2;
235
236         smp_mb();
237
238         __asm__ __volatile__(
239 "1:     ldrex   %0, [%2]\n"
240 "       sub     %0, %0, #1\n"
241 "       strex   %1, %0, [%2]\n"
242 "       teq     %1, #0\n"
243 "       bne     1b"
244         : "=&r" (tmp), "=&r" (tmp2)
245         : "r" (&rw->lock)
246         : "cc");
247
248         if (tmp == 0)
249                 dsb_sev();
250 }
251
252 static inline int arch_read_trylock(arch_rwlock_t *rw)
253 {
254         unsigned long tmp, tmp2 = 1;
255
256         __asm__ __volatile__(
257 "       ldrex   %0, [%2]\n"
258 "       adds    %0, %0, #1\n"
259 "       strexpl %1, %0, [%2]\n"
260         : "=&r" (tmp), "+r" (tmp2)
261         : "r" (&rw->lock)
262         : "cc");
263
264         smp_mb();
265         return tmp2 == 0;
266 }
267
268 /* read_can_lock - would read_trylock() succeed? */
269 #define arch_read_can_lock(x)           ((x)->lock < 0x80000000)
270
271 #define arch_read_lock_flags(lock, flags) arch_read_lock(lock)
272 #define arch_write_lock_flags(lock, flags) arch_write_lock(lock)
273
274 #define arch_spin_relax(lock)   cpu_relax()
275 #define arch_read_relax(lock)   cpu_relax()
276 #define arch_write_relax(lock)  cpu_relax()
277
278 #endif /* __ASM_SPINLOCK_H */