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 111 112 113 /* 114 * We need a fairly large pool to avoid contention on large SMP systems, 115 * particularly multi-chip systems. 116 */ 117 /*#define SPINLOCK_NUM_POOL 8101*/ 118 #define SPINLOCK_NUM_POOL 8192 119 #define SPINLOCK_NUM_POOL_MASK (SPINLOCK_NUM_POOL - 1) 120 121 static __cachealign struct { 122 struct spinlock spin; 123 char filler[32 - sizeof(struct spinlock)]; 124 } pool_spinlocks[SPINLOCK_NUM_POOL]; 125 126 static int spin_indefinite_check(struct spinlock *spin, 127 struct indefinite_info *info); 128 129 /* 130 * We contested due to another exclusive lock holder. We lose. 131 * 132 * We have to unwind the attempt and may acquire the spinlock 133 * anyway while doing so. countb was incremented on our behalf. 134 */ 135 int 136 spin_trylock_contested(struct spinlock *spin) 137 { 138 globaldata_t gd = mycpu; 139 140 /*atomic_add_int(&spin->counta, -1);*/ 141 --gd->gd_spinlocks; 142 --gd->gd_curthread->td_critcount; 143 return (FALSE); 144 } 145 146 /* 147 * The spin_lock() inline was unable to acquire the lock. 148 * 149 * atomic_swap_int() is the absolute fastest spinlock instruction, at 150 * least on multi-socket systems. All instructions seem to be about 151 * the same on single-socket multi-core systems. However, atomic_swap_int() 152 * does not result in an even distribution of successful acquisitions. 153 * 154 * UNFORTUNATELY we cannot really use atomic_swap_int() when also implementing 155 * shared spin locks, so as we do a better job removing contention we've 156 * moved to atomic_cmpset_int() to be able handle multiple states. 157 * 158 * Another problem we have is that (at least on the 48-core opteron we test 159 * with) having all 48 cores contesting the same spin lock reduces 160 * performance to around 600,000 ops/sec, verses millions when fewer cores 161 * are going after the same lock. 162 * 163 * Backoff algorithms can create even worse starvation problems, and don't 164 * really improve performance when a lot of cores are contending. 165 * 166 * Our solution is to allow the data cache to lazy-update by reading it 167 * non-atomically and only attempting to acquire the lock if the lazy read 168 * looks good. This effectively limits cache bus bandwidth. A cpu_pause() 169 * (for intel/amd anyhow) is not strictly needed as cache bus resource use 170 * is governed by the lazy update. 171 * 172 * WARNING!!!! Performance matters here, by a huge margin. 173 * 174 * 48-core test with pre-read / -j 48 no-modules kernel compile 175 * with fanned-out inactive and active queues came in at 55 seconds. 176 * 177 * 48-core test with pre-read / -j 48 no-modules kernel compile 178 * came in at 75 seconds. Without pre-read it came in at 170 seconds. 179 * 180 * 4-core test with pre-read / -j 48 no-modules kernel compile 181 * came in at 83 seconds. Without pre-read it came in at 83 seconds 182 * as well (no difference). 183 */ 184 void 185 _spin_lock_contested(struct spinlock *spin, const char *ident) 186 { 187 struct indefinite_info info = { 0, 0, ident }; 188 int i; 189 190 /* 191 * Transfer our count to the high bits, then loop until we can 192 * acquire the low counter (== 1). No new shared lock can be 193 * acquired while we hold the EXCLWAIT bits. 194 * 195 * Force any existing shared locks to exclusive. The shared unlock 196 * understands that this may occur. 197 */ 198 atomic_add_int(&spin->counta, SPINLOCK_EXCLWAIT - 1); 199 atomic_clear_int(&spin->counta, SPINLOCK_SHARED); 200 201 #ifdef DEBUG_LOCKS_LATENCY 202 long j; 203 for (j = spinlocks_add_latency; j > 0; --j) 204 cpu_ccfence(); 205 #endif 206 #if defined(INVARIANTS) 207 if (spin_lock_test_mode > 10 && 208 spin->countb > spin_lock_test_mode && 209 (spin_lock_test_mode & 0xFF) == mycpu->gd_cpuid) { 210 spin->countb = 0; 211 print_backtrace(-1); 212 } 213 ++spin->countb; 214 #endif 215 i = 0; 216 217 /*logspin(beg, spin, 'w');*/ 218 for (;;) { 219 /* 220 * If the low bits are zero, try to acquire the exclusive lock 221 * by transfering our high bit counter to the low bits. 222 * 223 * NOTE: Reading spin->counta prior to the swap is extremely 224 * important on multi-chip/many-core boxes. On 48-core 225 * this one change improves fully concurrent all-cores 226 * compiles by 100% or better. 227 * 228 * I can't emphasize enough how important the pre-read 229 * is in preventing hw cache bus armageddon on 230 * multi-chip systems. And on single-chip/multi-core 231 * systems it just doesn't hurt. 232 */ 233 uint32_t ovalue = spin->counta; 234 cpu_ccfence(); 235 if ((ovalue & (SPINLOCK_EXCLWAIT - 1)) == 0 && 236 atomic_cmpset_int(&spin->counta, ovalue, 237 (ovalue - SPINLOCK_EXCLWAIT) | 1)) { 238 break; 239 } 240 if ((++i & 0x7F) == 0x7F) { 241 mycpu->gd_cnt.v_lock_name[0] = 'X'; 242 strncpy(mycpu->gd_cnt.v_lock_name + 1, 243 ident, 244 sizeof(mycpu->gd_cnt.v_lock_name) - 2); 245 ++mycpu->gd_cnt.v_lock_colls; 246 #if defined(INVARIANTS) 247 ++spin->countb; 248 #endif 249 if (spin_indefinite_check(spin, &info)) 250 break; 251 } 252 #ifdef _KERNEL_VIRTUAL 253 pthread_yield(); 254 #endif 255 } 256 /*logspin(end, spin, 'w');*/ 257 } 258 259 /* 260 * Shared spinlock attempt was contested. 261 * 262 * The caller has not modified counta. 263 */ 264 void 265 _spin_lock_shared_contested(struct spinlock *spin, const char *ident) 266 { 267 struct indefinite_info info = { 0, 0, ident }; 268 int i; 269 270 #ifdef DEBUG_LOCKS_LATENCY 271 long j; 272 for (j = spinlocks_add_latency; j > 0; --j) 273 cpu_ccfence(); 274 #endif 275 #if defined(INVARIANTS) 276 if (spin_lock_test_mode > 10 && 277 spin->countb > spin_lock_test_mode && 278 (spin_lock_test_mode & 0xFF) == mycpu->gd_cpuid) { 279 spin->countb = 0; 280 print_backtrace(-1); 281 } 282 ++spin->countb; 283 #endif 284 i = 0; 285 286 /*logspin(beg, spin, 'w');*/ 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 defined(INVARIANTS) 324 ++spin->countb; 325 #endif 326 if (spin_indefinite_check(spin, &info)) 327 break; 328 } 329 #ifdef _KERNEL_VIRTUAL 330 pthread_yield(); 331 #endif 332 } 333 /*logspin(end, spin, 'w');*/ 334 } 335 336 /* 337 * Pool functions (SHARED SPINLOCKS NOT SUPPORTED) 338 */ 339 static __inline int 340 _spin_pool_hash(void *ptr) 341 { 342 int i; 343 344 i = ((int)(uintptr_t) ptr >> 5) ^ ((int)(uintptr_t)ptr >> 12); 345 i &= SPINLOCK_NUM_POOL_MASK; 346 return (i); 347 } 348 349 void 350 _spin_pool_lock(void *chan, const char *ident) 351 { 352 struct spinlock *sp; 353 354 sp = &pool_spinlocks[_spin_pool_hash(chan)].spin; 355 _spin_lock(sp, ident); 356 } 357 358 void 359 _spin_pool_unlock(void *chan) 360 { 361 struct spinlock *sp; 362 363 sp = &pool_spinlocks[_spin_pool_hash(chan)].spin; 364 spin_unlock(sp); 365 } 366 367 368 static 369 int 370 spin_indefinite_check(struct spinlock *spin, struct indefinite_info *info) 371 { 372 sysclock_t count; 373 374 cpu_spinlock_contested(); 375 376 count = sys_cputimer->count(); 377 if (info->secs == 0) { 378 info->base = count; 379 ++info->secs; 380 } else if (count - info->base > sys_cputimer->freq) { 381 kprintf("spin_lock: %s(%p), indefinite wait (%d secs)!\n", 382 info->ident, spin, info->secs); 383 info->base = count; 384 ++info->secs; 385 if (panicstr) 386 return (TRUE); 387 #if defined(INVARIANTS) 388 if (spin_lock_test_mode) { 389 print_backtrace(-1); 390 return (TRUE); 391 } 392 #endif 393 #if defined(INVARIANTS) 394 if (info->secs == 11) 395 print_backtrace(-1); 396 #endif 397 if (info->secs == 60) 398 panic("spin_lock: %s(%p), indefinite wait!", 399 info->ident, spin); 400 } 401 return (FALSE); 402 } 403 404 /* 405 * If INVARIANTS is enabled various spinlock timing tests can be run 406 * by setting debug.spin_lock_test: 407 * 408 * 1 Test the indefinite wait code 409 * 2 Time the best-case exclusive lock overhead (spin_test_count) 410 * 3 Time the best-case shared lock overhead (spin_test_count) 411 */ 412 413 #ifdef INVARIANTS 414 415 static int spin_test_count = 10000000; 416 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, 417 "Number of iterations to use for spinlock wait code test"); 418 419 static int 420 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS) 421 { 422 struct spinlock spin; 423 int error; 424 int value = 0; 425 int i; 426 427 if ((error = priv_check(curthread, PRIV_ROOT)) != 0) 428 return (error); 429 if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0) 430 return (error); 431 432 /* 433 * Indefinite wait test 434 */ 435 if (value == 1) { 436 spin_init(&spin, "sysctllock"); 437 spin_lock(&spin); /* force an indefinite wait */ 438 spin_lock_test_mode = 1; 439 spin_lock(&spin); 440 spin_unlock(&spin); /* Clean up the spinlock count */ 441 spin_unlock(&spin); 442 spin_lock_test_mode = 0; 443 } 444 445 /* 446 * Time best-case exclusive spinlocks 447 */ 448 if (value == 2) { 449 globaldata_t gd = mycpu; 450 451 spin_init(&spin, "sysctllocktest"); 452 for (i = spin_test_count; i > 0; --i) { 453 _spin_lock_quick(gd, &spin, "test"); 454 spin_unlock_quick(gd, &spin); 455 } 456 } 457 458 return (0); 459 } 460 461 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT, 462 0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code"); 463 464 #endif /* INVARIANTS */ 465