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 1.16 2008/09/11 01:11:42 y0netan1 Exp $
  33  */
  34
  35 #include "opt_ddb.h"
  36
  37 #include <sys/param.h>
  38 #include <sys/systm.h>
  39 #include <sys/types.h>
  40 #include <sys/kernel.h>
  41 #include <sys/sysctl.h>
  42 #ifdef INVARIANTS
  43 #include <sys/proc.h>
  44 #endif
  45 #include <ddb/ddb.h>
  46 #include <machine/atomic.h>
  47 #include <machine/cpufunc.h>
  48 #include <machine/specialreg.h>
  49 #include <machine/clock.h>
  50 #include <sys/spinlock.h>
  51 #include <sys/spinlock2.h>
  52 #include <sys/ktr.h>
  53
  54 #define BACKOFF_INITIAL 1
  55 #define BACKOFF_LIMIT   256
  56
  57 #ifdef SMP
  58
  59 /*
  60  * Kernal Trace
  61  */
  62 #if !defined(KTR_SPIN_CONTENTION)
  63 #define KTR_SPIN_CONTENTION     KTR_ALL
  64 #endif
  65 #define SPIN_STRING     "spin=%p type=%c"
  66 #define SPIN_ARG_SIZE   (sizeof(void *) + sizeof(int))
  67
  68 KTR_INFO_MASTER(spin);
  69 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE);
  70 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE);
  71 KTR_INFO(KTR_SPIN_CONTENTION, spin, backoff, 2,
  72          "spin=%p bo1=%d thr=%p bo=%d",
  73          ((2 * sizeof(void *)) + (2 * sizeof(int))));
  74 KTR_INFO(KTR_SPIN_CONTENTION, spin, bofail, 3, SPIN_STRING, SPIN_ARG_SIZE);
  75
  76 #define logspin(name, mtx, type)                        \
  77         KTR_LOG(spin_ ## name, mtx, type)
  78
  79 #define logspin_backoff(mtx, bo1, thr, bo)              \
  80         KTR_LOG(spin_backoff, mtx, bo1, thr, bo)
  81
  82 #ifdef INVARIANTS
  83 static int spin_lock_test_mode;
  84 #endif
  85
  86 static int64_t spinlocks_contested1;
  87 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested1, CTLFLAG_RD,
  88             &spinlocks_contested1, 0, "");
  89
  90 static int64_t spinlocks_contested2;
  91 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested2, CTLFLAG_RD,
  92             &spinlocks_contested2, 0, "");
  93
  94 static int spinlocks_backoff_limit = BACKOFF_LIMIT;
  95 SYSCTL_INT(_debug, OID_AUTO, spinlocks_bolim, CTLFLAG_RW,
  96            &spinlocks_backoff_limit, 0, "");
  97
  98 struct exponential_backoff {
  99         int backoff;
 100         int nsec;
 101         struct spinlock *mtx;
 102         sysclock_t base;
 103 };
 104 static int exponential_backoff(struct exponential_backoff *bo);
 105
 106 static __inline
 107 void
 108 exponential_init(struct exponential_backoff *bo, struct spinlock *mtx)
 109 {
 110         bo->backoff = BACKOFF_INITIAL;
 111         bo->nsec = 0;
 112         bo->mtx = mtx;
 113 }
 114
 115 /*
 116  * We were either contested due to another exclusive lock holder,
 117  * or due to the presence of shared locks.  We have to undo the mess
 118  * we created by returning the shared locks.
 119  *
 120  * If there was another exclusive lock holder only the exclusive bit
 121  * in value will be the only bit set.  We don't have to do anything since
 122  * restoration does not involve any work.
 123  *
 124  * Otherwise we successfully obtained the exclusive bit.  Attempt to
 125  * clear the shared bits.  If we are able to clear the shared bits
 126  * we win.  Otherwise we lose and we have to restore the shared bits
 127  * we couldn't clear (and also clear our exclusive bit).
 128  */
 129 int
 130 spin_trylock_wr_contested(globaldata_t gd, struct spinlock *mtx, int value)
 131 {
 132         int bit;
 133
 134         ++spinlocks_contested1;
 135         if ((value & SPINLOCK_EXCLUSIVE) == 0) {
 136                 while (value) {
 137                         bit = bsfl(value);
 138                         if (globaldata_find(bit)->gd_spinlock_rd == mtx) {
 139                                 atomic_swap_int(&mtx->lock, value);
 140                                 --gd->gd_spinlocks_wr;
 141                                 return (FALSE);
 142                         }
 143                         value &= ~(1 << bit);
 144                 }
 145                 return (TRUE);
 146         }
 147         --gd->gd_spinlocks_wr;
 148         return (FALSE);
 149 }
 150
 151 /*
 152  * We were either contested due to another exclusive lock holder,
 153  * or due to the presence of shared locks
 154  *
 155  * NOTE: If value indicates an exclusively held mutex, no shared bits
 156  * would have been set and we can throw away value.
 157  */
 158 void
 159 spin_lock_wr_contested(struct spinlock *mtx, int value)
 160 {
 161         struct exponential_backoff backoff;
 162         globaldata_t gd = mycpu;
 163         int bit;
 164         int mask;
 165
 166         /*
 167          * Wait until we can gain exclusive access vs another exclusive
 168          * holder.
 169          */
 170         exponential_init(&backoff, mtx);
 171         ++spinlocks_contested1;
 172         logspin(beg, mtx, 'w');
 173
 174         while (value & SPINLOCK_EXCLUSIVE) {
 175                 value = atomic_swap_int(&mtx->lock, SPINLOCK_EXCLUSIVE);
 176                 if (exponential_backoff(&backoff)) {
 177                         value &= ~SPINLOCK_EXCLUSIVE;
 178                         break;
 179                 }
 180         }
 181
 182         /*
 183          * Kill the cached shared bit for our own cpu.  This is the most
 184          * common case and there's no sense wasting cpu on it.  Since
 185          * spinlocks aren't recursive, we can't own a shared ref on the
 186          * spinlock while trying to get an exclusive one.
 187          *
 188          * If multiple bits are set do not stall on any single cpu.  Check
 189          * all cpus that have the cache bit set, then loop and check again,
 190          * until we've cleaned all the bits.
 191          */
 192         value &= ~gd->gd_cpumask;
 193
 194         while ((mask = value) != 0) {
 195                 while (mask) {
 196                         bit = bsfl(value);
 197                         if (globaldata_find(bit)->gd_spinlock_rd != mtx) {
 198                                 value &= ~(1 << bit);
 199                         } else if (exponential_backoff(&backoff)) {
 200                                 value = 0;
 201                                 break;
 202                         }
 203                         mask &= ~(1 << bit);
 204                 }
 205         }
 206         logspin(end, mtx, 'w');
 207 }
 208
 209 /*
 210  * The cache bit wasn't set for our cpu.  Loop until we can set the bit.
 211  * As with the spin_lock_rd() inline we need a memory fence after setting
 212  * gd_spinlock_rd to interlock against exclusive spinlocks waiting for
 213  * that field to clear.
 214  */
 215 void
 216 spin_lock_rd_contested(struct spinlock *mtx)
 217 {
 218         struct exponential_backoff backoff;
 219         globaldata_t gd = mycpu;
 220         int value = mtx->lock;
 221
 222         /*
 223          * Shortcut the op if we can just set the cache bit.  This case
 224          * occurs when the last lock was an exclusive lock.
 225          */
 226         while ((value & SPINLOCK_EXCLUSIVE) == 0) {
 227                 if (atomic_cmpset_int(&mtx->lock, value, value|gd->gd_cpumask))
 228                         return;
 229                 value = mtx->lock;
 230         }
 231
 232         exponential_init(&backoff, mtx);
 233         ++spinlocks_contested1;
 234
 235         logspin(beg, mtx, 'r');
 236
 237         while ((value & gd->gd_cpumask) == 0) {
 238                 if (value & SPINLOCK_EXCLUSIVE) {
 239                         gd->gd_spinlock_rd = NULL;
 240                         if (exponential_backoff(&backoff)) {
 241                                 gd->gd_spinlock_rd = mtx;
 242                                 break;
 243                         }
 244                         gd->gd_spinlock_rd = mtx;
 245                         cpu_mfence();
 246                 } else {
 247                         if (atomic_cmpset_int(&mtx->lock, value, value|gd->gd_cpumask))
 248                                 break;
 249                 }
 250                 value = mtx->lock;
 251         }
 252         logspin(end, mtx, 'r');
 253 }
 254
 255 /*
 256  * Handle exponential backoff and indefinite waits.
 257  *
 258  * If the system is handling a panic we hand the spinlock over to the caller
 259  * after 1 second.  After 10 seconds we attempt to print a debugger
 260  * backtrace.  We also run pending interrupts in order to allow a console
 261  * break into DDB.
 262  */
 263 static
 264 int
 265 exponential_backoff(struct exponential_backoff *bo)
 266 {
 267         sysclock_t count;
 268         int backoff;
 269
 270 #ifdef _RDTSC_SUPPORTED_
 271         if (cpu_feature & CPUID_TSC) {
 272                 backoff =
 273                 (((u_long)rdtsc() ^ (((u_long)curthread) >> 5)) &
 274                  (bo->backoff - 1)) + BACKOFF_INITIAL;
 275         } else
 276 #endif
 277                 backoff = bo->backoff;
 278         logspin_backoff(bo->mtx, bo->backoff, curthread, backoff);
 279
 280         /*
 281          * Quick backoff
 282          */
 283         for (; backoff; --backoff)
 284                 cpu_pause();
 285         if (bo->backoff < spinlocks_backoff_limit) {
 286                 bo->backoff <<= 1;
 287                 return (FALSE);
 288         } else {
 289                 bo->backoff = BACKOFF_INITIAL;
 290         }
 291
 292         logspin(bofail, bo->mtx, 'u');
 293
 294         /*
 295          * Indefinite
 296          */
 297         ++spinlocks_contested2;
 298         cpu_spinlock_contested();
 299         if (bo->nsec == 0) {
 300                 bo->base = sys_cputimer->count();
 301                 bo->nsec = 1;
 302         }
 303
 304         count = sys_cputimer->count();
 305         if (count - bo->base > sys_cputimer->freq) {
 306                 kprintf("spin_lock: %p, indefinite wait!\n", bo->mtx);
 307                 if (panicstr)
 308                         return (TRUE);
 309 #if defined(INVARIANTS) && defined(DDB)
 310                 if (spin_lock_test_mode) {
 311                         db_print_backtrace();
 312                         return (TRUE);
 313                 }
 314 #endif
 315                 ++bo->nsec;
 316 #if defined(INVARIANTS) && defined(DDB)
 317                 if (bo->nsec == 11)
 318                         db_print_backtrace();
 319 #endif
 320                 if (bo->nsec == 60)
 321                         panic("spin_lock: %p, indefinite wait!\n", bo->mtx);
 322                 splz();
 323                 bo->base = count;
 324         }
 325         return (FALSE);
 326 }
 327
 328 /*
 329  * If INVARIANTS is enabled various spinlock timing tests can be run
 330  * by setting debug.spin_lock_test:
 331  *
 332  *      1       Test the indefinite wait code
 333  *      2       Time the best-case exclusive lock overhead (spin_test_count)
 334  *      3       Time the best-case shared lock overhead (spin_test_count)
 335  */
 336
 337 #ifdef INVARIANTS
 338
 339 static int spin_test_count = 10000000;
 340 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, "");
 341
 342 static int
 343 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS)
 344 {
 345         struct spinlock mtx;
 346         int error;
 347         int value = 0;
 348         int i;
 349
 350         if ((error = suser(curthread)) != 0)
 351                 return (error);
 352         if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0)
 353                 return (error);
 354
 355         /*
 356          * Indefinite wait test
 357          */
 358         if (value == 1) {
 359                 spin_init(&mtx);
 360                 spin_lock_wr(&mtx);     /* force an indefinite wait */
 361                 spin_lock_test_mode = 1;
 362                 spin_lock_wr(&mtx);
 363                 spin_unlock_wr(&mtx);   /* Clean up the spinlock count */
 364                 spin_unlock_wr(&mtx);
 365                 spin_lock_test_mode = 0;
 366         }
 367
 368         /*
 369          * Time best-case exclusive spinlocks
 370          */
 371         if (value == 2) {
 372                 globaldata_t gd = mycpu;
 373
 374                 spin_init(&mtx);
 375                 for (i = spin_test_count; i > 0; --i) {
 376                     spin_lock_wr_quick(gd, &mtx);
 377                     spin_unlock_wr_quick(gd, &mtx);
 378                 }
 379         }
 380
 381         /*
 382          * Time best-case shared spinlocks
 383          */
 384         if (value == 3) {
 385                 globaldata_t gd = mycpu;
 386
 387                 spin_init(&mtx);
 388                 for (i = spin_test_count; i > 0; --i) {
 389                     spin_lock_rd_quick(gd, &mtx);
 390                     spin_unlock_rd_quick(gd, &mtx);
 391                 }
 392         }
 393         return (0);
 394 }
 395
 396 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT,
 397         0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code");
 398
 399 #endif  /* INVARIANTS */
 400 #endif  /* SMP */