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 33 /* 34 * The implementation is designed to avoid looping when compatible operations 35 * are executed. 36 * 37 * To acquire a spinlock we first increment counta. Then we check if counta 38 * meets our requirements. For an exclusive spinlock it must be 1, of a 39 * shared spinlock it must either be 1 or the SHARED_SPINLOCK bit must be set. 40 * 41 * Shared spinlock failure case: Decrement the count, loop until we can 42 * transition from 0 to SHARED_SPINLOCK|1, or until we find SHARED_SPINLOCK 43 * is set and increment the count. 44 * 45 * Exclusive spinlock failure case: While maintaining the count, clear the 46 * SHARED_SPINLOCK flag unconditionally. Then use an atomic add to transfer 47 * the count from the low bits to the high bits of counta. Then loop until 48 * all low bits are 0. Once the low bits drop to 0 we can transfer the 49 * count back with an atomic_cmpset_int(), atomically, and return. 50 */ 51 #include <sys/param.h> 52 #include <sys/systm.h> 53 #include <sys/types.h> 54 #include <sys/kernel.h> 55 #include <sys/sysctl.h> 56 #ifdef INVARIANTS 57 #include <sys/proc.h> 58 #endif 59 #include <sys/priv.h> 60 #include <machine/atomic.h> 61 #include <machine/cpu.h> 62 #include <machine/cpufunc.h> 63 #include <machine/specialreg.h> 64 #include <machine/clock.h> 65 #include <sys/indefinite2.h> 66 #include <sys/spinlock.h> 67 #include <sys/spinlock2.h> 68 #include <sys/ktr.h> 69 70 #ifdef _KERNEL_VIRTUAL 71 #include <pthread.h> 72 #endif 73 74 struct spinlock pmap_spin = SPINLOCK_INITIALIZER(pmap_spin, "pmap_spin"); 75 76 /* 77 * Kernal Trace 78 */ 79 #if !defined(KTR_SPIN_CONTENTION) 80 #define KTR_SPIN_CONTENTION KTR_ALL 81 #endif 82 #define SPIN_STRING "spin=%p type=%c" 83 #define SPIN_ARG_SIZE (sizeof(void *) + sizeof(int)) 84 85 KTR_INFO_MASTER(spin); 86 #if 0 87 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE); 88 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE); 89 #endif 90 91 #define logspin(name, spin, type) \ 92 KTR_LOG(spin_ ## name, spin, type) 93 94 #ifdef INVARIANTS 95 static int spin_lock_test_mode; 96 #endif 97 98 #ifdef DEBUG_LOCKS_LATENCY 99 100 static long spinlocks_add_latency; 101 SYSCTL_LONG(_debug, OID_AUTO, spinlocks_add_latency, CTLFLAG_RW, 102 &spinlocks_add_latency, 0, 103 "Add spinlock latency"); 104 105 #endif 106 107 /* 108 * We contested due to another exclusive lock holder. We lose. 109 * 110 * We have to unwind the attempt and may acquire the spinlock 111 * anyway while doing so. 112 */ 113 int 114 spin_trylock_contested(struct spinlock *spin) 115 { 116 globaldata_t gd = mycpu; 117 118 /* 119 * Handle degenerate case, else fail. 120 */ 121 if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED|0, 1)) 122 return TRUE; 123 /*atomic_add_int(&spin->counta, -1);*/ 124 --gd->gd_spinlocks; 125 crit_exit_raw(gd->gd_curthread); 126 127 return (FALSE); 128 } 129 130 /* 131 * The spin_lock() inline was unable to acquire the lock and calls this 132 * function with spin->counta already incremented, passing (spin->counta - 1) 133 * to the function (the result of the inline's fetchadd). 134 * 135 * Note that we implement both exclusive and shared spinlocks, so we cannot 136 * use atomic_swap_int(). Instead, we try to use atomic_fetchadd_int() 137 * to put most of the burden on the cpu. Atomic_cmpset_int() (cmpxchg) 138 * can cause a lot of unnecessary looping in situations where it is just 139 * trying to increment the count. 140 * 141 * Similarly, we leave the SHARED flag intact and incur slightly more 142 * overhead when switching from shared to exclusive. This allows us to 143 * use atomic_fetchadd_int() for both spinlock types in the critical 144 * path. 145 * 146 * Backoff algorithms can create even worse starvation problems, particularly 147 * on multi-socket cpus, and don't really improve performance when a lot 148 * of cores are contending. However, if we are contested on an exclusive 149 * lock due to a large number of shared locks being present, we throw in 150 * extra cpu_pause()'s to account for the necessary time it will take other 151 * cores to contend among themselves and release their shared locks. 152 */ 153 void 154 _spin_lock_contested(struct spinlock *spin, const char *ident, int value) 155 { 156 indefinite_info_t info; 157 158 /* 159 * WARNING! Caller has already incremented the lock. We must 160 * increment the count value (from the inline's fetch-add) 161 * to match. 162 * 163 * Handle the degenerate case where the spinlock is flagged SHARED 164 * with only our reference. We can convert it to EXCLUSIVE. 165 */ 166 ++value; 167 if (value == (SPINLOCK_SHARED | 1)) { 168 if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED | 1, 1)) 169 return; 170 } 171 indefinite_init(&info, ident, 0, 'S'); 172 173 /* 174 * Transfer our exclusive request to the high bits and clear the 175 * SPINLOCK_SHARED bit if it was set. This makes the spinlock 176 * appear exclusive, preventing any NEW shared or exclusive 177 * spinlocks from being obtained while we wait for existing 178 * shared or exclusive holders to unlock. 179 * 180 * Don't tread on earlier exclusive waiters by stealing the lock 181 * away early if the low bits happen to now be 1. 182 * 183 * The shared unlock understands that this may occur. 184 */ 185 atomic_add_int(&spin->counta, SPINLOCK_EXCLWAIT - 1); 186 if (value & SPINLOCK_SHARED) 187 atomic_clear_int(&spin->counta, SPINLOCK_SHARED); 188 189 /* 190 * Spin until we can acquire a low-count of 1. 191 */ 192 for (;;) { 193 /* 194 * If the low bits are zero, try to acquire the exclusive lock 195 * by transfering our high bit reservation to the low bits. 196 * 197 * NOTE: Reading spin->counta prior to the swap is extremely 198 * important on multi-chip/many-core boxes. On 48-core 199 * this one change improves fully concurrent all-cores 200 * compiles by 100% or better. 201 * 202 * I can't emphasize enough how important the pre-read 203 * is in preventing hw cache bus armageddon on 204 * multi-chip systems. And on single-chip/multi-core 205 * systems it just doesn't hurt. 206 */ 207 uint32_t ovalue = spin->counta; 208 cpu_ccfence(); 209 if ((ovalue & (SPINLOCK_EXCLWAIT - 1)) == 0 && 210 atomic_cmpset_int(&spin->counta, ovalue, 211 (ovalue - SPINLOCK_EXCLWAIT) | 1)) { 212 break; 213 } 214 215 /* 216 * Throw in extra cpu_pause()'s when we are waiting on 217 * multiple other shared lock holders to release (the 218 * indefinite_check() also throws one in). 219 * 220 * We know these are shared lock holders when the count 221 * is larger than 1, because an exclusive lock holder can 222 * only have one count. Do this optimization only when 223 * the number of shared lock holders is 3 or greater. 224 */ 225 ovalue &= SPINLOCK_EXCLWAIT - 1; 226 while (ovalue > 2) { 227 cpu_pause(); 228 cpu_pause(); 229 --ovalue; 230 } 231 232 if (indefinite_check(&info)) 233 break; 234 } 235 indefinite_done(&info); 236 } 237 238 /* 239 * The spin_lock_shared() inline was unable to acquire the lock and calls 240 * this function with spin->counta already incremented. 241 * 242 * This is not in the critical path unless there is contention between 243 * shared and exclusive holders. 244 */ 245 void 246 _spin_lock_shared_contested(struct spinlock *spin, const char *ident) 247 { 248 indefinite_info_t info; 249 250 indefinite_init(&info, ident, 0, 's'); 251 252 /* 253 * Undo the inline's increment. 254 */ 255 atomic_add_int(&spin->counta, -1); 256 257 #ifdef DEBUG_LOCKS_LATENCY 258 long j; 259 for (j = spinlocks_add_latency; j > 0; --j) 260 cpu_ccfence(); 261 #endif 262 263 for (;;) { 264 /* 265 * Loop until we can acquire the shared spinlock. Note that 266 * the low bits can be zero while the high EXCLWAIT bits are 267 * non-zero. In this situation exclusive requesters have 268 * priority (otherwise shared users on multiple cpus can hog 269 * the spinlnock). 270 * 271 * NOTE: Reading spin->counta prior to the swap is extremely 272 * important on multi-chip/many-core boxes. On 48-core 273 * this one change improves fully concurrent all-cores 274 * compiles by 100% or better. 275 * 276 * I can't emphasize enough how important the pre-read 277 * is in preventing hw cache bus armageddon on 278 * multi-chip systems. And on single-chip/multi-core 279 * systems it just doesn't hurt. 280 */ 281 uint32_t ovalue = spin->counta; 282 283 cpu_ccfence(); 284 if (ovalue == 0) { 285 if (atomic_cmpset_int(&spin->counta, 0, 286 SPINLOCK_SHARED | 1)) 287 break; 288 } else if (ovalue & SPINLOCK_SHARED) { 289 if (atomic_cmpset_int(&spin->counta, ovalue, 290 ovalue + 1)) 291 break; 292 } 293 if (indefinite_check(&info)) 294 break; 295 } 296 indefinite_done(&info); 297 } 298 299 /* 300 * If INVARIANTS is enabled various spinlock timing tests can be run 301 * by setting debug.spin_lock_test: 302 * 303 * 1 Test the indefinite wait code 304 * 2 Time the best-case exclusive lock overhead (spin_test_count) 305 * 3 Time the best-case shared lock overhead (spin_test_count) 306 */ 307 308 #ifdef INVARIANTS 309 310 static int spin_test_count = 10000000; 311 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, 312 "Number of iterations to use for spinlock wait code test"); 313 314 static int 315 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS) 316 { 317 struct spinlock spin; 318 int error; 319 int value = 0; 320 int i; 321 322 if ((error = priv_check(curthread, PRIV_ROOT)) != 0) 323 return (error); 324 if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0) 325 return (error); 326 327 /* 328 * Indefinite wait test 329 */ 330 if (value == 1) { 331 spin_init(&spin, "sysctllock"); 332 spin_lock(&spin); /* force an indefinite wait */ 333 spin_lock_test_mode = 1; 334 spin_lock(&spin); 335 spin_unlock(&spin); /* Clean up the spinlock count */ 336 spin_unlock(&spin); 337 spin_lock_test_mode = 0; 338 } 339 340 /* 341 * Time best-case exclusive spinlocks 342 */ 343 if (value == 2) { 344 globaldata_t gd = mycpu; 345 346 spin_init(&spin, "sysctllocktest"); 347 for (i = spin_test_count; i > 0; --i) { 348 _spin_lock_quick(gd, &spin, "test"); 349 spin_unlock_quick(gd, &spin); 350 } 351 } 352 353 return (0); 354 } 355 356 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT, 357 0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code"); 358 359 #endif /* INVARIANTS */ 360