Commit | Line | Data |
---|---|---|
867e359b CM |
1 | /* |
2 | * Copyright 2010 Tilera Corporation. All Rights Reserved. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or | |
5 | * modify it under the terms of the GNU General Public License | |
6 | * as published by the Free Software Foundation, version 2. | |
7 | * | |
8 | * This program is distributed in the hope that it will be useful, but | |
9 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
10 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or | |
11 | * NON INFRINGEMENT. See the GNU General Public License for | |
12 | * more details. | |
13 | */ | |
14 | ||
15 | #include <linux/cache.h> | |
16 | #include <linux/delay.h> | |
17 | #include <linux/uaccess.h> | |
18 | #include <linux/module.h> | |
19 | #include <linux/mm.h> | |
20 | #include <asm/atomic.h> | |
0707ad30 | 21 | #include <asm/futex.h> |
867e359b CM |
22 | #include <arch/chip.h> |
23 | ||
0707ad30 | 24 | /* See <asm/atomic_32.h> */ |
867e359b CM |
25 | #if ATOMIC_LOCKS_FOUND_VIA_TABLE() |
26 | ||
27 | /* | |
28 | * A block of memory containing locks for atomic ops. Each instance of this | |
29 | * struct will be homed on a different CPU. | |
30 | */ | |
31 | struct atomic_locks_on_cpu { | |
32 | int lock[ATOMIC_HASH_L2_SIZE]; | |
33 | } __attribute__((aligned(ATOMIC_HASH_L2_SIZE * 4))); | |
34 | ||
35 | static DEFINE_PER_CPU(struct atomic_locks_on_cpu, atomic_lock_pool); | |
36 | ||
37 | /* The locks we'll use until __init_atomic_per_cpu is called. */ | |
38 | static struct atomic_locks_on_cpu __initdata initial_atomic_locks; | |
39 | ||
40 | /* Hash into this vector to get a pointer to lock for the given atomic. */ | |
41 | struct atomic_locks_on_cpu *atomic_lock_ptr[ATOMIC_HASH_L1_SIZE] | |
42 | __write_once = { | |
43 | [0 ... ATOMIC_HASH_L1_SIZE-1] (&initial_atomic_locks) | |
44 | }; | |
45 | ||
46 | #else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ | |
47 | ||
48 | /* This page is remapped on startup to be hash-for-home. */ | |
49 | int atomic_locks[PAGE_SIZE / sizeof(int) /* Only ATOMIC_HASH_SIZE is used */] | |
50 | __attribute__((aligned(PAGE_SIZE), section(".bss.page_aligned"))); | |
51 | ||
52 | #endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ | |
53 | ||
54 | static inline int *__atomic_hashed_lock(volatile void *v) | |
55 | { | |
56 | /* NOTE: this code must match "sys_cmpxchg" in kernel/intvec.S */ | |
57 | #if ATOMIC_LOCKS_FOUND_VIA_TABLE() | |
58 | unsigned long i = | |
59 | (unsigned long) v & ((PAGE_SIZE-1) & -sizeof(long long)); | |
60 | unsigned long n = __insn_crc32_32(0, i); | |
61 | ||
62 | /* Grab high bits for L1 index. */ | |
63 | unsigned long l1_index = n >> ((sizeof(n) * 8) - ATOMIC_HASH_L1_SHIFT); | |
64 | /* Grab low bits for L2 index. */ | |
65 | unsigned long l2_index = n & (ATOMIC_HASH_L2_SIZE - 1); | |
66 | ||
67 | return &atomic_lock_ptr[l1_index]->lock[l2_index]; | |
68 | #else | |
69 | /* | |
70 | * Use bits [3, 3 + ATOMIC_HASH_SHIFT) as the lock index. | |
71 | * Using mm works here because atomic_locks is page aligned. | |
72 | */ | |
73 | unsigned long ptr = __insn_mm((unsigned long)v >> 1, | |
74 | (unsigned long)atomic_locks, | |
75 | 2, (ATOMIC_HASH_SHIFT + 2) - 1); | |
76 | return (int *)ptr; | |
77 | #endif | |
78 | } | |
79 | ||
80 | #ifdef CONFIG_SMP | |
81 | /* Return whether the passed pointer is a valid atomic lock pointer. */ | |
82 | static int is_atomic_lock(int *p) | |
83 | { | |
84 | #if ATOMIC_LOCKS_FOUND_VIA_TABLE() | |
85 | int i; | |
86 | for (i = 0; i < ATOMIC_HASH_L1_SIZE; ++i) { | |
87 | ||
88 | if (p >= &atomic_lock_ptr[i]->lock[0] && | |
89 | p < &atomic_lock_ptr[i]->lock[ATOMIC_HASH_L2_SIZE]) { | |
90 | return 1; | |
91 | } | |
92 | } | |
93 | return 0; | |
94 | #else | |
95 | return p >= &atomic_locks[0] && p < &atomic_locks[ATOMIC_HASH_SIZE]; | |
96 | #endif | |
97 | } | |
98 | ||
99 | void __atomic_fault_unlock(int *irqlock_word) | |
100 | { | |
101 | BUG_ON(!is_atomic_lock(irqlock_word)); | |
102 | BUG_ON(*irqlock_word != 1); | |
103 | *irqlock_word = 0; | |
104 | } | |
105 | ||
106 | #endif /* CONFIG_SMP */ | |
107 | ||
108 | static inline int *__atomic_setup(volatile void *v) | |
109 | { | |
110 | /* Issue a load to the target to bring it into cache. */ | |
111 | *(volatile int *)v; | |
112 | return __atomic_hashed_lock(v); | |
113 | } | |
114 | ||
115 | int _atomic_xchg(atomic_t *v, int n) | |
116 | { | |
117 | return __atomic_xchg(&v->counter, __atomic_setup(v), n).val; | |
118 | } | |
119 | EXPORT_SYMBOL(_atomic_xchg); | |
120 | ||
121 | int _atomic_xchg_add(atomic_t *v, int i) | |
122 | { | |
123 | return __atomic_xchg_add(&v->counter, __atomic_setup(v), i).val; | |
124 | } | |
125 | EXPORT_SYMBOL(_atomic_xchg_add); | |
126 | ||
127 | int _atomic_xchg_add_unless(atomic_t *v, int a, int u) | |
128 | { | |
129 | /* | |
130 | * Note: argument order is switched here since it is easier | |
131 | * to use the first argument consistently as the "old value" | |
132 | * in the assembly, as is done for _atomic_cmpxchg(). | |
133 | */ | |
134 | return __atomic_xchg_add_unless(&v->counter, __atomic_setup(v), u, a) | |
135 | .val; | |
136 | } | |
137 | EXPORT_SYMBOL(_atomic_xchg_add_unless); | |
138 | ||
139 | int _atomic_cmpxchg(atomic_t *v, int o, int n) | |
140 | { | |
141 | return __atomic_cmpxchg(&v->counter, __atomic_setup(v), o, n).val; | |
142 | } | |
143 | EXPORT_SYMBOL(_atomic_cmpxchg); | |
144 | ||
145 | unsigned long _atomic_or(volatile unsigned long *p, unsigned long mask) | |
146 | { | |
147 | return __atomic_or((int *)p, __atomic_setup(p), mask).val; | |
148 | } | |
149 | EXPORT_SYMBOL(_atomic_or); | |
150 | ||
151 | unsigned long _atomic_andn(volatile unsigned long *p, unsigned long mask) | |
152 | { | |
153 | return __atomic_andn((int *)p, __atomic_setup(p), mask).val; | |
154 | } | |
155 | EXPORT_SYMBOL(_atomic_andn); | |
156 | ||
157 | unsigned long _atomic_xor(volatile unsigned long *p, unsigned long mask) | |
158 | { | |
159 | return __atomic_xor((int *)p, __atomic_setup(p), mask).val; | |
160 | } | |
161 | EXPORT_SYMBOL(_atomic_xor); | |
162 | ||
163 | ||
164 | u64 _atomic64_xchg(atomic64_t *v, u64 n) | |
165 | { | |
166 | return __atomic64_xchg(&v->counter, __atomic_setup(v), n); | |
167 | } | |
168 | EXPORT_SYMBOL(_atomic64_xchg); | |
169 | ||
170 | u64 _atomic64_xchg_add(atomic64_t *v, u64 i) | |
171 | { | |
172 | return __atomic64_xchg_add(&v->counter, __atomic_setup(v), i); | |
173 | } | |
174 | EXPORT_SYMBOL(_atomic64_xchg_add); | |
175 | ||
176 | u64 _atomic64_xchg_add_unless(atomic64_t *v, u64 a, u64 u) | |
177 | { | |
178 | /* | |
179 | * Note: argument order is switched here since it is easier | |
180 | * to use the first argument consistently as the "old value" | |
181 | * in the assembly, as is done for _atomic_cmpxchg(). | |
182 | */ | |
183 | return __atomic64_xchg_add_unless(&v->counter, __atomic_setup(v), | |
184 | u, a); | |
185 | } | |
186 | EXPORT_SYMBOL(_atomic64_xchg_add_unless); | |
187 | ||
188 | u64 _atomic64_cmpxchg(atomic64_t *v, u64 o, u64 n) | |
189 | { | |
190 | return __atomic64_cmpxchg(&v->counter, __atomic_setup(v), o, n); | |
191 | } | |
192 | EXPORT_SYMBOL(_atomic64_cmpxchg); | |
193 | ||
194 | ||
0707ad30 | 195 | static inline int *__futex_setup(int __user *v) |
867e359b CM |
196 | { |
197 | /* | |
198 | * Issue a prefetch to the counter to bring it into cache. | |
199 | * As for __atomic_setup, but we can't do a read into the L1 | |
200 | * since it might fault; instead we do a prefetch into the L2. | |
201 | */ | |
202 | __insn_prefetch(v); | |
0707ad30 | 203 | return __atomic_hashed_lock((int __force *)v); |
867e359b CM |
204 | } |
205 | ||
0707ad30 | 206 | struct __get_user futex_set(int __user *v, int i) |
867e359b | 207 | { |
0707ad30 | 208 | return __atomic_xchg((int __force *)v, __futex_setup(v), i); |
867e359b CM |
209 | } |
210 | ||
0707ad30 | 211 | struct __get_user futex_add(int __user *v, int n) |
867e359b | 212 | { |
0707ad30 | 213 | return __atomic_xchg_add((int __force *)v, __futex_setup(v), n); |
867e359b CM |
214 | } |
215 | ||
0707ad30 | 216 | struct __get_user futex_or(int __user *v, int n) |
867e359b | 217 | { |
0707ad30 | 218 | return __atomic_or((int __force *)v, __futex_setup(v), n); |
867e359b CM |
219 | } |
220 | ||
0707ad30 | 221 | struct __get_user futex_andn(int __user *v, int n) |
867e359b | 222 | { |
0707ad30 | 223 | return __atomic_andn((int __force *)v, __futex_setup(v), n); |
867e359b CM |
224 | } |
225 | ||
0707ad30 | 226 | struct __get_user futex_xor(int __user *v, int n) |
867e359b | 227 | { |
0707ad30 | 228 | return __atomic_xor((int __force *)v, __futex_setup(v), n); |
867e359b CM |
229 | } |
230 | ||
0707ad30 | 231 | struct __get_user futex_cmpxchg(int __user *v, int o, int n) |
867e359b | 232 | { |
0707ad30 | 233 | return __atomic_cmpxchg((int __force *)v, __futex_setup(v), o, n); |
867e359b CM |
234 | } |
235 | ||
236 | /* | |
237 | * If any of the atomic or futex routines hit a bad address (not in | |
238 | * the page tables at kernel PL) this routine is called. The futex | |
239 | * routines are never used on kernel space, and the normal atomics and | |
240 | * bitops are never used on user space. So a fault on kernel space | |
241 | * must be fatal, but a fault on userspace is a futex fault and we | |
242 | * need to return -EFAULT. Note that the context this routine is | |
243 | * invoked in is the context of the "_atomic_xxx()" routines called | |
244 | * by the functions in this file. | |
245 | */ | |
0707ad30 | 246 | struct __get_user __atomic_bad_address(int __user *addr) |
867e359b CM |
247 | { |
248 | if (unlikely(!access_ok(VERIFY_WRITE, addr, sizeof(int)))) | |
249 | panic("Bad address used for kernel atomic op: %p\n", addr); | |
250 | return (struct __get_user) { .err = -EFAULT }; | |
251 | } | |
252 | ||
253 | ||
254 | #if CHIP_HAS_CBOX_HOME_MAP() | |
255 | static int __init noatomichash(char *str) | |
256 | { | |
0707ad30 | 257 | pr_warning("noatomichash is deprecated.\n"); |
867e359b CM |
258 | return 1; |
259 | } | |
260 | __setup("noatomichash", noatomichash); | |
261 | #endif | |
262 | ||
263 | void __init __init_atomic_per_cpu(void) | |
264 | { | |
265 | #if ATOMIC_LOCKS_FOUND_VIA_TABLE() | |
266 | ||
267 | unsigned int i; | |
268 | int actual_cpu; | |
269 | ||
270 | /* | |
271 | * Before this is called from setup, we just have one lock for | |
272 | * all atomic objects/operations. Here we replace the | |
273 | * elements of atomic_lock_ptr so that they point at per_cpu | |
274 | * integers. This seemingly over-complex approach stems from | |
275 | * the fact that DEFINE_PER_CPU defines an entry for each cpu | |
276 | * in the grid, not each cpu from 0..ATOMIC_HASH_SIZE-1. But | |
277 | * for efficient hashing of atomics to their locks we want a | |
278 | * compile time constant power of 2 for the size of this | |
279 | * table, so we use ATOMIC_HASH_SIZE. | |
280 | * | |
281 | * Here we populate atomic_lock_ptr from the per cpu | |
282 | * atomic_lock_pool, interspersing by actual cpu so that | |
283 | * subsequent elements are homed on consecutive cpus. | |
284 | */ | |
285 | ||
286 | actual_cpu = cpumask_first(cpu_possible_mask); | |
287 | ||
288 | for (i = 0; i < ATOMIC_HASH_L1_SIZE; ++i) { | |
289 | /* | |
290 | * Preincrement to slightly bias against using cpu 0, | |
291 | * which has plenty of stuff homed on it already. | |
292 | */ | |
293 | actual_cpu = cpumask_next(actual_cpu, cpu_possible_mask); | |
294 | if (actual_cpu >= nr_cpu_ids) | |
295 | actual_cpu = cpumask_first(cpu_possible_mask); | |
296 | ||
297 | atomic_lock_ptr[i] = &per_cpu(atomic_lock_pool, actual_cpu); | |
298 | } | |
299 | ||
300 | #else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ | |
301 | ||
302 | /* Validate power-of-two and "bigger than cpus" assumption */ | |
303 | BUG_ON(ATOMIC_HASH_SIZE & (ATOMIC_HASH_SIZE-1)); | |
304 | BUG_ON(ATOMIC_HASH_SIZE < nr_cpu_ids); | |
305 | ||
306 | /* | |
307 | * On TILEPro we prefer to use a single hash-for-home | |
308 | * page, since this means atomic operations are less | |
309 | * likely to encounter a TLB fault and thus should | |
310 | * in general perform faster. You may wish to disable | |
311 | * this in situations where few hash-for-home tiles | |
312 | * are configured. | |
313 | */ | |
314 | BUG_ON((unsigned long)atomic_locks % PAGE_SIZE != 0); | |
315 | ||
316 | /* The locks must all fit on one page. */ | |
317 | BUG_ON(ATOMIC_HASH_SIZE * sizeof(int) > PAGE_SIZE); | |
318 | ||
319 | /* | |
320 | * We use the page offset of the atomic value's address as | |
321 | * an index into atomic_locks, excluding the low 3 bits. | |
322 | * That should not produce more indices than ATOMIC_HASH_SIZE. | |
323 | */ | |
324 | BUG_ON((PAGE_SIZE >> 3) > ATOMIC_HASH_SIZE); | |
325 | ||
326 | #endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */ | |
327 | ||
328 | /* The futex code makes this assumption, so we validate it here. */ | |
329 | BUG_ON(sizeof(atomic_t) != sizeof(int)); | |
330 | } |