sys/kern/kern_spinlock.c

   1 /*
   2  * Copyright (c) 2005 Jeffrey M. Hsu.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Jeffrey M. Hsu. and Matthew Dillon
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Neither the name of The DragonFly Project nor the names of its
  16  *    contributors may be used to endorse or promote products derived
  17  *    from this software without specific, prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  22  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  23  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  24  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  25  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  26  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  27  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  28  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  29  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  * $DragonFly: src/sys/kern/kern_spinlock.c,v 1.11 2007/07/02 16:51:58 dillon Exp $
  33  */
  34
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/types.h>
  38 #include <sys/kernel.h>
  39 #include <sys/sysctl.h>
  40 #ifdef INVARIANTS
  41 #include <sys/proc.h>
  42 #endif
  43 #include <ddb/ddb.h>
  44 #include <machine/atomic.h>
  45 #include <machine/cpufunc.h>
  46 #include <machine/clock.h>
  47 #include <sys/spinlock.h>
  48 #include <sys/spinlock2.h>
  49 #include <sys/ktr.h>
  50
  51 #define BACKOFF_INITIAL 1
  52 #define BACKOFF_LIMIT   256
  53
  54 #ifdef SMP
  55
  56 /*
  57  * Kernal Trace
  58  */
  59 #if !defined(KTR_SPIN_CONTENTION)
  60 #define KTR_SPIN_CONTENTION     KTR_ALL
  61 #endif
  62 #define SPIN_STRING     "spin=%p type=%c"
  63 #define SPIN_ARG_SIZE   (sizeof(void *) + sizeof(int))
  64
  65 KTR_INFO_MASTER(spin);
  66 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE);
  67 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE);
  68
  69 #define logspin(name, mtx, type)                        \
  70         KTR_LOG(spin_ ## name, mtx, type)
  71
  72 #ifdef INVARIANTS
  73 static int spin_lock_test_mode;
  74 #endif
  75
  76 static int64_t spinlocks_contested1;
  77 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested1, CTLFLAG_RD, &spinlocks_contested1, 0, "");
  78 static int64_t spinlocks_contested2;
  79 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested2, CTLFLAG_RD, &spinlocks_contested2, 0, "");
  80
  81 struct exponential_backoff {
  82         int backoff;
  83         int nsec;
  84         struct spinlock *mtx;
  85         sysclock_t base;
  86 };
  87 static int exponential_backoff(struct exponential_backoff *bo);
  88
  89 static __inline
  90 void
  91 exponential_init(struct exponential_backoff *bo, struct spinlock *mtx)
  92 {
  93         bo->backoff = BACKOFF_INITIAL;
  94         bo->nsec = 0;
  95         bo->mtx = mtx;
  96 }
  97
  98 /*
  99  * We were either contested due to another exclusive lock holder,
 100  * or due to the presence of shared locks.  We have to undo the mess
 101  * we created by returning the shared locks.
 102  *
 103  * If there was another exclusive lock holder only the exclusive bit
 104  * in value will be the only bit set.  We don't have to do anything since
 105  * restoration does not involve any work.
 106  *
 107  * Otherwise we successfully obtained the exclusive bit.  Attempt to
 108  * clear the shared bits.  If we are able to clear the shared bits
 109  * we win.  Otherwise we lose and we have to restore the shared bits
 110  * we couldn't clear (and also clear our exclusive bit).
 111  */
 112 int
 113 spin_trylock_wr_contested(struct spinlock *mtx, int value)
 114 {
 115         int bit;
 116
 117         ++spinlocks_contested1;
 118         if ((value & SPINLOCK_EXCLUSIVE) == 0) {
 119                 while (value) {
 120                         bit = bsfl(value);
 121                         if (globaldata_find(bit)->gd_spinlock_rd != mtx) {
 122                                 atomic_swap_int(&mtx->lock, value);
 123                                 return (FALSE);
 124                         }
 125                         value &= ~(1 << bit);
 126                 }
 127                 return (TRUE);
 128         }
 129         return (FALSE);
 130 }
 131
 132 /*
 133  * We were either contested due to another exclusive lock holder,
 134  * or due to the presence of shared locks
 135  *
 136  * NOTE: If value indicates an exclusively held mutex, no shared bits
 137  * would have been set and we can throw away value.
 138  */
 139 void
 140 spin_lock_wr_contested(struct spinlock *mtx, int value)
 141 {
 142         struct exponential_backoff backoff;
 143         globaldata_t gd = mycpu;
 144         int bit;
 145         int mask;
 146
 147         /*
 148          * Wait until we can gain exclusive access vs another exclusive
 149          * holder.
 150          */
 151         exponential_init(&backoff, mtx);
 152         ++spinlocks_contested1;
 153         logspin(beg, mtx, 'w');
 154
 155         while (value & SPINLOCK_EXCLUSIVE) {
 156                 value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE);
 157                 if (exponential_backoff(&backoff)) {
 158                         value &= ~SPINLOCK_EXCLUSIVE;
 159                         break;
 160                 }
 161         }
 162
 163         /*
 164          * Kill the cached shared bit for our own cpu.  This is the most
 165          * common case and there's no sense wasting cpu on it.  Since
 166          * spinlocks aren't recursive, we can't own a shared ref on the
 167          * spinlock while trying to get an exclusive one.
 168          *
 169          * If multiple bits are set do not stall on any single cpu.  Check
 170          * all cpus that have the cache bit set, then loop and check again,
 171          * until we've cleaned all the bits.
 172          */
 173         value &= ~gd->gd_cpumask;
 174
 175         while ((mask = value) != 0) {
 176                 while (mask) {
 177                         bit = bsfl(value);
 178                         if (globaldata_find(bit)->gd_spinlock_rd != mtx) {
 179                                 value &= ~(1 << bit);
 180                         } else if (exponential_backoff(&backoff)) {
 181                                 value = 0;
 182                                 break;
 183                         }
 184                         mask &= ~(1 << bit);
 185                 }
 186         }
 187         logspin(end, mtx, 'w');
 188 }
 189
 190 /*
 191  * The cache bit wasn't set for our cpu.  Loop until we can set the bit.
 192  * As with the spin_lock_rd() inline we need a memory fence after setting
 193  * gd_spinlock_rd to interlock against exclusive spinlocks waiting for
 194  * that field to clear.
 195  */
 196 void
 197 spin_lock_rd_contested(struct spinlock *mtx)
 198 {
 199         struct exponential_backoff backoff;
 200         globaldata_t gd = mycpu;
 201         int value = mtx->lock;
 202
 203         /*
 204          * Shortcut the op if we can just set the cache bit.  This case
 205          * occurs when the last lock was an exclusive lock.
 206          */
 207         while ((value & SPINLOCK_EXCLUSIVE) == 0) {
 208                 if (atomic_cmpset_int(&mtx->lock, value, value|gd->gd_cpumask))
 209                         return;
 210                 value = mtx->lock;
 211         }
 212
 213         exponential_init(&backoff, mtx);
 214         ++spinlocks_contested1;
 215
 216         logspin(beg, mtx, 'r');
 217
 218         while ((value & gd->gd_cpumask) == 0) {
 219                 if (value & SPINLOCK_EXCLUSIVE) {
 220                         gd->gd_spinlock_rd = NULL;
 221                         if (exponential_backoff(&backoff)) {
 222                                 gd->gd_spinlock_rd = mtx;
 223                                 break;
 224                         }
 225                         gd->gd_spinlock_rd = mtx;
 226                         cpu_mfence();
 227                 } else {
 228                         if (atomic_cmpset_int(&mtx->lock, value, value|gd->gd_cpumask))
 229                                 break;
 230                 }
 231                 value = mtx->lock;
 232         }
 233         logspin(end, mtx, 'r');
 234 }
 235
 236 /*
 237  * Handle exponential backoff and indefinite waits.
 238  *
 239  * If the system is handling a panic we hand the spinlock over to the caller
 240  * after 1 second.  After 10 seconds we attempt to print a debugger
 241  * backtrace.  We also run pending interrupts in order to allow a console
 242  * break into DDB.
 243  */
 244 static
 245 int
 246 exponential_backoff(struct exponential_backoff *bo)
 247 {
 248         sysclock_t count;
 249         int i;
 250
 251         /*
 252          * Quick backoff
 253          */
 254         for (i = 0; i < bo->backoff; ++i)
 255                 cpu_nop();
 256         if (bo->backoff < BACKOFF_LIMIT) {
 257                 bo->backoff <<= 1;
 258                 return (FALSE);
 259         }
 260
 261         /*
 262          * Indefinite
 263          */
 264         ++spinlocks_contested2;
 265         cpu_spinlock_contested();
 266         if (bo->nsec == 0) {
 267                 bo->base = sys_cputimer->count();
 268                 bo->nsec = 1;
 269         }
 270
 271         count = sys_cputimer->count();
 272         if (count - bo->base > sys_cputimer->freq) {
 273                 kprintf("spin_lock: %p, indefinite wait!\n", bo->mtx);
 274                 if (panicstr)
 275                         return (TRUE);
 276 #if defined(INVARIANTS) && defined(DDB)
 277                 if (spin_lock_test_mode) {
 278                         db_print_backtrace();
 279                         return (TRUE);
 280                 }
 281 #endif
 282                 ++bo->nsec;
 283 #if defined(INVARIANTS) && defined(DDB)
 284                 if (bo->nsec == 11)
 285                         db_print_backtrace();
 286 #endif
 287                 if (bo->nsec == 60)
 288                         panic("spin_lock: %p, indefinite wait!\n", bo->mtx);
 289                 splz();
 290                 bo->base = count;
 291         }
 292         return (FALSE);
 293 }
 294
 295 /*
 296  * If INVARIANTS is enabled various spinlock timing tests can be run
 297  * by setting debug.spin_lock_test:
 298  *
 299  *      1       Test the indefinite wait code
 300  *      2       Time the best-case exclusive lock overhead (spin_test_count)
 301  *      3       Time the best-case shared lock overhead (spin_test_count)
 302  */
 303
 304 #ifdef INVARIANTS
 305
 306 static int spin_test_count = 10000000;
 307 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, "");
 308
 309 static int
 310 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS)
 311 {
 312         struct spinlock mtx;
 313         int error;
 314         int value = 0;
 315         int i;
 316
 317         if ((error = suser(curthread)) != 0)
 318                 return (error);
 319         if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0)
 320                 return (error);
 321
 322         /*
 323          * Indefinite wait test
 324          */
 325         if (value == 1) {
 326                 spin_init(&mtx);
 327                 spin_lock_wr(&mtx);     /* force an indefinite wait */
 328                 spin_lock_test_mode = 1;
 329                 spin_lock_wr(&mtx);
 330                 spin_unlock_wr(&mtx);   /* Clean up the spinlock count */
 331                 spin_unlock_wr(&mtx);
 332                 spin_lock_test_mode = 0;
 333         }
 334
 335         /*
 336          * Time best-case exclusive spinlocks
 337          */
 338         if (value == 2) {
 339                 globaldata_t gd = mycpu;
 340
 341                 spin_init(&mtx);
 342                 for (i = spin_test_count; i > 0; --i) {
 343                     spin_lock_wr_quick(gd, &mtx);
 344                     spin_unlock_wr_quick(gd, &mtx);
 345                 }
 346         }
 347
 348         /*
 349          * Time best-case shared spinlocks
 350          */
 351         if (value == 3) {
 352                 globaldata_t gd = mycpu;
 353
 354                 spin_init(&mtx);
 355                 for (i = spin_test_count; i > 0; --i) {
 356                     spin_lock_rd_quick(gd, &mtx);
 357                     spin_unlock_rd_quick(gd, &mtx);
 358                 }
 359         }
 360         return (0);
 361 }
 362
 363 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT,
 364         0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code");
 365
 366 #endif  /* INVARIANTS */
 367 #endif  /* SMP */