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/spinlock.h> 66 #include <sys/spinlock2.h> 67 #include <sys/ktr.h> 68 69 #ifdef _KERNEL_VIRTUAL 70 #include <pthread.h> 71 #endif 72 73 struct spinlock pmap_spin = SPINLOCK_INITIALIZER(pmap_spin, "pmap_spin"); 74 75 struct indefinite_info { 76 sysclock_t base; 77 int secs; 78 const char *ident; 79 }; 80 81 /* 82 * Kernal Trace 83 */ 84 #if !defined(KTR_SPIN_CONTENTION) 85 #define KTR_SPIN_CONTENTION KTR_ALL 86 #endif 87 #define SPIN_STRING "spin=%p type=%c" 88 #define SPIN_ARG_SIZE (sizeof(void *) + sizeof(int)) 89 90 KTR_INFO_MASTER(spin); 91 #if 0 92 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE); 93 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE); 94 #endif 95 96 #define logspin(name, spin, type) \ 97 KTR_LOG(spin_ ## name, spin, type) 98 99 #ifdef INVARIANTS 100 static int spin_lock_test_mode; 101 #endif 102 103 #ifdef DEBUG_LOCKS_LATENCY 104 105 static long spinlocks_add_latency; 106 SYSCTL_LONG(_debug, OID_AUTO, spinlocks_add_latency, CTLFLAG_RW, 107 &spinlocks_add_latency, 0, 108 "Add spinlock latency"); 109 110 #endif 111 112 static int spin_indefinite_check(struct spinlock *spin, 113 struct indefinite_info *info); 114 115 /* 116 * We contested due to another exclusive lock holder. We lose. 117 * 118 * We have to unwind the attempt and may acquire the spinlock 119 * anyway while doing so. 120 */ 121 int 122 spin_trylock_contested(struct spinlock *spin) 123 { 124 globaldata_t gd = mycpu; 125 126 /* 127 * Handle degenerate case, else fail. 128 */ 129 if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED|0, 1)) 130 return TRUE; 131 /*atomic_add_int(&spin->counta, -1);*/ 132 --gd->gd_spinlocks; 133 --gd->gd_curthread->td_critcount; 134 return (FALSE); 135 } 136 137 /* 138 * The spin_lock() inline was unable to acquire the lock and calls this 139 * function with spin->counta already incremented, passing (spin->counta - 1) 140 * to the function (the result of the inline's fetchadd). 141 * 142 * atomic_swap_int() is the absolute fastest spinlock instruction, at 143 * least on multi-socket systems. All instructions seem to be about 144 * the same on single-socket multi-core systems. However, atomic_swap_int() 145 * does not result in an even distribution of successful acquisitions. 146 * 147 * UNFORTUNATELY we cannot really use atomic_swap_int() when also implementing 148 * shared spin locks, so as we do a better job removing contention we've 149 * moved to atomic_cmpset_int() to be able handle multiple states. 150 * 151 * Another problem we have is that (at least on the 48-core opteron we test 152 * with) having all 48 cores contesting the same spin lock reduces 153 * performance to around 600,000 ops/sec, verses millions when fewer cores 154 * are going after the same lock. 155 * 156 * Backoff algorithms can create even worse starvation problems, and don't 157 * really improve performance when a lot of cores are contending. 158 * 159 * Our solution is to allow the data cache to lazy-update by reading it 160 * non-atomically and only attempting to acquire the lock if the lazy read 161 * looks good. This effectively limits cache bus bandwidth. A cpu_pause() 162 * (for intel/amd anyhow) is not strictly needed as cache bus resource use 163 * is governed by the lazy update. 164 * 165 * WARNING!!!! Performance matters here, by a huge margin. 166 * 167 * 48-core test with pre-read / -j 48 no-modules kernel compile 168 * with fanned-out inactive and active queues came in at 55 seconds. 169 * 170 * 48-core test with pre-read / -j 48 no-modules kernel compile 171 * came in at 75 seconds. Without pre-read it came in at 170 seconds. 172 * 173 * 4-core test with pre-read / -j 48 no-modules kernel compile 174 * came in at 83 seconds. Without pre-read it came in at 83 seconds 175 * as well (no difference). 176 */ 177 void 178 _spin_lock_contested(struct spinlock *spin, const char *ident, int value) 179 { 180 struct indefinite_info info = { 0, 0, ident }; 181 int i; 182 183 /* 184 * WARNING! Caller has already incremented the lock. We must 185 * increment the count value (from the inline's fetch-add) 186 * to match. 187 * 188 * Handle the degenerate case where the spinlock is flagged SHARED 189 * with only our reference. We can convert it to EXCLUSIVE. 190 */ 191 ++value; 192 if (value == (SPINLOCK_SHARED | 1)) { 193 if (atomic_cmpset_int(&spin->counta, SPINLOCK_SHARED | 1, 1)) 194 return; 195 } 196 197 /* 198 * Transfer our exclusive request to the high bits and clear the 199 * SPINLOCK_SHARED bit if it was set. This makes the spinlock 200 * appear exclusive, preventing any NEW shared or exclusive 201 * spinlocks from being obtained while we wait for existing 202 * shared or exclusive holders to unlock. 203 * 204 * Don't tread on earlier exclusive waiters by stealing the lock 205 * away early if the low bits happen to now be 1. 206 * 207 * The shared unlock understands that this may occur. 208 */ 209 atomic_add_int(&spin->counta, SPINLOCK_EXCLWAIT - 1); 210 if (value & SPINLOCK_SHARED) 211 atomic_clear_int(&spin->counta, SPINLOCK_SHARED); 212 213 #ifdef DEBUG_LOCKS_LATENCY 214 long j; 215 for (j = spinlocks_add_latency; j > 0; --j) 216 cpu_ccfence(); 217 #endif 218 /* 219 * Spin until we can acquire a low-count of 1. 220 */ 221 i = 0; 222 /*logspin(beg, spin, 'w');*/ 223 for (;;) { 224 /* 225 * If the low bits are zero, try to acquire the exclusive lock 226 * by transfering our high bit reservation to the low bits. 227 * 228 * NOTE: Reading spin->counta prior to the swap is extremely 229 * important on multi-chip/many-core boxes. On 48-core 230 * this one change improves fully concurrent all-cores 231 * compiles by 100% or better. 232 * 233 * I can't emphasize enough how important the pre-read 234 * is in preventing hw cache bus armageddon on 235 * multi-chip systems. And on single-chip/multi-core 236 * systems it just doesn't hurt. 237 */ 238 uint32_t ovalue = spin->counta; 239 cpu_ccfence(); 240 if ((ovalue & (SPINLOCK_EXCLWAIT - 1)) == 0 && 241 atomic_cmpset_int(&spin->counta, ovalue, 242 (ovalue - SPINLOCK_EXCLWAIT) | 1)) { 243 break; 244 } 245 if ((++i & 0x7F) == 0x7F) { 246 mycpu->gd_cnt.v_lock_name[0] = 'X'; 247 strncpy(mycpu->gd_cnt.v_lock_name + 1, 248 ident, 249 sizeof(mycpu->gd_cnt.v_lock_name) - 2); 250 ++mycpu->gd_cnt.v_lock_colls; 251 if (spin_indefinite_check(spin, &info)) 252 break; 253 } 254 #ifdef _KERNEL_VIRTUAL 255 pthread_yield(); 256 #endif 257 } 258 /*logspin(end, spin, 'w');*/ 259 } 260 261 /* 262 * The spin_lock_shared() inline was unable to acquire the lock and calls 263 * this function with spin->counta already incremented. 264 * 265 * This is not in the critical path unless there is contention between 266 * shared and exclusive holders. 267 */ 268 void 269 _spin_lock_shared_contested(struct spinlock *spin, const char *ident) 270 { 271 struct indefinite_info info = { 0, 0, ident }; 272 int i; 273 274 /* 275 * Undo the inline's increment. 276 */ 277 atomic_add_int(&spin->counta, -1); 278 279 #ifdef DEBUG_LOCKS_LATENCY 280 long j; 281 for (j = spinlocks_add_latency; j > 0; --j) 282 cpu_ccfence(); 283 #endif 284 285 /*logspin(beg, spin, 'w');*/ 286 i = 0; 287 for (;;) { 288 /* 289 * Loop until we can acquire the shared spinlock. Note that 290 * the low bits can be zero while the high EXCLWAIT bits are 291 * non-zero. In this situation exclusive requesters have 292 * priority (otherwise shared users on multiple cpus can hog 293 * the spinlnock). 294 * 295 * NOTE: Reading spin->counta prior to the swap is extremely 296 * important on multi-chip/many-core boxes. On 48-core 297 * this one change improves fully concurrent all-cores 298 * compiles by 100% or better. 299 * 300 * I can't emphasize enough how important the pre-read 301 * is in preventing hw cache bus armageddon on 302 * multi-chip systems. And on single-chip/multi-core 303 * systems it just doesn't hurt. 304 */ 305 uint32_t ovalue = spin->counta; 306 307 cpu_ccfence(); 308 if (ovalue == 0) { 309 if (atomic_cmpset_int(&spin->counta, 0, 310 SPINLOCK_SHARED | 1)) 311 break; 312 } else if (ovalue & SPINLOCK_SHARED) { 313 if (atomic_cmpset_int(&spin->counta, ovalue, 314 ovalue + 1)) 315 break; 316 } 317 if ((++i & 0x7F) == 0x7F) { 318 mycpu->gd_cnt.v_lock_name[0] = 'S'; 319 strncpy(mycpu->gd_cnt.v_lock_name + 1, 320 ident, 321 sizeof(mycpu->gd_cnt.v_lock_name) - 2); 322 ++mycpu->gd_cnt.v_lock_colls; 323 if (spin_indefinite_check(spin, &info)) 324 break; 325 } 326 #ifdef _KERNEL_VIRTUAL 327 pthread_yield(); 328 #endif 329 } 330 /*logspin(end, spin, 'w');*/ 331 } 332 333 static 334 int 335 spin_indefinite_check(struct spinlock *spin, struct indefinite_info *info) 336 { 337 sysclock_t count; 338 339 cpu_spinlock_contested(); 340 341 count = sys_cputimer->count(); 342 if (info->secs == 0) { 343 info->base = count; 344 ++info->secs; 345 } else if (count - info->base > sys_cputimer->freq) { 346 kprintf("spin_lock: %s(%p), indefinite wait (%d secs)!\n", 347 info->ident, spin, info->secs); 348 info->base = count; 349 ++info->secs; 350 if (panicstr) 351 return (TRUE); 352 #if defined(INVARIANTS) 353 if (spin_lock_test_mode) { 354 print_backtrace(-1); 355 return (TRUE); 356 } 357 #endif 358 #if defined(INVARIANTS) 359 if (info->secs == 11) 360 print_backtrace(-1); 361 #endif 362 if (info->secs == 60) 363 panic("spin_lock: %s(%p), indefinite wait!", 364 info->ident, spin); 365 } 366 return (FALSE); 367 } 368 369 /* 370 * If INVARIANTS is enabled various spinlock timing tests can be run 371 * by setting debug.spin_lock_test: 372 * 373 * 1 Test the indefinite wait code 374 * 2 Time the best-case exclusive lock overhead (spin_test_count) 375 * 3 Time the best-case shared lock overhead (spin_test_count) 376 */ 377 378 #ifdef INVARIANTS 379 380 static int spin_test_count = 10000000; 381 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, 382 "Number of iterations to use for spinlock wait code test"); 383 384 static int 385 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS) 386 { 387 struct spinlock spin; 388 int error; 389 int value = 0; 390 int i; 391 392 if ((error = priv_check(curthread, PRIV_ROOT)) != 0) 393 return (error); 394 if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0) 395 return (error); 396 397 /* 398 * Indefinite wait test 399 */ 400 if (value == 1) { 401 spin_init(&spin, "sysctllock"); 402 spin_lock(&spin); /* force an indefinite wait */ 403 spin_lock_test_mode = 1; 404 spin_lock(&spin); 405 spin_unlock(&spin); /* Clean up the spinlock count */ 406 spin_unlock(&spin); 407 spin_lock_test_mode = 0; 408 } 409 410 /* 411 * Time best-case exclusive spinlocks 412 */ 413 if (value == 2) { 414 globaldata_t gd = mycpu; 415 416 spin_init(&spin, "sysctllocktest"); 417 for (i = spin_test_count; i > 0; --i) { 418 _spin_lock_quick(gd, &spin, "test"); 419 spin_unlock_quick(gd, &spin); 420 } 421 } 422 423 return (0); 424 } 425 426 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT, 427 0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code"); 428 429 #endif /* INVARIANTS */ 430