| 1 | /*- |
| 2 | * Copyright (c) 1982, 1986, 1990, 1991, 1993 |
| 3 | * The Regents of the University of California. All rights reserved. |
| 4 | * (c) UNIX System Laboratories, Inc. |
| 5 | * All or some portions of this file are derived from material licensed |
| 6 | * to the University of California by American Telephone and Telegraph |
| 7 | * Co. or Unix System Laboratories, Inc. and are reproduced herein with |
| 8 | * the permission of UNIX System Laboratories, Inc. |
| 9 | * |
| 10 | * Redistribution and use in source and binary forms, with or without |
| 11 | * modification, are permitted provided that the following conditions |
| 12 | * are met: |
| 13 | * 1. Redistributions of source code must retain the above copyright |
| 14 | * notice, this list of conditions and the following disclaimer. |
| 15 | * 2. Redistributions in binary form must reproduce the above copyright |
| 16 | * notice, this list of conditions and the following disclaimer in the |
| 17 | * documentation and/or other materials provided with the distribution. |
| 18 | * 3. All advertising materials mentioning features or use of this software |
| 19 | * must display the following acknowledgement: |
| 20 | * This product includes software developed by the University of |
| 21 | * California, Berkeley and its contributors. |
| 22 | * 4. Neither the name of the University nor the names of its contributors |
| 23 | * may be used to endorse or promote products derived from this software |
| 24 | * without specific prior written permission. |
| 25 | * |
| 26 | * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND |
| 27 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 28 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 29 | * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE |
| 30 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 31 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 32 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 33 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 34 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 35 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 36 | * SUCH DAMAGE. |
| 37 | * |
| 38 | * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 |
| 39 | * $FreeBSD: src/sys/kern/kern_synch.c,v 1.87.2.6 2002/10/13 07:29:53 kbyanc Exp $ |
| 40 | * $DragonFly: src/sys/kern/kern_synch.c,v 1.20 2003/08/03 10:07:41 hmp Exp $ |
| 41 | */ |
| 42 | |
| 43 | #include "opt_ktrace.h" |
| 44 | |
| 45 | #include <sys/param.h> |
| 46 | #include <sys/systm.h> |
| 47 | #include <sys/proc.h> |
| 48 | #include <sys/kernel.h> |
| 49 | #include <sys/signalvar.h> |
| 50 | #include <sys/resourcevar.h> |
| 51 | #include <sys/vmmeter.h> |
| 52 | #include <sys/sysctl.h> |
| 53 | #include <sys/thread2.h> |
| 54 | #ifdef KTRACE |
| 55 | #include <sys/uio.h> |
| 56 | #include <sys/ktrace.h> |
| 57 | #endif |
| 58 | #include <sys/xwait.h> |
| 59 | |
| 60 | #include <machine/cpu.h> |
| 61 | #include <machine/ipl.h> |
| 62 | #include <machine/smp.h> |
| 63 | |
| 64 | static void sched_setup __P((void *dummy)); |
| 65 | SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) |
| 66 | |
| 67 | int hogticks; |
| 68 | int lbolt; |
| 69 | int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ |
| 70 | int ncpus; |
| 71 | |
| 72 | static struct callout loadav_callout; |
| 73 | |
| 74 | struct loadavg averunnable = |
| 75 | { {0, 0, 0}, FSCALE }; /* load average, of runnable procs */ |
| 76 | /* |
| 77 | * Constants for averages over 1, 5, and 15 minutes |
| 78 | * when sampling at 5 second intervals. |
| 79 | */ |
| 80 | static fixpt_t cexp[3] = { |
| 81 | 0.9200444146293232 * FSCALE, /* exp(-1/12) */ |
| 82 | 0.9834714538216174 * FSCALE, /* exp(-1/60) */ |
| 83 | 0.9944598480048967 * FSCALE, /* exp(-1/180) */ |
| 84 | }; |
| 85 | |
| 86 | static void endtsleep __P((void *)); |
| 87 | static void loadav __P((void *arg)); |
| 88 | static void maybe_resched __P((struct proc *chk)); |
| 89 | static void roundrobin __P((void *arg)); |
| 90 | static void schedcpu __P((void *arg)); |
| 91 | static void updatepri __P((struct proc *p)); |
| 92 | static void crit_panicints(void); |
| 93 | |
| 94 | static int |
| 95 | sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) |
| 96 | { |
| 97 | int error, new_val; |
| 98 | |
| 99 | new_val = sched_quantum * tick; |
| 100 | error = sysctl_handle_int(oidp, &new_val, 0, req); |
| 101 | if (error != 0 || req->newptr == NULL) |
| 102 | return (error); |
| 103 | if (new_val < tick) |
| 104 | return (EINVAL); |
| 105 | sched_quantum = new_val / tick; |
| 106 | hogticks = 2 * sched_quantum; |
| 107 | return (0); |
| 108 | } |
| 109 | |
| 110 | SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, |
| 111 | 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); |
| 112 | |
| 113 | /* |
| 114 | * Arrange to reschedule if necessary by checking to see if the current |
| 115 | * process is on the highest priority user scheduling queue. This may |
| 116 | * be run from an interrupt so we have to follow any preemption chains |
| 117 | * back to the original process. |
| 118 | */ |
| 119 | static void |
| 120 | maybe_resched(struct proc *chk) |
| 121 | { |
| 122 | struct proc *cur = lwkt_preempted_proc(); |
| 123 | |
| 124 | if (cur == NULL) |
| 125 | return; |
| 126 | |
| 127 | /* |
| 128 | * Check the user queue (realtime, normal, idle). Lower numbers |
| 129 | * indicate higher priority queues. Lower numbers are also better |
| 130 | * for p_priority. |
| 131 | */ |
| 132 | if (chk->p_rtprio.type < cur->p_rtprio.type) { |
| 133 | need_resched(); |
| 134 | } else if (chk->p_rtprio.type == cur->p_rtprio.type) { |
| 135 | if (chk->p_rtprio.type == RTP_PRIO_NORMAL) { |
| 136 | if (chk->p_priority / PPQ < cur->p_priority / PPQ) |
| 137 | need_resched(); |
| 138 | } else { |
| 139 | if (chk->p_rtprio.prio < cur->p_rtprio.prio) |
| 140 | need_resched(); |
| 141 | } |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | int |
| 146 | roundrobin_interval(void) |
| 147 | { |
| 148 | return (sched_quantum); |
| 149 | } |
| 150 | |
| 151 | /* |
| 152 | * Force switch among equal priority processes every 100ms. |
| 153 | */ |
| 154 | #ifdef SMP |
| 155 | |
| 156 | static void |
| 157 | roundrobin_remote(void *arg) |
| 158 | { |
| 159 | struct proc *p = lwkt_preempted_proc(); |
| 160 | if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type)) |
| 161 | need_resched(); |
| 162 | } |
| 163 | |
| 164 | #endif |
| 165 | |
| 166 | static void |
| 167 | roundrobin(void *arg) |
| 168 | { |
| 169 | struct proc *p = lwkt_preempted_proc(); |
| 170 | if (p == NULL || RTP_PRIO_NEED_RR(p->p_rtprio.type)) |
| 171 | need_resched(); |
| 172 | #ifdef SMP |
| 173 | lwkt_send_ipiq_mask(mycpu->gd_other_cpus, roundrobin_remote, NULL); |
| 174 | #endif |
| 175 | timeout(roundrobin, NULL, sched_quantum); |
| 176 | } |
| 177 | |
| 178 | #ifdef SMP |
| 179 | |
| 180 | void |
| 181 | resched_cpus(u_int32_t mask) |
| 182 | { |
| 183 | lwkt_send_ipiq_mask(mask, roundrobin_remote, NULL); |
| 184 | } |
| 185 | |
| 186 | #endif |
| 187 | |
| 188 | /* |
| 189 | * Constants for digital decay and forget: |
| 190 | * 90% of (p_estcpu) usage in 5 * loadav time |
| 191 | * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) |
| 192 | * Note that, as ps(1) mentions, this can let percentages |
| 193 | * total over 100% (I've seen 137.9% for 3 processes). |
| 194 | * |
| 195 | * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. |
| 196 | * |
| 197 | * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. |
| 198 | * That is, the system wants to compute a value of decay such |
| 199 | * that the following for loop: |
| 200 | * for (i = 0; i < (5 * loadavg); i++) |
| 201 | * p_estcpu *= decay; |
| 202 | * will compute |
| 203 | * p_estcpu *= 0.1; |
| 204 | * for all values of loadavg: |
| 205 | * |
| 206 | * Mathematically this loop can be expressed by saying: |
| 207 | * decay ** (5 * loadavg) ~= .1 |
| 208 | * |
| 209 | * The system computes decay as: |
| 210 | * decay = (2 * loadavg) / (2 * loadavg + 1) |
| 211 | * |
| 212 | * We wish to prove that the system's computation of decay |
| 213 | * will always fulfill the equation: |
| 214 | * decay ** (5 * loadavg) ~= .1 |
| 215 | * |
| 216 | * If we compute b as: |
| 217 | * b = 2 * loadavg |
| 218 | * then |
| 219 | * decay = b / (b + 1) |
| 220 | * |
| 221 | * We now need to prove two things: |
| 222 | * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) |
| 223 | * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) |
| 224 | * |
| 225 | * Facts: |
| 226 | * For x close to zero, exp(x) =~ 1 + x, since |
| 227 | * exp(x) = 0! + x**1/1! + x**2/2! + ... . |
| 228 | * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. |
| 229 | * For x close to zero, ln(1+x) =~ x, since |
| 230 | * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 |
| 231 | * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). |
| 232 | * ln(.1) =~ -2.30 |
| 233 | * |
| 234 | * Proof of (1): |
| 235 | * Solve (factor)**(power) =~ .1 given power (5*loadav): |
| 236 | * solving for factor, |
| 237 | * ln(factor) =~ (-2.30/5*loadav), or |
| 238 | * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = |
| 239 | * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED |
| 240 | * |
| 241 | * Proof of (2): |
| 242 | * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): |
| 243 | * solving for power, |
| 244 | * power*ln(b/(b+1)) =~ -2.30, or |
| 245 | * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED |
| 246 | * |
| 247 | * Actual power values for the implemented algorithm are as follows: |
| 248 | * loadav: 1 2 3 4 |
| 249 | * power: 5.68 10.32 14.94 19.55 |
| 250 | */ |
| 251 | |
| 252 | /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ |
| 253 | #define loadfactor(loadav) (2 * (loadav)) |
| 254 | #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) |
| 255 | |
| 256 | /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ |
| 257 | static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ |
| 258 | SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); |
| 259 | |
| 260 | /* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ |
| 261 | static int fscale __unused = FSCALE; |
| 262 | SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); |
| 263 | |
| 264 | /* |
| 265 | * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the |
| 266 | * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below |
| 267 | * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). |
| 268 | * |
| 269 | * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: |
| 270 | * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). |
| 271 | * |
| 272 | * If you don't want to bother with the faster/more-accurate formula, you |
| 273 | * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate |
| 274 | * (more general) method of calculating the %age of CPU used by a process. |
| 275 | */ |
| 276 | #define CCPU_SHIFT 11 |
| 277 | |
| 278 | /* |
| 279 | * Recompute process priorities, every hz ticks. |
| 280 | */ |
| 281 | /* ARGSUSED */ |
| 282 | static void |
| 283 | schedcpu(void *arg) |
| 284 | { |
| 285 | fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); |
| 286 | struct proc *p; |
| 287 | struct proc *curp; |
| 288 | int realstathz, s; |
| 289 | |
| 290 | curp = lwkt_preempted_proc(); /* YYY temporary hack */ |
| 291 | |
| 292 | realstathz = stathz ? stathz : hz; |
| 293 | FOREACH_PROC_IN_SYSTEM(p) { |
| 294 | /* |
| 295 | * Increment time in/out of memory and sleep time |
| 296 | * (if sleeping). We ignore overflow; with 16-bit int's |
| 297 | * (remember them?) overflow takes 45 days. |
| 298 | */ |
| 299 | p->p_swtime++; |
| 300 | if (p->p_stat == SSLEEP || p->p_stat == SSTOP) |
| 301 | p->p_slptime++; |
| 302 | p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; |
| 303 | /* |
| 304 | * If the process has slept the entire second, |
| 305 | * stop recalculating its priority until it wakes up. |
| 306 | */ |
| 307 | if (p->p_slptime > 1) |
| 308 | continue; |
| 309 | s = splhigh(); /* prevent state changes and protect run queue */ |
| 310 | /* |
| 311 | * p_pctcpu is only for ps. |
| 312 | */ |
| 313 | #if (FSHIFT >= CCPU_SHIFT) |
| 314 | p->p_pctcpu += (realstathz == 100)? |
| 315 | ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): |
| 316 | 100 * (((fixpt_t) p->p_cpticks) |
| 317 | << (FSHIFT - CCPU_SHIFT)) / realstathz; |
| 318 | #else |
| 319 | p->p_pctcpu += ((FSCALE - ccpu) * |
| 320 | (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; |
| 321 | #endif |
| 322 | p->p_cpticks = 0; |
| 323 | p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); |
| 324 | resetpriority(p); |
| 325 | splx(s); |
| 326 | } |
| 327 | wakeup((caddr_t)&lbolt); |
| 328 | timeout(schedcpu, (void *)0, hz); |
| 329 | } |
| 330 | |
| 331 | /* |
| 332 | * Recalculate the priority of a process after it has slept for a while. |
| 333 | * For all load averages >= 1 and max p_estcpu of 255, sleeping for at |
| 334 | * least six times the loadfactor will decay p_estcpu to zero. |
| 335 | */ |
| 336 | static void |
| 337 | updatepri(struct proc *p) |
| 338 | { |
| 339 | unsigned int newcpu = p->p_estcpu; |
| 340 | fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); |
| 341 | |
| 342 | if (p->p_slptime > 5 * loadfac) { |
| 343 | p->p_estcpu = 0; |
| 344 | } else { |
| 345 | p->p_slptime--; /* the first time was done in schedcpu */ |
| 346 | while (newcpu && --p->p_slptime) |
| 347 | newcpu = decay_cpu(loadfac, newcpu); |
| 348 | p->p_estcpu = newcpu; |
| 349 | } |
| 350 | resetpriority(p); |
| 351 | } |
| 352 | |
| 353 | /* |
| 354 | * We're only looking at 7 bits of the address; everything is |
| 355 | * aligned to 4, lots of things are aligned to greater powers |
| 356 | * of 2. Shift right by 8, i.e. drop the bottom 256 worth. |
| 357 | */ |
| 358 | #define TABLESIZE 128 |
| 359 | static TAILQ_HEAD(slpquehead, thread) slpque[TABLESIZE]; |
| 360 | #define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) |
| 361 | |
| 362 | /* |
| 363 | * During autoconfiguration or after a panic, a sleep will simply |
| 364 | * lower the priority briefly to allow interrupts, then return. |
| 365 | * The priority to be used (safepri) is machine-dependent, thus this |
| 366 | * value is initialized and maintained in the machine-dependent layers. |
| 367 | * This priority will typically be 0, or the lowest priority |
| 368 | * that is safe for use on the interrupt stack; it can be made |
| 369 | * higher to block network software interrupts after panics. |
| 370 | */ |
| 371 | int safepri; |
| 372 | |
| 373 | void |
| 374 | sleepinit(void) |
| 375 | { |
| 376 | int i; |
| 377 | |
| 378 | sched_quantum = hz/10; |
| 379 | hogticks = 2 * sched_quantum; |
| 380 | for (i = 0; i < TABLESIZE; i++) |
| 381 | TAILQ_INIT(&slpque[i]); |
| 382 | } |
| 383 | |
| 384 | /* |
| 385 | * General sleep call. Suspends the current process until a wakeup is |
| 386 | * performed on the specified identifier. The process will then be made |
| 387 | * runnable with the specified priority. Sleeps at most timo/hz seconds |
| 388 | * (0 means no timeout). If flags includes PCATCH flag, signals are checked |
| 389 | * before and after sleeping, else signals are not checked. Returns 0 if |
| 390 | * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a |
| 391 | * signal needs to be delivered, ERESTART is returned if the current system |
| 392 | * call should be restarted if possible, and EINTR is returned if the system |
| 393 | * call should be interrupted by the signal (return EINTR). |
| 394 | * |
| 395 | * If the process has P_CURPROC set mi_switch() will not re-queue it to |
| 396 | * the userland scheduler queues because we are in a SSLEEP state. If |
| 397 | * we are not the current process then we have to remove ourselves from |
| 398 | * the scheduler queues. |
| 399 | * |
| 400 | * YYY priority now unused |
| 401 | */ |
| 402 | int |
| 403 | tsleep(ident, flags, wmesg, timo) |
| 404 | void *ident; |
| 405 | int flags, timo; |
| 406 | const char *wmesg; |
| 407 | { |
| 408 | struct thread *td = curthread; |
| 409 | struct proc *p = td->td_proc; /* may be NULL */ |
| 410 | int s, sig = 0, catch = flags & PCATCH; |
| 411 | int id = LOOKUP(ident); |
| 412 | struct callout_handle thandle; |
| 413 | |
| 414 | /* |
| 415 | * NOTE: removed KTRPOINT, it could cause races due to blocking |
| 416 | * even in stable. Just scrap it for now. |
| 417 | */ |
| 418 | if (cold || panicstr) { |
| 419 | /* |
| 420 | * After a panic, or during autoconfiguration, |
| 421 | * just give interrupts a chance, then just return; |
| 422 | * don't run any other procs or panic below, |
| 423 | * in case this is the idle process and already asleep. |
| 424 | */ |
| 425 | crit_panicints(); |
| 426 | return (0); |
| 427 | } |
| 428 | KKASSERT(td != &mycpu->gd_idlethread); /* you must be kidding! */ |
| 429 | s = splhigh(); |
| 430 | KASSERT(ident != NULL, ("tsleep: no ident")); |
| 431 | KASSERT(p == NULL || p->p_stat == SRUN, ("tsleep %p %s %d", |
| 432 | ident, wmesg, p->p_stat)); |
| 433 | |
| 434 | crit_enter(); |
| 435 | td->td_wchan = ident; |
| 436 | td->td_wmesg = wmesg; |
| 437 | if (p) |
| 438 | p->p_slptime = 0; |
| 439 | lwkt_deschedule_self(); |
| 440 | TAILQ_INSERT_TAIL(&slpque[id], td, td_threadq); |
| 441 | if (timo) |
| 442 | thandle = timeout(endtsleep, (void *)td, timo); |
| 443 | /* |
| 444 | * We put ourselves on the sleep queue and start our timeout |
| 445 | * before calling CURSIG, as we could stop there, and a wakeup |
| 446 | * or a SIGCONT (or both) could occur while we were stopped. |
| 447 | * A SIGCONT would cause us to be marked as SSLEEP |
| 448 | * without resuming us, thus we must be ready for sleep |
| 449 | * when CURSIG is called. If the wakeup happens while we're |
| 450 | * stopped, td->td_wchan will be 0 upon return from CURSIG. |
| 451 | */ |
| 452 | if (p) { |
| 453 | if (catch) { |
| 454 | p->p_flag |= P_SINTR; |
| 455 | if ((sig = CURSIG(p))) { |
| 456 | if (td->td_wchan) { |
| 457 | unsleep(td); |
| 458 | lwkt_schedule_self(); |
| 459 | } |
| 460 | p->p_stat = SRUN; |
| 461 | goto resume; |
| 462 | } |
| 463 | if (td->td_wchan == NULL) { |
| 464 | catch = 0; |
| 465 | goto resume; |
| 466 | } |
| 467 | } else { |
| 468 | sig = 0; |
| 469 | } |
| 470 | |
| 471 | /* |
| 472 | * If we are not the current process we have to remove ourself |
| 473 | * from the run queue. |
| 474 | */ |
| 475 | KASSERT(p->p_stat == SRUN, ("PSTAT NOT SRUN %d %d", p->p_pid, p->p_stat)); |
| 476 | /* |
| 477 | * If this is the current 'user' process schedule another one. |
| 478 | */ |
| 479 | clrrunnable(p, SSLEEP); |
| 480 | p->p_stats->p_ru.ru_nvcsw++; |
| 481 | KKASSERT(td->td_release || (p->p_flag & P_CURPROC) == 0); |
| 482 | mi_switch(); |
| 483 | KASSERT(p->p_stat == SRUN, ("tsleep: stat not srun")); |
| 484 | } else { |
| 485 | lwkt_switch(); |
| 486 | } |
| 487 | resume: |
| 488 | crit_exit(); |
| 489 | if (p) |
| 490 | p->p_flag &= ~P_SINTR; |
| 491 | splx(s); |
| 492 | if (td->td_flags & TDF_TIMEOUT) { |
| 493 | td->td_flags &= ~TDF_TIMEOUT; |
| 494 | if (sig == 0) |
| 495 | return (EWOULDBLOCK); |
| 496 | } else if (timo) { |
| 497 | untimeout(endtsleep, (void *)td, thandle); |
| 498 | } |
| 499 | if (p) { |
| 500 | if (catch && (sig != 0 || (sig = CURSIG(p)))) { |
| 501 | if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) |
| 502 | return (EINTR); |
| 503 | return (ERESTART); |
| 504 | } |
| 505 | } |
| 506 | return (0); |
| 507 | } |
| 508 | |
| 509 | /* |
| 510 | * Implement the timeout for tsleep. We interlock against |
| 511 | * wchan when setting TDF_TIMEOUT. For processes we remove |
| 512 | * the sleep if the process is stopped rather then sleeping, |
| 513 | * so it remains stopped. |
| 514 | */ |
| 515 | static void |
| 516 | endtsleep(void *arg) |
| 517 | { |
| 518 | thread_t td = arg; |
| 519 | struct proc *p; |
| 520 | int s; |
| 521 | |
| 522 | s = splhigh(); |
| 523 | if (td->td_wchan) { |
| 524 | td->td_flags |= TDF_TIMEOUT; |
| 525 | if ((p = td->td_proc) != NULL) { |
| 526 | if (p->p_stat == SSLEEP) |
| 527 | setrunnable(p); |
| 528 | else |
| 529 | unsleep(td); |
| 530 | } else { |
| 531 | unsleep(td); |
| 532 | lwkt_schedule(td); |
| 533 | } |
| 534 | } |
| 535 | splx(s); |
| 536 | } |
| 537 | |
| 538 | /* |
| 539 | * Remove a process from its wait queue |
| 540 | */ |
| 541 | void |
| 542 | unsleep(struct thread *td) |
| 543 | { |
| 544 | int s; |
| 545 | |
| 546 | s = splhigh(); |
| 547 | if (td->td_wchan) { |
| 548 | #if 0 |
| 549 | if (p->p_flag & P_XSLEEP) { |
| 550 | struct xwait *w = p->p_wchan; |
| 551 | TAILQ_REMOVE(&w->waitq, p, p_procq); |
| 552 | p->p_flag &= ~P_XSLEEP; |
| 553 | } else |
| 554 | #endif |
| 555 | TAILQ_REMOVE(&slpque[LOOKUP(td->td_wchan)], td, td_threadq); |
| 556 | td->td_wchan = NULL; |
| 557 | } |
| 558 | splx(s); |
| 559 | } |
| 560 | |
| 561 | #if 0 |
| 562 | /* |
| 563 | * Make all processes sleeping on the explicit lock structure runnable. |
| 564 | */ |
| 565 | void |
| 566 | xwakeup(struct xwait *w) |
| 567 | { |
| 568 | struct proc *p; |
| 569 | int s; |
| 570 | |
| 571 | s = splhigh(); |
| 572 | ++w->gen; |
| 573 | while ((p = TAILQ_FIRST(&w->waitq)) != NULL) { |
| 574 | TAILQ_REMOVE(&w->waitq, p, p_procq); |
| 575 | KASSERT(p->p_wchan == w && (p->p_flag & P_XSLEEP), |
| 576 | ("xwakeup: wchan mismatch for %p (%p/%p) %08x", p, p->p_wchan, w, p->p_flag & P_XSLEEP)); |
| 577 | p->p_wchan = NULL; |
| 578 | p->p_flag &= ~P_XSLEEP; |
| 579 | if (p->p_stat == SSLEEP) { |
| 580 | /* OPTIMIZED EXPANSION OF setrunnable(p); */ |
| 581 | if (p->p_slptime > 1) |
| 582 | updatepri(p); |
| 583 | p->p_slptime = 0; |
| 584 | p->p_stat = SRUN; |
| 585 | if (p->p_flag & P_INMEM) { |
| 586 | setrunqueue(p); |
| 587 | maybe_resched(p); |
| 588 | } else { |
| 589 | p->p_flag |= P_SWAPINREQ; |
| 590 | wakeup((caddr_t)&proc0); |
| 591 | } |
| 592 | } |
| 593 | } |
| 594 | splx(s); |
| 595 | } |
| 596 | #endif |
| 597 | |
| 598 | /* |
| 599 | * Make all processes sleeping on the specified identifier runnable. |
| 600 | */ |
| 601 | static void |
| 602 | _wakeup(void *ident, int count) |
| 603 | { |
| 604 | struct slpquehead *qp; |
| 605 | struct thread *td; |
| 606 | struct thread *ntd; |
| 607 | struct proc *p; |
| 608 | int s; |
| 609 | int id = LOOKUP(ident); |
| 610 | |
| 611 | s = splhigh(); |
| 612 | qp = &slpque[id]; |
| 613 | restart: |
| 614 | for (td = TAILQ_FIRST(qp); td != NULL; td = ntd) { |
| 615 | ntd = TAILQ_NEXT(td, td_threadq); |
| 616 | if (td->td_wchan == ident) { |
| 617 | TAILQ_REMOVE(qp, td, td_threadq); |
| 618 | td->td_wchan = NULL; |
| 619 | if ((p = td->td_proc) != NULL && p->p_stat == SSLEEP) { |
| 620 | /* OPTIMIZED EXPANSION OF setrunnable(p); */ |
| 621 | if (p->p_slptime > 1) |
| 622 | updatepri(p); |
| 623 | p->p_slptime = 0; |
| 624 | p->p_stat = SRUN; |
| 625 | if (p->p_flag & P_INMEM) { |
| 626 | setrunqueue(p); |
| 627 | if (p->p_flag & P_CURPROC) |
| 628 | maybe_resched(p); |
| 629 | } else { |
| 630 | p->p_flag |= P_SWAPINREQ; |
| 631 | wakeup((caddr_t)&proc0); |
| 632 | } |
| 633 | /* END INLINE EXPANSION */ |
| 634 | } else if (p == NULL) { |
| 635 | lwkt_schedule(td); |
| 636 | } |
| 637 | if (--count == 0) |
| 638 | break; |
| 639 | goto restart; |
| 640 | } |
| 641 | } |
| 642 | splx(s); |
| 643 | } |
| 644 | |
| 645 | void |
| 646 | wakeup(void *ident) |
| 647 | { |
| 648 | _wakeup(ident, 0); |
| 649 | } |
| 650 | |
| 651 | void |
| 652 | wakeup_one(void *ident) |
| 653 | { |
| 654 | _wakeup(ident, 1); |
| 655 | } |
| 656 | |
| 657 | /* |
| 658 | * The machine independent parts of mi_switch(). |
| 659 | * Must be called at splstatclock() or higher. |
| 660 | */ |
| 661 | void |
| 662 | mi_switch() |
| 663 | { |
| 664 | struct thread *td = curthread; |
| 665 | struct proc *p = td->td_proc; /* XXX */ |
| 666 | struct rlimit *rlim; |
| 667 | int x; |
| 668 | u_int64_t ttime; |
| 669 | |
| 670 | /* |
| 671 | * XXX this spl is almost unnecessary. It is partly to allow for |
| 672 | * sloppy callers that don't do it (issignal() via CURSIG() is the |
| 673 | * main offender). It is partly to work around a bug in the i386 |
| 674 | * cpu_switch() (the ipl is not preserved). We ran for years |
| 675 | * without it. I think there was only a interrupt latency problem. |
| 676 | * The main caller, tsleep(), does an splx() a couple of instructions |
| 677 | * after calling here. The buggy caller, issignal(), usually calls |
| 678 | * here at spl0() and sometimes returns at splhigh(). The process |
| 679 | * then runs for a little too long at splhigh(). The ipl gets fixed |
| 680 | * when the process returns to user mode (or earlier). |
| 681 | * |
| 682 | * It would probably be better to always call here at spl0(). Callers |
| 683 | * are prepared to give up control to another process, so they must |
| 684 | * be prepared to be interrupted. The clock stuff here may not |
| 685 | * actually need splstatclock(). |
| 686 | */ |
| 687 | x = splstatclock(); |
| 688 | clear_resched(); |
| 689 | |
| 690 | /* |
| 691 | * Check if the process exceeds its cpu resource allocation. |
| 692 | * If over max, kill it. Time spent in interrupts is not |
| 693 | * included. YYY 64 bit match is expensive. Ick. |
| 694 | */ |
| 695 | ttime = td->td_sticks + td->td_uticks; |
| 696 | if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && |
| 697 | ttime > p->p_limit->p_cpulimit) { |
| 698 | rlim = &p->p_rlimit[RLIMIT_CPU]; |
| 699 | if (ttime / (rlim_t)1000000 >= rlim->rlim_max) { |
| 700 | killproc(p, "exceeded maximum CPU limit"); |
| 701 | } else { |
| 702 | psignal(p, SIGXCPU); |
| 703 | if (rlim->rlim_cur < rlim->rlim_max) { |
| 704 | /* XXX: we should make a private copy */ |
| 705 | rlim->rlim_cur += 5; |
| 706 | } |
| 707 | } |
| 708 | } |
| 709 | |
| 710 | /* |
| 711 | * Pick a new current process and record its start time. If we |
| 712 | * are in a SSTOPped state we deschedule ourselves. YYY this needs |
| 713 | * to be cleaned up, remember that LWKTs stay on their run queue |
| 714 | * which works differently then the user scheduler which removes |
| 715 | * the process from the runq when it runs it. |
| 716 | */ |
| 717 | mycpu->gd_cnt.v_swtch++; |
| 718 | if (p->p_stat == SSTOP) |
| 719 | lwkt_deschedule_self(); |
| 720 | lwkt_switch(); |
| 721 | |
| 722 | splx(x); |
| 723 | } |
| 724 | |
| 725 | /* |
| 726 | * Change process state to be runnable, |
| 727 | * placing it on the run queue if it is in memory, |
| 728 | * and awakening the swapper if it isn't in memory. |
| 729 | */ |
| 730 | void |
| 731 | setrunnable(struct proc *p) |
| 732 | { |
| 733 | int s; |
| 734 | |
| 735 | s = splhigh(); |
| 736 | switch (p->p_stat) { |
| 737 | case 0: |
| 738 | case SRUN: |
| 739 | case SZOMB: |
| 740 | default: |
| 741 | panic("setrunnable"); |
| 742 | case SSTOP: |
| 743 | case SSLEEP: |
| 744 | unsleep(p->p_thread); /* e.g. when sending signals */ |
| 745 | break; |
| 746 | |
| 747 | case SIDL: |
| 748 | break; |
| 749 | } |
| 750 | p->p_stat = SRUN; |
| 751 | if (p->p_flag & P_INMEM) |
| 752 | setrunqueue(p); |
| 753 | splx(s); |
| 754 | if (p->p_slptime > 1) |
| 755 | updatepri(p); |
| 756 | p->p_slptime = 0; |
| 757 | if ((p->p_flag & P_INMEM) == 0) { |
| 758 | p->p_flag |= P_SWAPINREQ; |
| 759 | wakeup((caddr_t)&proc0); |
| 760 | } else { |
| 761 | maybe_resched(p); |
| 762 | } |
| 763 | } |
| 764 | |
| 765 | /* |
| 766 | * Change the process state to NOT be runnable, removing it from the run |
| 767 | * queue. If P_CURPROC is not set and we are in SRUN the process is on the |
| 768 | * run queue (If P_INMEM is not set then it isn't because it is swapped). |
| 769 | */ |
| 770 | void |
| 771 | clrrunnable(struct proc *p, int stat) |
| 772 | { |
| 773 | int s; |
| 774 | |
| 775 | s = splhigh(); |
| 776 | switch(p->p_stat) { |
| 777 | case SRUN: |
| 778 | if (p->p_flag & P_ONRUNQ) |
| 779 | remrunqueue(p); |
| 780 | break; |
| 781 | default: |
| 782 | break; |
| 783 | } |
| 784 | p->p_stat = stat; |
| 785 | splx(s); |
| 786 | } |
| 787 | |
| 788 | /* |
| 789 | * Compute the priority of a process when running in user mode. |
| 790 | * Arrange to reschedule if the resulting priority is better |
| 791 | * than that of the current process. |
| 792 | * |
| 793 | * YYY real time / idle procs do not use p_priority XXX |
| 794 | */ |
| 795 | void |
| 796 | resetpriority(struct proc *p) |
| 797 | { |
| 798 | unsigned int newpriority; |
| 799 | int opq; |
| 800 | int npq; |
| 801 | |
| 802 | if (p->p_rtprio.type != RTP_PRIO_NORMAL) |
| 803 | return; |
| 804 | newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + |
| 805 | NICE_WEIGHT * p->p_nice; |
| 806 | newpriority = min(newpriority, MAXPRI); |
| 807 | npq = newpriority / PPQ; |
| 808 | crit_enter(); |
| 809 | opq = p->p_priority / PPQ; |
| 810 | if (p->p_stat == SRUN && (p->p_flag & P_ONRUNQ) && opq != npq) { |
| 811 | /* |
| 812 | * We have to move the process to another queue |
| 813 | */ |
| 814 | remrunqueue(p); |
| 815 | p->p_priority = newpriority; |
| 816 | setrunqueue(p); |
| 817 | } else { |
| 818 | /* |
| 819 | * We can just adjust the priority and it will be picked |
| 820 | * up later. |
| 821 | */ |
| 822 | KKASSERT(opq == npq || (p->p_flag & P_ONRUNQ) == 0); |
| 823 | p->p_priority = newpriority; |
| 824 | } |
| 825 | crit_exit(); |
| 826 | maybe_resched(p); |
| 827 | } |
| 828 | |
| 829 | /* |
| 830 | * Compute a tenex style load average of a quantity on |
| 831 | * 1, 5 and 15 minute intervals. |
| 832 | */ |
| 833 | static void |
| 834 | loadav(void *arg) |
| 835 | { |
| 836 | int i, nrun; |
| 837 | struct loadavg *avg; |
| 838 | struct proc *p; |
| 839 | |
| 840 | avg = &averunnable; |
| 841 | nrun = 0; |
| 842 | FOREACH_PROC_IN_SYSTEM(p) { |
| 843 | switch (p->p_stat) { |
| 844 | case SRUN: |
| 845 | case SIDL: |
| 846 | nrun++; |
| 847 | } |
| 848 | } |
| 849 | for (i = 0; i < 3; i++) |
| 850 | avg->ldavg[i] = (cexp[i] * avg->ldavg[i] + |
| 851 | nrun * FSCALE * (FSCALE - cexp[i])) >> FSHIFT; |
| 852 | |
| 853 | /* |
| 854 | * Schedule the next update to occur after 5 seconds, but add a |
| 855 | * random variation to avoid synchronisation with processes that |
| 856 | * run at regular intervals. |
| 857 | */ |
| 858 | callout_reset(&loadav_callout, hz * 4 + (int)(random() % (hz * 2 + 1)), |
| 859 | loadav, NULL); |
| 860 | } |
| 861 | |
| 862 | /* ARGSUSED */ |
| 863 | static void |
| 864 | sched_setup(dummy) |
| 865 | void *dummy; |
| 866 | { |
| 867 | |
| 868 | callout_init(&loadav_callout); |
| 869 | |
| 870 | /* Kick off timeout driven events by calling first time. */ |
| 871 | roundrobin(NULL); |
| 872 | schedcpu(NULL); |
| 873 | loadav(NULL); |
| 874 | } |
| 875 | |
| 876 | /* |
| 877 | * We adjust the priority of the current process. The priority of |
| 878 | * a process gets worse as it accumulates CPU time. The cpu usage |
| 879 | * estimator (p_estcpu) is increased here. resetpriority() will |
| 880 | * compute a different priority each time p_estcpu increases by |
| 881 | * INVERSE_ESTCPU_WEIGHT |
| 882 | * (until MAXPRI is reached). The cpu usage estimator ramps up |
| 883 | * quite quickly when the process is running (linearly), and decays |
| 884 | * away exponentially, at a rate which is proportionally slower when |
| 885 | * the system is busy. The basic principle is that the system will |
| 886 | * 90% forget that the process used a lot of CPU time in 5 * loadav |
| 887 | * seconds. This causes the system to favor processes which haven't |
| 888 | * run much recently, and to round-robin among other processes. |
| 889 | */ |
| 890 | void |
| 891 | schedclock(p) |
| 892 | struct proc *p; |
| 893 | { |
| 894 | |
| 895 | p->p_cpticks++; |
| 896 | p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); |
| 897 | if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) |
| 898 | resetpriority(p); |
| 899 | } |
| 900 | |
| 901 | static |
| 902 | void |
| 903 | crit_panicints(void) |
| 904 | { |
| 905 | int s; |
| 906 | int cpri; |
| 907 | |
| 908 | s = splhigh(); |
| 909 | cpri = crit_panic_save(); |
| 910 | splx(safepri); |
| 911 | crit_panic_restore(cpri); |
| 912 | splx(s); |
| 913 | } |
| 914 | |