arch/tile/lib/spinlock_32.c

   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         u32 iterations = 0;
 171
 172         /*
 173          * Wait until there are no readers, then bump up the next
 174          * field and capture the ticket value.
 175          */
 176         for (;;) {
 177                 if (!(val & 1)) {
 178                         if ((val >> RD_COUNT_SHIFT) == 0)
 179                                 break;
 180                         rwlock->lock = val;
 181                 }
 182                 delay_backoff(iterations++);
 183                 val = __insn_tns((int *)&rwlock->lock);
 184         }
 185
 186         /* Take out the next ticket and extract my ticket value. */
 187         rwlock->lock = __insn_addb(val, 1 << WR_NEXT_SHIFT);
 188         my_ticket_ = val >> WR_NEXT_SHIFT;
 189
 190         /* Wait until the "current" field matches our ticket. */
 191         for (;;) {
 192                 u32 curr_ = val >> WR_CURR_SHIFT;
 193                 u32 delta = ((my_ticket_ - curr_) & WR_MASK);
 194                 if (likely(delta == 0))
 195                         break;
 196
 197                 /* Delay based on how many lock-holders are still out there. */
 198                 relax((256 / CYCLES_PER_RELAX_LOOP) * delta);
 199
 200                 /*
 201                  * Get a non-tns value to check; we don't need to tns
 202                  * it ourselves.  Since we're not tns'ing, we retry
 203                  * more rapidly to get a valid value.
 204                  */
 205                 while ((val = rwlock->lock) & 1)
 206                         relax(4);
 207         }
 208 }
 209 EXPORT_SYMBOL(arch_write_lock_slow);
 210
 211 int __tns_atomic_acquire(atomic_t *lock)
 212 {
 213         int ret;
 214         u32 iterations = 0;
 215
 216         BUG_ON(__insn_mfspr(SPR_INTERRUPT_CRITICAL_SECTION));
 217         __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 1);
 218
 219         while ((ret = __insn_tns((void *)&lock->counter)) == 1)
 220                 delay_backoff(iterations++);
 221         return ret;
 222 }
 223
 224 void __tns_atomic_release(atomic_t *p, int v)
 225 {
 226         p->counter = v;
 227         __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0);
 228 }