diff options
Diffstat (limited to 'arch/tile/lib/spinlock_32.c')
| -rw-r--r-- | arch/tile/lib/spinlock_32.c | 221 |
1 files changed, 221 insertions, 0 deletions
diff --git a/arch/tile/lib/spinlock_32.c b/arch/tile/lib/spinlock_32.c new file mode 100644 index 000000000000..485e24d62c6b --- /dev/null +++ b/arch/tile/lib/spinlock_32.c | |||
| @@ -0,0 +1,221 @@ | |||
| 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/spinlock.h> | ||
| 16 | #include <linux/module.h> | ||
| 17 | #include <asm/processor.h> | ||
| 18 | |||
| 19 | #include "spinlock_common.h" | ||
| 20 | |||
| 21 | void arch_spin_lock(arch_spinlock_t *lock) | ||
| 22 | { | ||
| 23 | int my_ticket; | ||
| 24 | int iterations = 0; | ||
| 25 | int delta; | ||
| 26 | |||
| 27 | while ((my_ticket = __insn_tns((void *)&lock->next_ticket)) & 1) | ||
| 28 | delay_backoff(iterations++); | ||
| 29 | |||
| 30 | /* Increment the next ticket number, implicitly releasing tns lock. */ | ||
| 31 | lock->next_ticket = my_ticket + TICKET_QUANTUM; | ||
| 32 | |||
| 33 | /* Wait until it's our turn. */ | ||
| 34 | while ((delta = my_ticket - lock->current_ticket) != 0) | ||
| 35 | relax((128 / CYCLES_PER_RELAX_LOOP) * delta); | ||
| 36 | } | ||
| 37 | EXPORT_SYMBOL(arch_spin_lock); | ||
| 38 | |||
| 39 | int arch_spin_trylock(arch_spinlock_t *lock) | ||
| 40 | { | ||
| 41 | /* | ||
| 42 | * Grab a ticket; no need to retry if it's busy, we'll just | ||
| 43 | * treat that the same as "locked", since someone else | ||
| 44 | * will lock it momentarily anyway. | ||
| 45 | */ | ||
| 46 | int my_ticket = __insn_tns((void *)&lock->next_ticket); | ||
| 47 | |||
| 48 | if (my_ticket == lock->current_ticket) { | ||
| 49 | /* Not currently locked, so lock it by keeping this ticket. */ | ||
| 50 | lock->next_ticket = my_ticket + TICKET_QUANTUM; | ||
| 51 | /* Success! */ | ||
| 52 | return 1; | ||
| 53 | } | ||
| 54 | |||
| 55 | if (!(my_ticket & 1)) { | ||
| 56 | /* Release next_ticket. */ | ||
| 57 | lock->next_ticket = my_ticket; | ||
| 58 | } | ||
| 59 | |||
| 60 | return 0; | ||
| 61 | } | ||
| 62 | EXPORT_SYMBOL(arch_spin_trylock); | ||
| 63 | |||
| 64 | void arch_spin_unlock_wait(arch_spinlock_t *lock) | ||
| 65 | { | ||
| 66 | u32 iterations = 0; | ||
| 67 | while (arch_spin_is_locked(lock)) | ||
| 68 | delay_backoff(iterations++); | ||
| 69 | } | ||
| 70 | EXPORT_SYMBOL(arch_spin_unlock_wait); | ||
| 71 | |||
| 72 | /* | ||
| 73 | * The low byte is always reserved to be the marker for a "tns" operation | ||
| 74 | * since the low bit is set to "1" by a tns. The next seven bits are | ||
| 75 | * zeroes. The next byte holds the "next" writer value, i.e. the ticket | ||
| 76 | * available for the next task that wants to write. The third byte holds | ||
| 77 | * the current writer value, i.e. the writer who holds the current ticket. | ||
| 78 | * If current == next == 0, there are no interested writers. | ||
| 79 | */ | ||
| 80 | #define WR_NEXT_SHIFT _WR_NEXT_SHIFT | ||
| 81 | #define WR_CURR_SHIFT _WR_CURR_SHIFT | ||
| 82 | #define WR_WIDTH _WR_WIDTH | ||
| 83 | #define WR_MASK ((1 << WR_WIDTH) - 1) | ||
| 84 | |||
| 85 | /* | ||
| 86 | * The last eight bits hold the active reader count. This has to be | ||
| 87 | * zero before a writer can start to write. | ||
| 88 | */ | ||
| 89 | #define RD_COUNT_SHIFT _RD_COUNT_SHIFT | ||
| 90 | #define RD_COUNT_WIDTH _RD_COUNT_WIDTH | ||
| 91 | #define RD_COUNT_MASK ((1 << RD_COUNT_WIDTH) - 1) | ||
| 92 | |||
| 93 | |||
| 94 | /* Lock the word, spinning until there are no tns-ers. */ | ||
| 95 | static inline u32 get_rwlock(arch_rwlock_t *rwlock) | ||
| 96 | { | ||
| 97 | u32 iterations = 0; | ||
| 98 | for (;;) { | ||
| 99 | u32 val = __insn_tns((int *)&rwlock->lock); | ||
| 100 | if (unlikely(val & 1)) { | ||
| 101 | delay_backoff(iterations++); | ||
| 102 | continue; | ||
| 103 | } | ||
| 104 | return val; | ||
| 105 | } | ||
| 106 | } | ||
| 107 | |||
| 108 | int arch_read_trylock_slow(arch_rwlock_t *rwlock) | ||
| 109 | { | ||
| 110 | u32 val = get_rwlock(rwlock); | ||
| 111 | int locked = (val << RD_COUNT_WIDTH) == 0; | ||
| 112 | rwlock->lock = val + (locked << RD_COUNT_SHIFT); | ||
| 113 | return locked; | ||
| 114 | } | ||
| 115 | EXPORT_SYMBOL(arch_read_trylock_slow); | ||
| 116 | |||
| 117 | void arch_read_unlock_slow(arch_rwlock_t *rwlock) | ||
| 118 | { | ||
| 119 | u32 val = get_rwlock(rwlock); | ||
| 120 | rwlock->lock = val - (1 << RD_COUNT_SHIFT); | ||
| 121 | } | ||
| 122 | EXPORT_SYMBOL(arch_read_unlock_slow); | ||
| 123 | |||
| 124 | void arch_write_unlock_slow(arch_rwlock_t *rwlock, u32 val) | ||
| 125 | { | ||
| 126 | u32 eq, mask = 1 << WR_CURR_SHIFT; | ||
| 127 | while (unlikely(val & 1)) { | ||
| 128 | /* Limited backoff since we are the highest-priority task. */ | ||
| 129 | relax(4); | ||
| 130 | val = __insn_tns((int *)&rwlock->lock); | ||
| 131 | } | ||
| 132 | val = __insn_addb(val, mask); | ||
| 133 | eq = __insn_seqb(val, val << (WR_CURR_SHIFT - WR_NEXT_SHIFT)); | ||
| 134 | val = __insn_mz(eq & mask, val); | ||
| 135 | rwlock->lock = val; | ||
| 136 | } | ||
| 137 | EXPORT_SYMBOL(arch_write_unlock_slow); | ||
| 138 | |||
| 139 | /* | ||
| 140 | * We spin until everything but the reader bits (which are in the high | ||
| 141 | * part of the word) are zero, i.e. no active or waiting writers, no tns. | ||
| 142 | * | ||
| 143 | * ISSUE: This approach can permanently starve readers. A reader who sees | ||
| 144 | * a writer could instead take a ticket lock (just like a writer would), | ||
| 145 | * and atomically enter read mode (with 1 reader) when it gets the ticket. | ||
| 146 | * This way both readers and writers will always make forward progress | ||
| 147 | * in a finite time. | ||
| 148 | */ | ||
| 149 | void arch_read_lock_slow(arch_rwlock_t *rwlock, u32 val) | ||
| 150 | { | ||
| 151 | u32 iterations = 0; | ||
| 152 | do { | ||
| 153 | if (!(val & 1)) | ||
| 154 | rwlock->lock = val; | ||
| 155 | delay_backoff(iterations++); | ||
| 156 | val = __insn_tns((int *)&rwlock->lock); | ||
| 157 | } while ((val << RD_COUNT_WIDTH) != 0); | ||
| 158 | rwlock->lock = val + (1 << RD_COUNT_SHIFT); | ||
| 159 | } | ||
| 160 | EXPORT_SYMBOL(arch_read_lock_slow); | ||
| 161 | |||
| 162 | void arch_write_lock_slow(arch_rwlock_t *rwlock, u32 val) | ||
| 163 | { | ||
| 164 | /* | ||
| 165 | * The trailing underscore on this variable (and curr_ below) | ||
| 166 | * reminds us that the high bits are garbage; we mask them out | ||
| 167 | * when we compare them. | ||
| 168 | */ | ||
| 169 | u32 my_ticket_; | ||
| 170 | |||
| 171 | /* Take out the next ticket; this will also stop would-be readers. */ | ||
| 172 | if (val & 1) | ||
| 173 | val = get_rwlock(rwlock); | ||
| 174 | rwlock->lock = __insn_addb(val, 1 << WR_NEXT_SHIFT); | ||
| 175 | |||
| 176 | /* Extract my ticket value from the original word. */ | ||
| 177 | my_ticket_ = val >> WR_NEXT_SHIFT; | ||
| 178 | |||
| 179 | /* | ||
| 180 | * Wait until the "current" field matches our ticket, and | ||
| 181 | * there are no remaining readers. | ||
| 182 | */ | ||
| 183 | for (;;) { | ||
| 184 | u32 curr_ = val >> WR_CURR_SHIFT; | ||
| 185 | u32 readers = val >> RD_COUNT_SHIFT; | ||
| 186 | u32 delta = ((my_ticket_ - curr_) & WR_MASK) + !!readers; | ||
| 187 | if (likely(delta == 0)) | ||
| 188 | break; | ||
| 189 | |||
| 190 | /* Delay based on how many lock-holders are still out there. */ | ||
| 191 | relax((256 / CYCLES_PER_RELAX_LOOP) * delta); | ||
| 192 | |||
| 193 | /* | ||
| 194 | * Get a non-tns value to check; we don't need to tns | ||
| 195 | * it ourselves. Since we're not tns'ing, we retry | ||
| 196 | * more rapidly to get a valid value. | ||
| 197 | */ | ||
| 198 | while ((val = rwlock->lock) & 1) | ||
| 199 | relax(4); | ||
| 200 | } | ||
| 201 | } | ||
| 202 | EXPORT_SYMBOL(arch_write_lock_slow); | ||
| 203 | |||
| 204 | int __tns_atomic_acquire(atomic_t *lock) | ||
| 205 | { | ||
| 206 | int ret; | ||
| 207 | u32 iterations = 0; | ||
| 208 | |||
| 209 | BUG_ON(__insn_mfspr(SPR_INTERRUPT_CRITICAL_SECTION)); | ||
| 210 | __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1); | ||
| 211 | |||
| 212 | while ((ret = __insn_tns((void *)&lock->counter)) == 1) | ||
| 213 | delay_backoff(iterations++); | ||
| 214 | return ret; | ||
| 215 | } | ||
| 216 | |||
| 217 | void __tns_atomic_release(atomic_t *p, int v) | ||
| 218 | { | ||
| 219 | p->counter = v; | ||
| 220 | __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); | ||
| 221 | } | ||
