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 * The spinlock code utilizes two counters to form a virtual FIFO, allowing 34 * a spinlock to allocate a slot and then only issue memory read operations 35 * until it is handed the lock (if it is not the next owner for the lock). 36 */ 37 38 #include <sys/param.h> 39 #include <sys/systm.h> 40 #include <sys/types.h> 41 #include <sys/kernel.h> 42 #include <sys/sysctl.h> 43 #ifdef INVARIANTS 44 #include <sys/proc.h> 45 #endif 46 #include <sys/priv.h> 47 #include <machine/atomic.h> 48 #include <machine/cpu.h> 49 #include <machine/cpufunc.h> 50 #include <machine/specialreg.h> 51 #include <machine/clock.h> 52 #include <sys/spinlock.h> 53 #include <sys/spinlock2.h> 54 #include <sys/ktr.h> 55 56 struct spinlock pmap_spin = SPINLOCK_INITIALIZER(pmap_spin); 57 58 #ifdef SMP 59 60 struct indefinite_info { 61 sysclock_t base; 62 int secs; 63 }; 64 65 /* 66 * Kernal Trace 67 */ 68 #if !defined(KTR_SPIN_CONTENTION) 69 #define KTR_SPIN_CONTENTION KTR_ALL 70 #endif 71 #define SPIN_STRING "spin=%p type=%c" 72 #define SPIN_ARG_SIZE (sizeof(void *) + sizeof(int)) 73 74 KTR_INFO_MASTER(spin); 75 #if 0 76 KTR_INFO(KTR_SPIN_CONTENTION, spin, beg, 0, SPIN_STRING, SPIN_ARG_SIZE); 77 KTR_INFO(KTR_SPIN_CONTENTION, spin, end, 1, SPIN_STRING, SPIN_ARG_SIZE); 78 #endif 79 80 #define logspin(name, spin, type) \ 81 KTR_LOG(spin_ ## name, spin, type) 82 83 #ifdef INVARIANTS 84 static int spin_lock_test_mode; 85 #endif 86 87 static int64_t spinlocks_contested1; 88 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested1, CTLFLAG_RD, 89 &spinlocks_contested1, 0, 90 "Spinlock contention count due to collisions with exclusive lock holders"); 91 92 static int64_t spinlocks_contested2; 93 SYSCTL_QUAD(_debug, OID_AUTO, spinlocks_contested2, CTLFLAG_RD, 94 &spinlocks_contested2, 0, 95 "Serious spinlock contention count"); 96 97 #ifdef DEBUG_LOCKS_LATENCY 98 99 static long spinlocks_add_latency; 100 SYSCTL_LONG(_debug, OID_AUTO, spinlocks_add_latency, CTLFLAG_RW, 101 &spinlocks_add_latency, 0, 102 "Add spinlock latency"); 103 104 #endif 105 106 107 /* 108 * We need a fairly large pool to avoid contention on large SMP systems, 109 * particularly multi-chip systems. 110 */ 111 /*#define SPINLOCK_NUM_POOL 8101*/ 112 #define SPINLOCK_NUM_POOL 8192 113 #define SPINLOCK_NUM_POOL_MASK (SPINLOCK_NUM_POOL - 1) 114 115 static __cachealign struct { 116 struct spinlock spin; 117 char filler[32 - sizeof(struct spinlock)]; 118 } pool_spinlocks[SPINLOCK_NUM_POOL]; 119 120 static int spin_indefinite_check(struct spinlock *spin, 121 struct indefinite_info *info); 122 123 /* 124 * We contested due to another exclusive lock holder. We lose. 125 * 126 * We have to unwind the attempt and may acquire the spinlock 127 * anyway while doing so. countb was incremented on our behalf. 128 */ 129 int 130 spin_trylock_contested(struct spinlock *spin) 131 { 132 globaldata_t gd = mycpu; 133 134 /*++spinlocks_contested1;*/ 135 --gd->gd_spinlocks_wr; 136 --gd->gd_curthread->td_critcount; 137 return (FALSE); 138 } 139 140 /* 141 * The spin_lock() inline was unable to acquire the lock. 142 * 143 * atomic_swap_int() is the absolute fastest spinlock instruction, at 144 * least on multi-socket systems. All instructions seem to be about 145 * the same on single-socket multi-core systems. However, atomic_swap_int() 146 * does not result in an even distribution of successful acquisitions. 147 * 148 * Another problem we have is that (at least on the 48-core opteron we test 149 * with) having all 48 cores contesting the same spin lock reduces 150 * performance to around 600,000 ops/sec, verses millions when fewer cores 151 * are going after the same lock. 152 * 153 * Backoff algorithms can create even worse starvation problems, and don't 154 * really improve performance when a lot of cores are contending. 155 * 156 * Our solution is to allow the data cache to lazy-update by reading it 157 * non-atomically and only attempting to acquire the lock if the lazy read 158 * looks good. This effectively limits cache bus bandwidth. A cpu_pause() 159 * (for intel/amd anyhow) is not strictly needed as cache bus resource use 160 * is governed by the lazy update. 161 * 162 * WARNING!!!! Performance matters here, by a huge margin. 163 * 164 * 48-core test with pre-read / -j 48 no-modules kernel compile 165 * with fanned-out inactive and active queues came in at 55 seconds. 166 * 167 * 48-core test with pre-read / -j 48 no-modules kernel compile 168 * came in at 75 seconds. Without pre-read it came in at 170 seconds. 169 * 170 * 4-core test with pre-read / -j 48 no-modules kernel compile 171 * came in at 83 seconds. Without pre-read it came in at 83 seconds 172 * as well (no difference). 173 */ 174 void 175 spin_lock_contested(struct spinlock *spin) 176 { 177 struct indefinite_info info = { 0, 0 }; 178 int i; 179 180 #ifdef DEBUG_LOCKS_LATENCY 181 long j; 182 for (j = spinlocks_add_latency; j > 0; --j) 183 cpu_ccfence(); 184 #endif 185 186 i = 0; 187 ++spin->countb; 188 189 /*logspin(beg, spin, 'w');*/ 190 for (;;) { 191 /* 192 * NOTE: Reading spin->counta prior to the swap is extremely 193 * important on multi-chip/many-core boxes. On 48-core 194 * this one change improves fully concurrent all-cores 195 * compiles by 100% or better. 196 * 197 * I can't emphasize enough how important the pre-read is in 198 * preventing hw cache bus armageddon on multi-chip systems. 199 * And on single-chip/multi-core systems it just doesn't hurt. 200 */ 201 if (spin->counta == 0 && atomic_swap_int(&spin->counta, 1) == 0) 202 break; 203 if ((++i & 0x7F) == 0x7F) { 204 ++spin->countb; 205 if (spin_indefinite_check(spin, &info)) 206 break; 207 } 208 } 209 /*logspin(end, spin, 'w');*/ 210 } 211 212 static __inline int 213 _spin_pool_hash(void *ptr) 214 { 215 int i; 216 217 i = ((int)(uintptr_t) ptr >> 5) ^ ((int)(uintptr_t)ptr >> 12); 218 i &= SPINLOCK_NUM_POOL_MASK; 219 return (i); 220 } 221 222 void 223 _spin_pool_lock(void *chan) 224 { 225 struct spinlock *sp; 226 227 sp = &pool_spinlocks[_spin_pool_hash(chan)].spin; 228 spin_lock(sp); 229 } 230 231 void 232 _spin_pool_unlock(void *chan) 233 { 234 struct spinlock *sp; 235 236 sp = &pool_spinlocks[_spin_pool_hash(chan)].spin; 237 spin_unlock(sp); 238 } 239 240 241 static 242 int 243 spin_indefinite_check(struct spinlock *spin, struct indefinite_info *info) 244 { 245 sysclock_t count; 246 247 cpu_spinlock_contested(); 248 249 count = sys_cputimer->count(); 250 if (info->secs == 0) { 251 info->base = count; 252 ++info->secs; 253 } else if (count - info->base > sys_cputimer->freq) { 254 kprintf("spin_lock: %p, indefinite wait (%d secs)!\n", 255 spin, info->secs); 256 info->base = count; 257 ++info->secs; 258 if (panicstr) 259 return (TRUE); 260 #if defined(INVARIANTS) 261 if (spin_lock_test_mode) { 262 print_backtrace(-1); 263 return (TRUE); 264 } 265 #endif 266 #if defined(INVARIANTS) 267 if (info->secs == 11) 268 print_backtrace(-1); 269 #endif 270 if (info->secs == 60) 271 panic("spin_lock: %p, indefinite wait!", spin); 272 } 273 return (FALSE); 274 } 275 276 /* 277 * If INVARIANTS is enabled various spinlock timing tests can be run 278 * by setting debug.spin_lock_test: 279 * 280 * 1 Test the indefinite wait code 281 * 2 Time the best-case exclusive lock overhead (spin_test_count) 282 * 3 Time the best-case shared lock overhead (spin_test_count) 283 */ 284 285 #ifdef INVARIANTS 286 287 static int spin_test_count = 10000000; 288 SYSCTL_INT(_debug, OID_AUTO, spin_test_count, CTLFLAG_RW, &spin_test_count, 0, 289 "Number of iterations to use for spinlock wait code test"); 290 291 static int 292 sysctl_spin_lock_test(SYSCTL_HANDLER_ARGS) 293 { 294 struct spinlock spin; 295 int error; 296 int value = 0; 297 int i; 298 299 if ((error = priv_check(curthread, PRIV_ROOT)) != 0) 300 return (error); 301 if ((error = SYSCTL_IN(req, &value, sizeof(value))) != 0) 302 return (error); 303 304 /* 305 * Indefinite wait test 306 */ 307 if (value == 1) { 308 spin_init(&spin); 309 spin_lock(&spin); /* force an indefinite wait */ 310 spin_lock_test_mode = 1; 311 spin_lock(&spin); 312 spin_unlock(&spin); /* Clean up the spinlock count */ 313 spin_unlock(&spin); 314 spin_lock_test_mode = 0; 315 } 316 317 /* 318 * Time best-case exclusive spinlocks 319 */ 320 if (value == 2) { 321 globaldata_t gd = mycpu; 322 323 spin_init(&spin); 324 for (i = spin_test_count; i > 0; --i) { 325 spin_lock_quick(gd, &spin); 326 spin_unlock_quick(gd, &spin); 327 } 328 } 329 330 return (0); 331 } 332 333 SYSCTL_PROC(_debug, KERN_PROC_ALL, spin_lock_test, CTLFLAG_RW|CTLTYPE_INT, 334 0, 0, sysctl_spin_lock_test, "I", "Test spinlock wait code"); 335 336 #endif /* INVARIANTS */ 337 #endif /* SMP */ 338