1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2003 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 /* 28 * Copyright (c) 2012, Joyent Inc. All rights reserved. 29 */ 30 31 #include <sys/timer.h> 32 #include <sys/systm.h> 33 #include <sys/param.h> 34 #include <sys/kmem.h> 35 #include <sys/debug.h> 36 #include <sys/cyclic.h> 37 #include <sys/cmn_err.h> 38 #include <sys/pset.h> 39 #include <sys/atomic.h> 40 #include <sys/policy.h> 41 42 static clock_backend_t clock_highres; 43 44 /*ARGSUSED*/ 45 static int 46 clock_highres_settime(timespec_t *ts) 47 { 48 return (EINVAL); 49 } 50 51 static int 52 clock_highres_gettime(timespec_t *ts) 53 { 54 hrt2ts(gethrtime(), (timestruc_t *)ts); 55 56 return (0); 57 } 58 59 static int 60 clock_highres_getres(timespec_t *ts) 61 { 62 hrt2ts(cyclic_getres(), (timestruc_t *)ts); 63 64 return (0); 65 } 66 67 /*ARGSUSED*/ 68 static int 69 clock_highres_timer_create(itimer_t *it, struct sigevent *ev) 70 { 71 /* 72 * CLOCK_HIGHRES timers of sufficiently high resolution can deny 73 * service; only allow privileged users to create such timers. 74 * Sites that do not wish to have this restriction should 75 * give users the "proc_clock_highres" privilege. 76 */ 77 if (secpolicy_clock_highres(CRED()) != 0) { 78 it->it_arg = NULL; 79 return (EPERM); 80 } 81 82 it->it_arg = kmem_zalloc(sizeof (cyclic_id_t), KM_SLEEP); 83 84 return (0); 85 } 86 87 static void 88 clock_highres_fire(void *arg) 89 { 90 itimer_t *it = (itimer_t *)arg; 91 hrtime_t *addr = &it->it_hrtime; 92 hrtime_t old = *addr, new = gethrtime(); 93 94 do { 95 old = *addr; 96 } while (atomic_cas_64((uint64_t *)addr, old, new) != old); 97 98 timer_fire(it); 99 } 100 101 static int 102 clock_highres_timer_settime(itimer_t *it, int flags, 103 const struct itimerspec *when) 104 { 105 cyclic_id_t cyc, *cycp = it->it_arg; 106 proc_t *p = curproc; 107 kthread_t *t = curthread; 108 cyc_time_t cyctime; 109 cyc_handler_t hdlr; 110 cpu_t *cpu; 111 cpupart_t *cpupart; 112 int pset; 113 114 cyctime.cyt_when = ts2hrt(&when->it_value); 115 cyctime.cyt_interval = ts2hrt(&when->it_interval); 116 117 if (cyctime.cyt_when != 0 && cyctime.cyt_interval == 0 && 118 it->it_itime.it_interval.tv_sec == 0 && 119 it->it_itime.it_interval.tv_nsec == 0 && 120 (cyc = *cycp) != CYCLIC_NONE) { 121 /* 122 * If our existing timer is a one-shot and our new timer is a 123 * one-shot, we'll save ourselves a world of grief and just 124 * reprogram the cyclic. 125 */ 126 it->it_itime = *when; 127 128 if (!(flags & TIMER_ABSTIME)) 129 cyctime.cyt_when += gethrtime(); 130 131 hrt2ts(cyctime.cyt_when, &it->it_itime.it_value); 132 (void) cyclic_reprogram(cyc, cyctime.cyt_when); 133 return (0); 134 } 135 136 mutex_enter(&cpu_lock); 137 if ((cyc = *cycp) != CYCLIC_NONE) { 138 cyclic_remove(cyc); 139 *cycp = CYCLIC_NONE; 140 } 141 142 if (cyctime.cyt_when == 0) { 143 mutex_exit(&cpu_lock); 144 return (0); 145 } 146 147 if (!(flags & TIMER_ABSTIME)) 148 cyctime.cyt_when += gethrtime(); 149 150 /* 151 * Now we will check for overflow (that is, we will check to see 152 * that the start time plus the interval time doesn't exceed 153 * INT64_MAX). The astute code reviewer will observe that this 154 * one-time check doesn't guarantee that a future expiration 155 * will not wrap. We wish to prove, then, that if a future 156 * expiration does wrap, the earliest the problem can be encountered 157 * is (INT64_MAX / 2) nanoseconds (191 years) after boot. Formally: 158 * 159 * Given: s + i < m s > 0 i > 0 160 * s + ni > m n > 1 161 * 162 * (where "s" is the start time, "i" is the interval, "n" is the 163 * number of times the cyclic has fired and "m" is INT64_MAX) 164 * 165 * Prove: 166 * (a) s + (n - 1)i > (m / 2) 167 * (b) s + (n - 1)i < m 168 * 169 * That is, prove that we must have fired at least once 191 years 170 * after boot. The proof is very straightforward; since the left 171 * side of (a) is minimized when i is small, it is sufficient to show 172 * that the statement is true for i's smallest possible value 173 * (((m - s) / n) + epsilon). The same goes for (b); showing that the 174 * statement is true for i's largest possible value (m - s + epsilon) 175 * is sufficient to prove the statement. 176 * 177 * The actual arithmetic manipulation is left up to reader. 178 */ 179 if (cyctime.cyt_when > INT64_MAX - cyctime.cyt_interval) { 180 mutex_exit(&cpu_lock); 181 return (EOVERFLOW); 182 } 183 184 if (cyctime.cyt_interval == 0) { 185 /* 186 * If this is a one-shot, then we set the interval to be 187 * inifinite. If this timer is never touched, this cyclic will 188 * simply consume space in the cyclic subsystem. As soon as 189 * timer_settime() or timer_delete() is called, the cyclic is 190 * removed (so it's not possible to run the machine out 191 * of resources by creating one-shots). 192 */ 193 cyctime.cyt_interval = CY_INFINITY; 194 } 195 196 it->it_itime = *when; 197 198 hrt2ts(cyctime.cyt_when, &it->it_itime.it_value); 199 200 hdlr.cyh_func = (cyc_func_t)clock_highres_fire; 201 hdlr.cyh_arg = it; 202 hdlr.cyh_level = CY_LOW_LEVEL; 203 204 if (cyctime.cyt_when != 0) 205 *cycp = cyc = cyclic_add(&hdlr, &cyctime); 206 207 /* 208 * Now that we have the cyclic created, we need to bind it to our 209 * bound CPU and processor set (if any). 210 */ 211 mutex_enter(&p->p_lock); 212 cpu = t->t_bound_cpu; 213 cpupart = t->t_cpupart; 214 pset = t->t_bind_pset; 215 216 mutex_exit(&p->p_lock); 217 218 cyclic_bind(cyc, cpu, pset == PS_NONE ? NULL : cpupart); 219 220 mutex_exit(&cpu_lock); 221 222 return (0); 223 } 224 225 static int 226 clock_highres_timer_gettime(itimer_t *it, struct itimerspec *when) 227 { 228 /* 229 * CLOCK_HIGHRES doesn't update it_itime. 230 */ 231 hrtime_t start = ts2hrt(&it->it_itime.it_value); 232 hrtime_t interval = ts2hrt(&it->it_itime.it_interval); 233 hrtime_t diff, now = gethrtime(); 234 hrtime_t *addr = &it->it_hrtime; 235 hrtime_t last; 236 237 /* 238 * We're using atomic_cas_64() here only to assure that we slurp the 239 * entire timestamp atomically. 240 */ 241 last = atomic_cas_64((uint64_t *)addr, 0, 0); 242 243 *when = it->it_itime; 244 245 if (!timerspecisset(&when->it_value)) 246 return (0); 247 248 if (start > now) { 249 /* 250 * We haven't gone off yet... 251 */ 252 diff = start - now; 253 } else { 254 if (interval == 0) { 255 /* 256 * This is a one-shot which should have already 257 * fired; set it_value to 0. 258 */ 259 timerspecclear(&when->it_value); 260 return (0); 261 } 262 263 /* 264 * Calculate how far we are into this interval. 265 */ 266 diff = (now - start) % interval; 267 268 /* 269 * Now check to see if we've dealt with the last interval 270 * yet. 271 */ 272 if (now - diff > last) { 273 /* 274 * The last interval hasn't fired; set it_value to 0. 275 */ 276 timerspecclear(&when->it_value); 277 return (0); 278 } 279 280 /* 281 * The last interval _has_ fired; we can return the amount 282 * of time left in this interval. 283 */ 284 diff = interval - diff; 285 } 286 287 hrt2ts(diff, &when->it_value); 288 289 return (0); 290 } 291 292 static int 293 clock_highres_timer_delete(itimer_t *it) 294 { 295 cyclic_id_t cyc; 296 297 if (it->it_arg == NULL) { 298 /* 299 * This timer was never fully created; we must have failed 300 * in the clock_highres_timer_create() routine. 301 */ 302 return (0); 303 } 304 305 mutex_enter(&cpu_lock); 306 307 if ((cyc = *((cyclic_id_t *)it->it_arg)) != CYCLIC_NONE) 308 cyclic_remove(cyc); 309 310 mutex_exit(&cpu_lock); 311 312 kmem_free(it->it_arg, sizeof (cyclic_id_t)); 313 314 return (0); 315 } 316 317 static void 318 clock_highres_timer_lwpbind(itimer_t *it) 319 { 320 proc_t *p = curproc; 321 kthread_t *t = curthread; 322 cyclic_id_t cyc = *((cyclic_id_t *)it->it_arg); 323 cpu_t *cpu; 324 cpupart_t *cpupart; 325 int pset; 326 327 if (cyc == CYCLIC_NONE) 328 return; 329 330 mutex_enter(&cpu_lock); 331 mutex_enter(&p->p_lock); 332 333 /* 334 * Okay, now we can safely look at the bindings. 335 */ 336 cpu = t->t_bound_cpu; 337 cpupart = t->t_cpupart; 338 pset = t->t_bind_pset; 339 340 /* 341 * Now we drop p_lock. We haven't dropped cpu_lock; we're guaranteed 342 * that even if the bindings change, the CPU and/or processor set 343 * that this timer was bound to remain valid (and the combination 344 * remains self-consistent). 345 */ 346 mutex_exit(&p->p_lock); 347 348 cyclic_bind(cyc, cpu, pset == PS_NONE ? NULL : cpupart); 349 350 mutex_exit(&cpu_lock); 351 } 352 353 void 354 clock_highres_init() 355 { 356 clock_backend_t *be = &clock_highres; 357 struct sigevent *ev = &be->clk_default; 358 359 ev->sigev_signo = SIGALRM; 360 ev->sigev_notify = SIGEV_SIGNAL; 361 ev->sigev_value.sival_ptr = NULL; 362 363 be->clk_clock_settime = clock_highres_settime; 364 be->clk_clock_gettime = clock_highres_gettime; 365 be->clk_clock_getres = clock_highres_getres; 366 be->clk_timer_create = clock_highres_timer_create; 367 be->clk_timer_gettime = clock_highres_timer_gettime; 368 be->clk_timer_settime = clock_highres_timer_settime; 369 be->clk_timer_delete = clock_highres_timer_delete; 370 be->clk_timer_lwpbind = clock_highres_timer_lwpbind; 371 372 clock_add_backend(CLOCK_HIGHRES, &clock_highres); 373 } 374