17c478bd9Sstevel@tonic-gate /* 27c478bd9Sstevel@tonic-gate * CDDL HEADER START 37c478bd9Sstevel@tonic-gate * 47c478bd9Sstevel@tonic-gate * The contents of this file are subject to the terms of the 5ab761399Sesaxe * Common Development and Distribution License (the "License"). 6ab761399Sesaxe * You may not use this file except in compliance with the License. 77c478bd9Sstevel@tonic-gate * 87c478bd9Sstevel@tonic-gate * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 97c478bd9Sstevel@tonic-gate * or http://www.opensolaris.org/os/licensing. 107c478bd9Sstevel@tonic-gate * See the License for the specific language governing permissions 117c478bd9Sstevel@tonic-gate * and limitations under the License. 127c478bd9Sstevel@tonic-gate * 137c478bd9Sstevel@tonic-gate * When distributing Covered Code, include this CDDL HEADER in each 147c478bd9Sstevel@tonic-gate * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 157c478bd9Sstevel@tonic-gate * If applicable, add the following below this CDDL HEADER, with the 167c478bd9Sstevel@tonic-gate * fields enclosed by brackets "[]" replaced with your own identifying 177c478bd9Sstevel@tonic-gate * information: Portions Copyright [yyyy] [name of copyright owner] 187c478bd9Sstevel@tonic-gate * 197c478bd9Sstevel@tonic-gate * CDDL HEADER END 207c478bd9Sstevel@tonic-gate */ 217c478bd9Sstevel@tonic-gate /* 22*f2bd4627Sjohansen * Copyright 2007 Sun Microsystems, Inc. All rights reserved. 237c478bd9Sstevel@tonic-gate * Use is subject to license terms. 247c478bd9Sstevel@tonic-gate */ 257c478bd9Sstevel@tonic-gate 267c478bd9Sstevel@tonic-gate /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */ 277c478bd9Sstevel@tonic-gate /* All Rights Reserved */ 287c478bd9Sstevel@tonic-gate 297c478bd9Sstevel@tonic-gate 307c478bd9Sstevel@tonic-gate #pragma ident "%Z%%M% %I% %E% SMI" /* from SVr4.0 1.30 */ 317c478bd9Sstevel@tonic-gate 327c478bd9Sstevel@tonic-gate #include <sys/types.h> 337c478bd9Sstevel@tonic-gate #include <sys/param.h> 347c478bd9Sstevel@tonic-gate #include <sys/sysmacros.h> 357c478bd9Sstevel@tonic-gate #include <sys/signal.h> 367c478bd9Sstevel@tonic-gate #include <sys/user.h> 377c478bd9Sstevel@tonic-gate #include <sys/systm.h> 387c478bd9Sstevel@tonic-gate #include <sys/sysinfo.h> 397c478bd9Sstevel@tonic-gate #include <sys/var.h> 407c478bd9Sstevel@tonic-gate #include <sys/errno.h> 417c478bd9Sstevel@tonic-gate #include <sys/cmn_err.h> 427c478bd9Sstevel@tonic-gate #include <sys/debug.h> 437c478bd9Sstevel@tonic-gate #include <sys/inline.h> 447c478bd9Sstevel@tonic-gate #include <sys/disp.h> 457c478bd9Sstevel@tonic-gate #include <sys/class.h> 467c478bd9Sstevel@tonic-gate #include <sys/bitmap.h> 477c478bd9Sstevel@tonic-gate #include <sys/kmem.h> 487c478bd9Sstevel@tonic-gate #include <sys/cpuvar.h> 497c478bd9Sstevel@tonic-gate #include <sys/vtrace.h> 507c478bd9Sstevel@tonic-gate #include <sys/tnf.h> 517c478bd9Sstevel@tonic-gate #include <sys/cpupart.h> 527c478bd9Sstevel@tonic-gate #include <sys/lgrp.h> 537c478bd9Sstevel@tonic-gate #include <sys/chip.h> 547c478bd9Sstevel@tonic-gate #include <sys/schedctl.h> 557c478bd9Sstevel@tonic-gate #include <sys/atomic.h> 567c478bd9Sstevel@tonic-gate #include <sys/dtrace.h> 577c478bd9Sstevel@tonic-gate #include <sys/sdt.h> 587c478bd9Sstevel@tonic-gate 597c478bd9Sstevel@tonic-gate #include <vm/as.h> 607c478bd9Sstevel@tonic-gate 617c478bd9Sstevel@tonic-gate #define BOUND_CPU 0x1 627c478bd9Sstevel@tonic-gate #define BOUND_PARTITION 0x2 637c478bd9Sstevel@tonic-gate #define BOUND_INTR 0x4 647c478bd9Sstevel@tonic-gate 657c478bd9Sstevel@tonic-gate /* Dispatch queue allocation structure and functions */ 667c478bd9Sstevel@tonic-gate struct disp_queue_info { 677c478bd9Sstevel@tonic-gate disp_t *dp; 687c478bd9Sstevel@tonic-gate dispq_t *olddispq; 697c478bd9Sstevel@tonic-gate dispq_t *newdispq; 707c478bd9Sstevel@tonic-gate ulong_t *olddqactmap; 717c478bd9Sstevel@tonic-gate ulong_t *newdqactmap; 727c478bd9Sstevel@tonic-gate int oldnglobpris; 737c478bd9Sstevel@tonic-gate }; 747c478bd9Sstevel@tonic-gate static void disp_dq_alloc(struct disp_queue_info *dptr, int numpris, 757c478bd9Sstevel@tonic-gate disp_t *dp); 767c478bd9Sstevel@tonic-gate static void disp_dq_assign(struct disp_queue_info *dptr, int numpris); 777c478bd9Sstevel@tonic-gate static void disp_dq_free(struct disp_queue_info *dptr); 787c478bd9Sstevel@tonic-gate 797c478bd9Sstevel@tonic-gate /* platform-specific routine to call when processor is idle */ 807c478bd9Sstevel@tonic-gate static void generic_idle_cpu(); 817c478bd9Sstevel@tonic-gate void (*idle_cpu)() = generic_idle_cpu; 827c478bd9Sstevel@tonic-gate 837c478bd9Sstevel@tonic-gate /* routines invoked when a CPU enters/exits the idle loop */ 847c478bd9Sstevel@tonic-gate static void idle_enter(); 857c478bd9Sstevel@tonic-gate static void idle_exit(); 867c478bd9Sstevel@tonic-gate 877c478bd9Sstevel@tonic-gate /* platform-specific routine to call when thread is enqueued */ 887c478bd9Sstevel@tonic-gate static void generic_enq_thread(cpu_t *, int); 897c478bd9Sstevel@tonic-gate void (*disp_enq_thread)(cpu_t *, int) = generic_enq_thread; 907c478bd9Sstevel@tonic-gate 91685679f7Sakolb pri_t kpreemptpri; /* priority where kernel preemption applies */ 92685679f7Sakolb pri_t upreemptpri = 0; /* priority where normal preemption applies */ 93685679f7Sakolb pri_t intr_pri; /* interrupt thread priority base level */ 947c478bd9Sstevel@tonic-gate 95685679f7Sakolb #define KPQPRI -1 /* pri where cpu affinity is dropped for kpq */ 96685679f7Sakolb pri_t kpqpri = KPQPRI; /* can be set in /etc/system */ 97685679f7Sakolb disp_t cpu0_disp; /* boot CPU's dispatch queue */ 987c478bd9Sstevel@tonic-gate disp_lock_t swapped_lock; /* lock swapped threads and swap queue */ 99685679f7Sakolb int nswapped; /* total number of swapped threads */ 1007c478bd9Sstevel@tonic-gate void disp_swapped_enq(kthread_t *tp); 1017c478bd9Sstevel@tonic-gate static void disp_swapped_setrun(kthread_t *tp); 1027c478bd9Sstevel@tonic-gate static void cpu_resched(cpu_t *cp, pri_t tpri); 1037c478bd9Sstevel@tonic-gate 1047c478bd9Sstevel@tonic-gate /* 1057c478bd9Sstevel@tonic-gate * If this is set, only interrupt threads will cause kernel preemptions. 1067c478bd9Sstevel@tonic-gate * This is done by changing the value of kpreemptpri. kpreemptpri 1077c478bd9Sstevel@tonic-gate * will either be the max sysclass pri + 1 or the min interrupt pri. 1087c478bd9Sstevel@tonic-gate */ 1097c478bd9Sstevel@tonic-gate int only_intr_kpreempt; 1107c478bd9Sstevel@tonic-gate 1117c478bd9Sstevel@tonic-gate extern void set_idle_cpu(int cpun); 1127c478bd9Sstevel@tonic-gate extern void unset_idle_cpu(int cpun); 1137c478bd9Sstevel@tonic-gate static void setkpdq(kthread_t *tp, int borf); 1147c478bd9Sstevel@tonic-gate #define SETKP_BACK 0 1157c478bd9Sstevel@tonic-gate #define SETKP_FRONT 1 1167c478bd9Sstevel@tonic-gate /* 1177c478bd9Sstevel@tonic-gate * Parameter that determines how recently a thread must have run 1187c478bd9Sstevel@tonic-gate * on the CPU to be considered loosely-bound to that CPU to reduce 1197c478bd9Sstevel@tonic-gate * cold cache effects. The interval is in hertz. 1207c478bd9Sstevel@tonic-gate * 1217c478bd9Sstevel@tonic-gate * The platform may define a per physical processor adjustment of 1227c478bd9Sstevel@tonic-gate * this parameter. For efficiency, the effective rechoose interval 1237c478bd9Sstevel@tonic-gate * (rechoose_interval + per chip adjustment) is maintained in the 1247c478bd9Sstevel@tonic-gate * cpu structures. See cpu_choose() 1257c478bd9Sstevel@tonic-gate */ 1267c478bd9Sstevel@tonic-gate int rechoose_interval = RECHOOSE_INTERVAL; 1277c478bd9Sstevel@tonic-gate static cpu_t *cpu_choose(kthread_t *, pri_t); 1287c478bd9Sstevel@tonic-gate 129685679f7Sakolb /* 130685679f7Sakolb * Parameter that determines how long (in nanoseconds) a thread must 131685679f7Sakolb * be sitting on a run queue before it can be stolen by another CPU 132685679f7Sakolb * to reduce migrations. The interval is in nanoseconds. 133685679f7Sakolb * 134685679f7Sakolb * The nosteal_nsec should be set by a platform code to an appropriate value. 135685679f7Sakolb * 136685679f7Sakolb */ 137685679f7Sakolb hrtime_t nosteal_nsec = 0; 138685679f7Sakolb 139685679f7Sakolb /* 140685679f7Sakolb * Value of nosteal_nsec meaning that nosteal optimization should be disabled 141685679f7Sakolb */ 142685679f7Sakolb #define NOSTEAL_DISABLED 1 143685679f7Sakolb 1447c478bd9Sstevel@tonic-gate id_t defaultcid; /* system "default" class; see dispadmin(1M) */ 1457c478bd9Sstevel@tonic-gate 1467c478bd9Sstevel@tonic-gate disp_lock_t transition_lock; /* lock on transitioning threads */ 1477c478bd9Sstevel@tonic-gate disp_lock_t stop_lock; /* lock on stopped threads */ 1487c478bd9Sstevel@tonic-gate 149685679f7Sakolb static void cpu_dispqalloc(int numpris); 150685679f7Sakolb 151685679f7Sakolb /* 152685679f7Sakolb * This gets returned by disp_getwork/disp_getbest if we couldn't steal 153685679f7Sakolb * a thread because it was sitting on its run queue for a very short 154685679f7Sakolb * period of time. 155685679f7Sakolb */ 156685679f7Sakolb #define T_DONTSTEAL (kthread_t *)(-1) /* returned by disp_getwork/getbest */ 1577c478bd9Sstevel@tonic-gate 1587c478bd9Sstevel@tonic-gate static kthread_t *disp_getwork(cpu_t *to); 1597c478bd9Sstevel@tonic-gate static kthread_t *disp_getbest(disp_t *from); 1607c478bd9Sstevel@tonic-gate static kthread_t *disp_ratify(kthread_t *tp, disp_t *kpq); 1617c478bd9Sstevel@tonic-gate 1627c478bd9Sstevel@tonic-gate void swtch_to(kthread_t *); 1637c478bd9Sstevel@tonic-gate 1647c478bd9Sstevel@tonic-gate /* 1657c478bd9Sstevel@tonic-gate * dispatcher and scheduler initialization 1667c478bd9Sstevel@tonic-gate */ 1677c478bd9Sstevel@tonic-gate 1687c478bd9Sstevel@tonic-gate /* 1697c478bd9Sstevel@tonic-gate * disp_setup - Common code to calculate and allocate dispatcher 1707c478bd9Sstevel@tonic-gate * variables and structures based on the maximum priority. 1717c478bd9Sstevel@tonic-gate */ 1727c478bd9Sstevel@tonic-gate static void 1737c478bd9Sstevel@tonic-gate disp_setup(pri_t maxglobpri, pri_t oldnglobpris) 1747c478bd9Sstevel@tonic-gate { 1757c478bd9Sstevel@tonic-gate pri_t newnglobpris; 1767c478bd9Sstevel@tonic-gate 1777c478bd9Sstevel@tonic-gate ASSERT(MUTEX_HELD(&cpu_lock)); 1787c478bd9Sstevel@tonic-gate 1797c478bd9Sstevel@tonic-gate newnglobpris = maxglobpri + 1 + LOCK_LEVEL; 1807c478bd9Sstevel@tonic-gate 1817c478bd9Sstevel@tonic-gate if (newnglobpris > oldnglobpris) { 1827c478bd9Sstevel@tonic-gate /* 1837c478bd9Sstevel@tonic-gate * Allocate new kp queues for each CPU partition. 1847c478bd9Sstevel@tonic-gate */ 1857c478bd9Sstevel@tonic-gate cpupart_kpqalloc(newnglobpris); 1867c478bd9Sstevel@tonic-gate 1877c478bd9Sstevel@tonic-gate /* 1887c478bd9Sstevel@tonic-gate * Allocate new dispatch queues for each CPU. 1897c478bd9Sstevel@tonic-gate */ 1907c478bd9Sstevel@tonic-gate cpu_dispqalloc(newnglobpris); 1917c478bd9Sstevel@tonic-gate 1927c478bd9Sstevel@tonic-gate /* 1937c478bd9Sstevel@tonic-gate * compute new interrupt thread base priority 1947c478bd9Sstevel@tonic-gate */ 1957c478bd9Sstevel@tonic-gate intr_pri = maxglobpri; 1967c478bd9Sstevel@tonic-gate if (only_intr_kpreempt) { 1977c478bd9Sstevel@tonic-gate kpreemptpri = intr_pri + 1; 1987c478bd9Sstevel@tonic-gate if (kpqpri == KPQPRI) 1997c478bd9Sstevel@tonic-gate kpqpri = kpreemptpri; 2007c478bd9Sstevel@tonic-gate } 2017c478bd9Sstevel@tonic-gate v.v_nglobpris = newnglobpris; 2027c478bd9Sstevel@tonic-gate } 2037c478bd9Sstevel@tonic-gate } 2047c478bd9Sstevel@tonic-gate 2057c478bd9Sstevel@tonic-gate /* 2067c478bd9Sstevel@tonic-gate * dispinit - Called to initialize all loaded classes and the 2077c478bd9Sstevel@tonic-gate * dispatcher framework. 2087c478bd9Sstevel@tonic-gate */ 2097c478bd9Sstevel@tonic-gate void 2107c478bd9Sstevel@tonic-gate dispinit(void) 2117c478bd9Sstevel@tonic-gate { 2127c478bd9Sstevel@tonic-gate id_t cid; 2137c478bd9Sstevel@tonic-gate pri_t maxglobpri; 2147c478bd9Sstevel@tonic-gate pri_t cl_maxglobpri; 2157c478bd9Sstevel@tonic-gate 2167c478bd9Sstevel@tonic-gate maxglobpri = -1; 2177c478bd9Sstevel@tonic-gate 2187c478bd9Sstevel@tonic-gate /* 2197c478bd9Sstevel@tonic-gate * Initialize transition lock, which will always be set. 2207c478bd9Sstevel@tonic-gate */ 2217c478bd9Sstevel@tonic-gate DISP_LOCK_INIT(&transition_lock); 2227c478bd9Sstevel@tonic-gate disp_lock_enter_high(&transition_lock); 2237c478bd9Sstevel@tonic-gate DISP_LOCK_INIT(&stop_lock); 2247c478bd9Sstevel@tonic-gate 2257c478bd9Sstevel@tonic-gate mutex_enter(&cpu_lock); 2267c478bd9Sstevel@tonic-gate CPU->cpu_disp->disp_maxrunpri = -1; 2277c478bd9Sstevel@tonic-gate CPU->cpu_disp->disp_max_unbound_pri = -1; 2287c478bd9Sstevel@tonic-gate /* 2297c478bd9Sstevel@tonic-gate * Initialize the default CPU partition. 2307c478bd9Sstevel@tonic-gate */ 2317c478bd9Sstevel@tonic-gate cpupart_initialize_default(); 2327c478bd9Sstevel@tonic-gate /* 2337c478bd9Sstevel@tonic-gate * Call the class specific initialization functions for 2347c478bd9Sstevel@tonic-gate * all pre-installed schedulers. 2357c478bd9Sstevel@tonic-gate * 2367c478bd9Sstevel@tonic-gate * We pass the size of a class specific parameter 2377c478bd9Sstevel@tonic-gate * buffer to each of the initialization functions 2387c478bd9Sstevel@tonic-gate * to try to catch problems with backward compatibility 2397c478bd9Sstevel@tonic-gate * of class modules. 2407c478bd9Sstevel@tonic-gate * 2417c478bd9Sstevel@tonic-gate * For example a new class module running on an old system 2427c478bd9Sstevel@tonic-gate * which didn't provide sufficiently large parameter buffers 2437c478bd9Sstevel@tonic-gate * would be bad news. Class initialization modules can check for 2447c478bd9Sstevel@tonic-gate * this and take action if they detect a problem. 2457c478bd9Sstevel@tonic-gate */ 2467c478bd9Sstevel@tonic-gate 2477c478bd9Sstevel@tonic-gate for (cid = 0; cid < nclass; cid++) { 2487c478bd9Sstevel@tonic-gate sclass_t *sc; 2497c478bd9Sstevel@tonic-gate 2507c478bd9Sstevel@tonic-gate sc = &sclass[cid]; 2517c478bd9Sstevel@tonic-gate if (SCHED_INSTALLED(sc)) { 2527c478bd9Sstevel@tonic-gate cl_maxglobpri = sc->cl_init(cid, PC_CLPARMSZ, 2537c478bd9Sstevel@tonic-gate &sc->cl_funcs); 2547c478bd9Sstevel@tonic-gate if (cl_maxglobpri > maxglobpri) 2557c478bd9Sstevel@tonic-gate maxglobpri = cl_maxglobpri; 2567c478bd9Sstevel@tonic-gate } 2577c478bd9Sstevel@tonic-gate } 2587c478bd9Sstevel@tonic-gate kpreemptpri = (pri_t)v.v_maxsyspri + 1; 2597c478bd9Sstevel@tonic-gate if (kpqpri == KPQPRI) 2607c478bd9Sstevel@tonic-gate kpqpri = kpreemptpri; 2617c478bd9Sstevel@tonic-gate 2627c478bd9Sstevel@tonic-gate ASSERT(maxglobpri >= 0); 2637c478bd9Sstevel@tonic-gate disp_setup(maxglobpri, 0); 2647c478bd9Sstevel@tonic-gate 2657c478bd9Sstevel@tonic-gate mutex_exit(&cpu_lock); 2667c478bd9Sstevel@tonic-gate 2677c478bd9Sstevel@tonic-gate /* 2687c478bd9Sstevel@tonic-gate * Get the default class ID; this may be later modified via 2697c478bd9Sstevel@tonic-gate * dispadmin(1M). This will load the class (normally TS) and that will 2707c478bd9Sstevel@tonic-gate * call disp_add(), which is why we had to drop cpu_lock first. 2717c478bd9Sstevel@tonic-gate */ 2727c478bd9Sstevel@tonic-gate if (getcid(defaultclass, &defaultcid) != 0) { 2737c478bd9Sstevel@tonic-gate cmn_err(CE_PANIC, "Couldn't load default scheduling class '%s'", 2747c478bd9Sstevel@tonic-gate defaultclass); 2757c478bd9Sstevel@tonic-gate } 2767c478bd9Sstevel@tonic-gate } 2777c478bd9Sstevel@tonic-gate 2787c478bd9Sstevel@tonic-gate /* 2797c478bd9Sstevel@tonic-gate * disp_add - Called with class pointer to initialize the dispatcher 2807c478bd9Sstevel@tonic-gate * for a newly loaded class. 2817c478bd9Sstevel@tonic-gate */ 2827c478bd9Sstevel@tonic-gate void 2837c478bd9Sstevel@tonic-gate disp_add(sclass_t *clp) 2847c478bd9Sstevel@tonic-gate { 2857c478bd9Sstevel@tonic-gate pri_t maxglobpri; 2867c478bd9Sstevel@tonic-gate pri_t cl_maxglobpri; 2877c478bd9Sstevel@tonic-gate 2887c478bd9Sstevel@tonic-gate mutex_enter(&cpu_lock); 2897c478bd9Sstevel@tonic-gate /* 2907c478bd9Sstevel@tonic-gate * Initialize the scheduler class. 2917c478bd9Sstevel@tonic-gate */ 2927c478bd9Sstevel@tonic-gate maxglobpri = (pri_t)(v.v_nglobpris - LOCK_LEVEL - 1); 2937c478bd9Sstevel@tonic-gate cl_maxglobpri = clp->cl_init(clp - sclass, PC_CLPARMSZ, &clp->cl_funcs); 2947c478bd9Sstevel@tonic-gate if (cl_maxglobpri > maxglobpri) 2957c478bd9Sstevel@tonic-gate maxglobpri = cl_maxglobpri; 2967c478bd9Sstevel@tonic-gate 2977c478bd9Sstevel@tonic-gate /* 2987c478bd9Sstevel@tonic-gate * Save old queue information. Since we're initializing a 2997c478bd9Sstevel@tonic-gate * new scheduling class which has just been loaded, then 3007c478bd9Sstevel@tonic-gate * the size of the dispq may have changed. We need to handle 3017c478bd9Sstevel@tonic-gate * that here. 3027c478bd9Sstevel@tonic-gate */ 3037c478bd9Sstevel@tonic-gate disp_setup(maxglobpri, v.v_nglobpris); 3047c478bd9Sstevel@tonic-gate 3057c478bd9Sstevel@tonic-gate mutex_exit(&cpu_lock); 3067c478bd9Sstevel@tonic-gate } 3077c478bd9Sstevel@tonic-gate 3087c478bd9Sstevel@tonic-gate 3097c478bd9Sstevel@tonic-gate /* 3107c478bd9Sstevel@tonic-gate * For each CPU, allocate new dispatch queues 3117c478bd9Sstevel@tonic-gate * with the stated number of priorities. 3127c478bd9Sstevel@tonic-gate */ 3137c478bd9Sstevel@tonic-gate static void 3147c478bd9Sstevel@tonic-gate cpu_dispqalloc(int numpris) 3157c478bd9Sstevel@tonic-gate { 3167c478bd9Sstevel@tonic-gate cpu_t *cpup; 3177c478bd9Sstevel@tonic-gate struct disp_queue_info *disp_mem; 3187c478bd9Sstevel@tonic-gate int i, num; 3197c478bd9Sstevel@tonic-gate 3207c478bd9Sstevel@tonic-gate ASSERT(MUTEX_HELD(&cpu_lock)); 3217c478bd9Sstevel@tonic-gate 3227c478bd9Sstevel@tonic-gate disp_mem = kmem_zalloc(NCPU * 3237c478bd9Sstevel@tonic-gate sizeof (struct disp_queue_info), KM_SLEEP); 3247c478bd9Sstevel@tonic-gate 3257c478bd9Sstevel@tonic-gate /* 3267c478bd9Sstevel@tonic-gate * This routine must allocate all of the memory before stopping 3277c478bd9Sstevel@tonic-gate * the cpus because it must not sleep in kmem_alloc while the 3287c478bd9Sstevel@tonic-gate * CPUs are stopped. Locks they hold will not be freed until they 3297c478bd9Sstevel@tonic-gate * are restarted. 3307c478bd9Sstevel@tonic-gate */ 3317c478bd9Sstevel@tonic-gate i = 0; 3327c478bd9Sstevel@tonic-gate cpup = cpu_list; 3337c478bd9Sstevel@tonic-gate do { 3347c478bd9Sstevel@tonic-gate disp_dq_alloc(&disp_mem[i], numpris, cpup->cpu_disp); 3357c478bd9Sstevel@tonic-gate i++; 3367c478bd9Sstevel@tonic-gate cpup = cpup->cpu_next; 3377c478bd9Sstevel@tonic-gate } while (cpup != cpu_list); 3387c478bd9Sstevel@tonic-gate num = i; 3397c478bd9Sstevel@tonic-gate 3407c478bd9Sstevel@tonic-gate pause_cpus(NULL); 3417c478bd9Sstevel@tonic-gate for (i = 0; i < num; i++) 3427c478bd9Sstevel@tonic-gate disp_dq_assign(&disp_mem[i], numpris); 3437c478bd9Sstevel@tonic-gate start_cpus(); 3447c478bd9Sstevel@tonic-gate 3457c478bd9Sstevel@tonic-gate /* 3467c478bd9Sstevel@tonic-gate * I must free all of the memory after starting the cpus because 3477c478bd9Sstevel@tonic-gate * I can not risk sleeping in kmem_free while the cpus are stopped. 3487c478bd9Sstevel@tonic-gate */ 3497c478bd9Sstevel@tonic-gate for (i = 0; i < num; i++) 3507c478bd9Sstevel@tonic-gate disp_dq_free(&disp_mem[i]); 3517c478bd9Sstevel@tonic-gate 3527c478bd9Sstevel@tonic-gate kmem_free(disp_mem, NCPU * sizeof (struct disp_queue_info)); 3537c478bd9Sstevel@tonic-gate } 3547c478bd9Sstevel@tonic-gate 3557c478bd9Sstevel@tonic-gate static void 3567c478bd9Sstevel@tonic-gate disp_dq_alloc(struct disp_queue_info *dptr, int numpris, disp_t *dp) 3577c478bd9Sstevel@tonic-gate { 3587c478bd9Sstevel@tonic-gate dptr->newdispq = kmem_zalloc(numpris * sizeof (dispq_t), KM_SLEEP); 3597c478bd9Sstevel@tonic-gate dptr->newdqactmap = kmem_zalloc(((numpris / BT_NBIPUL) + 1) * 3607c478bd9Sstevel@tonic-gate sizeof (long), KM_SLEEP); 3617c478bd9Sstevel@tonic-gate dptr->dp = dp; 3627c478bd9Sstevel@tonic-gate } 3637c478bd9Sstevel@tonic-gate 3647c478bd9Sstevel@tonic-gate static void 3657c478bd9Sstevel@tonic-gate disp_dq_assign(struct disp_queue_info *dptr, int numpris) 3667c478bd9Sstevel@tonic-gate { 3677c478bd9Sstevel@tonic-gate disp_t *dp; 3687c478bd9Sstevel@tonic-gate 3697c478bd9Sstevel@tonic-gate dp = dptr->dp; 3707c478bd9Sstevel@tonic-gate dptr->olddispq = dp->disp_q; 3717c478bd9Sstevel@tonic-gate dptr->olddqactmap = dp->disp_qactmap; 3727c478bd9Sstevel@tonic-gate dptr->oldnglobpris = dp->disp_npri; 3737c478bd9Sstevel@tonic-gate 3747c478bd9Sstevel@tonic-gate ASSERT(dptr->oldnglobpris < numpris); 3757c478bd9Sstevel@tonic-gate 3767c478bd9Sstevel@tonic-gate if (dptr->olddispq != NULL) { 3777c478bd9Sstevel@tonic-gate /* 3787c478bd9Sstevel@tonic-gate * Use kcopy because bcopy is platform-specific 3797c478bd9Sstevel@tonic-gate * and could block while we might have paused the cpus. 3807c478bd9Sstevel@tonic-gate */ 3817c478bd9Sstevel@tonic-gate (void) kcopy(dptr->olddispq, dptr->newdispq, 3827c478bd9Sstevel@tonic-gate dptr->oldnglobpris * sizeof (dispq_t)); 3837c478bd9Sstevel@tonic-gate (void) kcopy(dptr->olddqactmap, dptr->newdqactmap, 3847c478bd9Sstevel@tonic-gate ((dptr->oldnglobpris / BT_NBIPUL) + 1) * 3857c478bd9Sstevel@tonic-gate sizeof (long)); 3867c478bd9Sstevel@tonic-gate } 3877c478bd9Sstevel@tonic-gate dp->disp_q = dptr->newdispq; 3887c478bd9Sstevel@tonic-gate dp->disp_qactmap = dptr->newdqactmap; 3897c478bd9Sstevel@tonic-gate dp->disp_q_limit = &dptr->newdispq[numpris]; 3907c478bd9Sstevel@tonic-gate dp->disp_npri = numpris; 3917c478bd9Sstevel@tonic-gate } 3927c478bd9Sstevel@tonic-gate 3937c478bd9Sstevel@tonic-gate static void 3947c478bd9Sstevel@tonic-gate disp_dq_free(struct disp_queue_info *dptr) 3957c478bd9Sstevel@tonic-gate { 3967c478bd9Sstevel@tonic-gate if (dptr->olddispq != NULL) 3977c478bd9Sstevel@tonic-gate kmem_free(dptr->olddispq, 3987c478bd9Sstevel@tonic-gate dptr->oldnglobpris * sizeof (dispq_t)); 3997c478bd9Sstevel@tonic-gate if (dptr->olddqactmap != NULL) 4007c478bd9Sstevel@tonic-gate kmem_free(dptr->olddqactmap, 4017c478bd9Sstevel@tonic-gate ((dptr->oldnglobpris / BT_NBIPUL) + 1) * sizeof (long)); 4027c478bd9Sstevel@tonic-gate } 4037c478bd9Sstevel@tonic-gate 4047c478bd9Sstevel@tonic-gate /* 4057c478bd9Sstevel@tonic-gate * For a newly created CPU, initialize the dispatch queue. 4067c478bd9Sstevel@tonic-gate * This is called before the CPU is known through cpu[] or on any lists. 4077c478bd9Sstevel@tonic-gate */ 4087c478bd9Sstevel@tonic-gate void 4097c478bd9Sstevel@tonic-gate disp_cpu_init(cpu_t *cp) 4107c478bd9Sstevel@tonic-gate { 4117c478bd9Sstevel@tonic-gate disp_t *dp; 4127c478bd9Sstevel@tonic-gate dispq_t *newdispq; 4137c478bd9Sstevel@tonic-gate ulong_t *newdqactmap; 4147c478bd9Sstevel@tonic-gate 4157c478bd9Sstevel@tonic-gate ASSERT(MUTEX_HELD(&cpu_lock)); /* protect dispatcher queue sizes */ 4167c478bd9Sstevel@tonic-gate 4177c478bd9Sstevel@tonic-gate if (cp == cpu0_disp.disp_cpu) 4187c478bd9Sstevel@tonic-gate dp = &cpu0_disp; 4197c478bd9Sstevel@tonic-gate else 4207c478bd9Sstevel@tonic-gate dp = kmem_alloc(sizeof (disp_t), KM_SLEEP); 4217c478bd9Sstevel@tonic-gate bzero(dp, sizeof (disp_t)); 4227c478bd9Sstevel@tonic-gate cp->cpu_disp = dp; 4237c478bd9Sstevel@tonic-gate dp->disp_cpu = cp; 4247c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = -1; 4257c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = -1; 4267c478bd9Sstevel@tonic-gate DISP_LOCK_INIT(&cp->cpu_thread_lock); 4277c478bd9Sstevel@tonic-gate /* 4287c478bd9Sstevel@tonic-gate * Allocate memory for the dispatcher queue headers 4297c478bd9Sstevel@tonic-gate * and the active queue bitmap. 4307c478bd9Sstevel@tonic-gate */ 4317c478bd9Sstevel@tonic-gate newdispq = kmem_zalloc(v.v_nglobpris * sizeof (dispq_t), KM_SLEEP); 4327c478bd9Sstevel@tonic-gate newdqactmap = kmem_zalloc(((v.v_nglobpris / BT_NBIPUL) + 1) * 4337c478bd9Sstevel@tonic-gate sizeof (long), KM_SLEEP); 4347c478bd9Sstevel@tonic-gate dp->disp_q = newdispq; 4357c478bd9Sstevel@tonic-gate dp->disp_qactmap = newdqactmap; 4367c478bd9Sstevel@tonic-gate dp->disp_q_limit = &newdispq[v.v_nglobpris]; 4377c478bd9Sstevel@tonic-gate dp->disp_npri = v.v_nglobpris; 4387c478bd9Sstevel@tonic-gate } 4397c478bd9Sstevel@tonic-gate 4407c478bd9Sstevel@tonic-gate void 4417c478bd9Sstevel@tonic-gate disp_cpu_fini(cpu_t *cp) 4427c478bd9Sstevel@tonic-gate { 4437c478bd9Sstevel@tonic-gate ASSERT(MUTEX_HELD(&cpu_lock)); 4447c478bd9Sstevel@tonic-gate 4457c478bd9Sstevel@tonic-gate disp_kp_free(cp->cpu_disp); 4467c478bd9Sstevel@tonic-gate if (cp->cpu_disp != &cpu0_disp) 4477c478bd9Sstevel@tonic-gate kmem_free(cp->cpu_disp, sizeof (disp_t)); 4487c478bd9Sstevel@tonic-gate } 4497c478bd9Sstevel@tonic-gate 4507c478bd9Sstevel@tonic-gate /* 4517c478bd9Sstevel@tonic-gate * Allocate new, larger kpreempt dispatch queue to replace the old one. 4527c478bd9Sstevel@tonic-gate */ 4537c478bd9Sstevel@tonic-gate void 4547c478bd9Sstevel@tonic-gate disp_kp_alloc(disp_t *dq, pri_t npri) 4557c478bd9Sstevel@tonic-gate { 4567c478bd9Sstevel@tonic-gate struct disp_queue_info mem_info; 4577c478bd9Sstevel@tonic-gate 4587c478bd9Sstevel@tonic-gate if (npri > dq->disp_npri) { 4597c478bd9Sstevel@tonic-gate /* 4607c478bd9Sstevel@tonic-gate * Allocate memory for the new array. 4617c478bd9Sstevel@tonic-gate */ 4627c478bd9Sstevel@tonic-gate disp_dq_alloc(&mem_info, npri, dq); 4637c478bd9Sstevel@tonic-gate 4647c478bd9Sstevel@tonic-gate /* 4657c478bd9Sstevel@tonic-gate * We need to copy the old structures to the new 4667c478bd9Sstevel@tonic-gate * and free the old. 4677c478bd9Sstevel@tonic-gate */ 4687c478bd9Sstevel@tonic-gate disp_dq_assign(&mem_info, npri); 4697c478bd9Sstevel@tonic-gate disp_dq_free(&mem_info); 4707c478bd9Sstevel@tonic-gate } 4717c478bd9Sstevel@tonic-gate } 4727c478bd9Sstevel@tonic-gate 4737c478bd9Sstevel@tonic-gate /* 4747c478bd9Sstevel@tonic-gate * Free dispatch queue. 4757c478bd9Sstevel@tonic-gate * Used for the kpreempt queues for a removed CPU partition and 4767c478bd9Sstevel@tonic-gate * for the per-CPU queues of deleted CPUs. 4777c478bd9Sstevel@tonic-gate */ 4787c478bd9Sstevel@tonic-gate void 4797c478bd9Sstevel@tonic-gate disp_kp_free(disp_t *dq) 4807c478bd9Sstevel@tonic-gate { 4817c478bd9Sstevel@tonic-gate struct disp_queue_info mem_info; 4827c478bd9Sstevel@tonic-gate 4837c478bd9Sstevel@tonic-gate mem_info.olddispq = dq->disp_q; 4847c478bd9Sstevel@tonic-gate mem_info.olddqactmap = dq->disp_qactmap; 4857c478bd9Sstevel@tonic-gate mem_info.oldnglobpris = dq->disp_npri; 4867c478bd9Sstevel@tonic-gate disp_dq_free(&mem_info); 4877c478bd9Sstevel@tonic-gate } 4887c478bd9Sstevel@tonic-gate 4897c478bd9Sstevel@tonic-gate /* 4907c478bd9Sstevel@tonic-gate * End dispatcher and scheduler initialization. 4917c478bd9Sstevel@tonic-gate */ 4927c478bd9Sstevel@tonic-gate 4937c478bd9Sstevel@tonic-gate /* 4947c478bd9Sstevel@tonic-gate * See if there's anything to do other than remain idle. 4957c478bd9Sstevel@tonic-gate * Return non-zero if there is. 4967c478bd9Sstevel@tonic-gate * 4977c478bd9Sstevel@tonic-gate * This function must be called with high spl, or with 4987c478bd9Sstevel@tonic-gate * kernel preemption disabled to prevent the partition's 4997c478bd9Sstevel@tonic-gate * active cpu list from changing while being traversed. 5007c478bd9Sstevel@tonic-gate * 5017c478bd9Sstevel@tonic-gate */ 5027c478bd9Sstevel@tonic-gate int 5037c478bd9Sstevel@tonic-gate disp_anywork(void) 5047c478bd9Sstevel@tonic-gate { 5057c478bd9Sstevel@tonic-gate cpu_t *cp = CPU; 5067c478bd9Sstevel@tonic-gate cpu_t *ocp; 5077c478bd9Sstevel@tonic-gate 5087c478bd9Sstevel@tonic-gate if (cp->cpu_disp->disp_nrunnable != 0) 5097c478bd9Sstevel@tonic-gate return (1); 5107c478bd9Sstevel@tonic-gate 5117c478bd9Sstevel@tonic-gate if (!(cp->cpu_flags & CPU_OFFLINE)) { 5127c478bd9Sstevel@tonic-gate if (CP_MAXRUNPRI(cp->cpu_part) >= 0) 5137c478bd9Sstevel@tonic-gate return (1); 5147c478bd9Sstevel@tonic-gate 5157c478bd9Sstevel@tonic-gate /* 5167c478bd9Sstevel@tonic-gate * Work can be taken from another CPU if: 5177c478bd9Sstevel@tonic-gate * - There is unbound work on the run queue 5187c478bd9Sstevel@tonic-gate * - That work isn't a thread undergoing a 5197c478bd9Sstevel@tonic-gate * - context switch on an otherwise empty queue. 5207c478bd9Sstevel@tonic-gate * - The CPU isn't running the idle loop. 5217c478bd9Sstevel@tonic-gate */ 5227c478bd9Sstevel@tonic-gate for (ocp = cp->cpu_next_part; ocp != cp; 5237c478bd9Sstevel@tonic-gate ocp = ocp->cpu_next_part) { 5247c478bd9Sstevel@tonic-gate ASSERT(CPU_ACTIVE(ocp)); 5257c478bd9Sstevel@tonic-gate 5267c478bd9Sstevel@tonic-gate if (ocp->cpu_disp->disp_max_unbound_pri != -1 && 5277c478bd9Sstevel@tonic-gate !((ocp->cpu_disp_flags & CPU_DISP_DONTSTEAL) && 5287c478bd9Sstevel@tonic-gate ocp->cpu_disp->disp_nrunnable == 1) && 5297c478bd9Sstevel@tonic-gate ocp->cpu_dispatch_pri != -1) 5307c478bd9Sstevel@tonic-gate return (1); 5317c478bd9Sstevel@tonic-gate } 5327c478bd9Sstevel@tonic-gate } 5337c478bd9Sstevel@tonic-gate return (0); 5347c478bd9Sstevel@tonic-gate } 5357c478bd9Sstevel@tonic-gate 5367c478bd9Sstevel@tonic-gate /* 5377c478bd9Sstevel@tonic-gate * Called when CPU enters the idle loop 5387c478bd9Sstevel@tonic-gate */ 5397c478bd9Sstevel@tonic-gate static void 5407c478bd9Sstevel@tonic-gate idle_enter() 5417c478bd9Sstevel@tonic-gate { 5427c478bd9Sstevel@tonic-gate cpu_t *cp = CPU; 5437c478bd9Sstevel@tonic-gate 544eda89462Sesolom new_cpu_mstate(CMS_IDLE, gethrtime_unscaled()); 5457c478bd9Sstevel@tonic-gate CPU_STATS_ADDQ(cp, sys, idlethread, 1); 5467c478bd9Sstevel@tonic-gate set_idle_cpu(cp->cpu_id); /* arch-dependent hook */ 5477c478bd9Sstevel@tonic-gate } 5487c478bd9Sstevel@tonic-gate 5497c478bd9Sstevel@tonic-gate /* 5507c478bd9Sstevel@tonic-gate * Called when CPU exits the idle loop 5517c478bd9Sstevel@tonic-gate */ 5527c478bd9Sstevel@tonic-gate static void 5537c478bd9Sstevel@tonic-gate idle_exit() 5547c478bd9Sstevel@tonic-gate { 5557c478bd9Sstevel@tonic-gate cpu_t *cp = CPU; 5567c478bd9Sstevel@tonic-gate 557eda89462Sesolom new_cpu_mstate(CMS_SYSTEM, gethrtime_unscaled()); 5587c478bd9Sstevel@tonic-gate unset_idle_cpu(cp->cpu_id); /* arch-dependent hook */ 5597c478bd9Sstevel@tonic-gate } 5607c478bd9Sstevel@tonic-gate 5617c478bd9Sstevel@tonic-gate /* 5627c478bd9Sstevel@tonic-gate * Idle loop. 5637c478bd9Sstevel@tonic-gate */ 5647c478bd9Sstevel@tonic-gate void 5657c478bd9Sstevel@tonic-gate idle() 5667c478bd9Sstevel@tonic-gate { 5677c478bd9Sstevel@tonic-gate struct cpu *cp = CPU; /* pointer to this CPU */ 5687c478bd9Sstevel@tonic-gate kthread_t *t; /* taken thread */ 5697c478bd9Sstevel@tonic-gate 5707c478bd9Sstevel@tonic-gate idle_enter(); 5717c478bd9Sstevel@tonic-gate 5727c478bd9Sstevel@tonic-gate /* 5737c478bd9Sstevel@tonic-gate * Uniprocessor version of idle loop. 5747c478bd9Sstevel@tonic-gate * Do this until notified that we're on an actual multiprocessor. 5757c478bd9Sstevel@tonic-gate */ 5767c478bd9Sstevel@tonic-gate while (ncpus == 1) { 5777c478bd9Sstevel@tonic-gate if (cp->cpu_disp->disp_nrunnable == 0) { 5787c478bd9Sstevel@tonic-gate (*idle_cpu)(); 5797c478bd9Sstevel@tonic-gate continue; 5807c478bd9Sstevel@tonic-gate } 5817c478bd9Sstevel@tonic-gate idle_exit(); 5827c478bd9Sstevel@tonic-gate swtch(); 5837c478bd9Sstevel@tonic-gate 5847c478bd9Sstevel@tonic-gate idle_enter(); /* returned from swtch */ 5857c478bd9Sstevel@tonic-gate } 5867c478bd9Sstevel@tonic-gate 5877c478bd9Sstevel@tonic-gate /* 5887c478bd9Sstevel@tonic-gate * Multiprocessor idle loop. 5897c478bd9Sstevel@tonic-gate */ 5907c478bd9Sstevel@tonic-gate for (;;) { 5917c478bd9Sstevel@tonic-gate /* 5927c478bd9Sstevel@tonic-gate * If CPU is completely quiesced by p_online(2), just wait 5937c478bd9Sstevel@tonic-gate * here with minimal bus traffic until put online. 5947c478bd9Sstevel@tonic-gate */ 5957c478bd9Sstevel@tonic-gate while (cp->cpu_flags & CPU_QUIESCED) 5967c478bd9Sstevel@tonic-gate (*idle_cpu)(); 5977c478bd9Sstevel@tonic-gate 5987c478bd9Sstevel@tonic-gate if (cp->cpu_disp->disp_nrunnable != 0) { 5997c478bd9Sstevel@tonic-gate idle_exit(); 6007c478bd9Sstevel@tonic-gate swtch(); 6017c478bd9Sstevel@tonic-gate } else { 6027c478bd9Sstevel@tonic-gate if (cp->cpu_flags & CPU_OFFLINE) 6037c478bd9Sstevel@tonic-gate continue; 6047c478bd9Sstevel@tonic-gate if ((t = disp_getwork(cp)) == NULL) { 6057c478bd9Sstevel@tonic-gate if (cp->cpu_chosen_level != -1) { 6067c478bd9Sstevel@tonic-gate disp_t *dp = cp->cpu_disp; 6077c478bd9Sstevel@tonic-gate disp_t *kpq; 6087c478bd9Sstevel@tonic-gate 6097c478bd9Sstevel@tonic-gate disp_lock_enter(&dp->disp_lock); 6107c478bd9Sstevel@tonic-gate /* 6117c478bd9Sstevel@tonic-gate * Set kpq under lock to prevent 6127c478bd9Sstevel@tonic-gate * migration between partitions. 6137c478bd9Sstevel@tonic-gate */ 6147c478bd9Sstevel@tonic-gate kpq = &cp->cpu_part->cp_kp_queue; 6157c478bd9Sstevel@tonic-gate if (kpq->disp_maxrunpri == -1) 6167c478bd9Sstevel@tonic-gate cp->cpu_chosen_level = -1; 6177c478bd9Sstevel@tonic-gate disp_lock_exit(&dp->disp_lock); 6187c478bd9Sstevel@tonic-gate } 6197c478bd9Sstevel@tonic-gate (*idle_cpu)(); 6207c478bd9Sstevel@tonic-gate continue; 6217c478bd9Sstevel@tonic-gate } 622685679f7Sakolb /* 623685679f7Sakolb * If there was a thread but we couldn't steal 624685679f7Sakolb * it, then keep trying. 625685679f7Sakolb */ 626685679f7Sakolb if (t == T_DONTSTEAL) 627685679f7Sakolb continue; 6287c478bd9Sstevel@tonic-gate idle_exit(); 6297c478bd9Sstevel@tonic-gate swtch_to(t); 6307c478bd9Sstevel@tonic-gate } 6317c478bd9Sstevel@tonic-gate idle_enter(); /* returned from swtch/swtch_to */ 6327c478bd9Sstevel@tonic-gate } 6337c478bd9Sstevel@tonic-gate } 6347c478bd9Sstevel@tonic-gate 6357c478bd9Sstevel@tonic-gate 6367c478bd9Sstevel@tonic-gate /* 6377c478bd9Sstevel@tonic-gate * Preempt the currently running thread in favor of the highest 6387c478bd9Sstevel@tonic-gate * priority thread. The class of the current thread controls 6397c478bd9Sstevel@tonic-gate * where it goes on the dispatcher queues. If panicking, turn 6407c478bd9Sstevel@tonic-gate * preemption off. 6417c478bd9Sstevel@tonic-gate */ 6427c478bd9Sstevel@tonic-gate void 6437c478bd9Sstevel@tonic-gate preempt() 6447c478bd9Sstevel@tonic-gate { 6457c478bd9Sstevel@tonic-gate kthread_t *t = curthread; 6467c478bd9Sstevel@tonic-gate klwp_t *lwp = ttolwp(curthread); 6477c478bd9Sstevel@tonic-gate 6487c478bd9Sstevel@tonic-gate if (panicstr) 6497c478bd9Sstevel@tonic-gate return; 6507c478bd9Sstevel@tonic-gate 6517c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_PREEMPT_START, "preempt_start"); 6527c478bd9Sstevel@tonic-gate 6537c478bd9Sstevel@tonic-gate thread_lock(t); 6547c478bd9Sstevel@tonic-gate 6557c478bd9Sstevel@tonic-gate if (t->t_state != TS_ONPROC || t->t_disp_queue != CPU->cpu_disp) { 6567c478bd9Sstevel@tonic-gate /* 6577c478bd9Sstevel@tonic-gate * this thread has already been chosen to be run on 6587c478bd9Sstevel@tonic-gate * another CPU. Clear kprunrun on this CPU since we're 6597c478bd9Sstevel@tonic-gate * already headed for swtch(). 6607c478bd9Sstevel@tonic-gate */ 6617c478bd9Sstevel@tonic-gate CPU->cpu_kprunrun = 0; 6627c478bd9Sstevel@tonic-gate thread_unlock_nopreempt(t); 6637c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_PREEMPT_END, "preempt_end"); 6647c478bd9Sstevel@tonic-gate } else { 6657c478bd9Sstevel@tonic-gate if (lwp != NULL) 6667c478bd9Sstevel@tonic-gate lwp->lwp_ru.nivcsw++; 6677c478bd9Sstevel@tonic-gate CPU_STATS_ADDQ(CPU, sys, inv_swtch, 1); 6687c478bd9Sstevel@tonic-gate THREAD_TRANSITION(t); 6697c478bd9Sstevel@tonic-gate CL_PREEMPT(t); 6707c478bd9Sstevel@tonic-gate DTRACE_SCHED(preempt); 6717c478bd9Sstevel@tonic-gate thread_unlock_nopreempt(t); 6727c478bd9Sstevel@tonic-gate 6737c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_PREEMPT_END, "preempt_end"); 6747c478bd9Sstevel@tonic-gate 6757c478bd9Sstevel@tonic-gate swtch(); /* clears CPU->cpu_runrun via disp() */ 6767c478bd9Sstevel@tonic-gate } 6777c478bd9Sstevel@tonic-gate } 6787c478bd9Sstevel@tonic-gate 6797c478bd9Sstevel@tonic-gate extern kthread_t *thread_unpin(); 6807c478bd9Sstevel@tonic-gate 6817c478bd9Sstevel@tonic-gate /* 6827c478bd9Sstevel@tonic-gate * disp() - find the highest priority thread for this processor to run, and 6837c478bd9Sstevel@tonic-gate * set it in TS_ONPROC state so that resume() can be called to run it. 6847c478bd9Sstevel@tonic-gate */ 6857c478bd9Sstevel@tonic-gate static kthread_t * 6867c478bd9Sstevel@tonic-gate disp() 6877c478bd9Sstevel@tonic-gate { 6887c478bd9Sstevel@tonic-gate cpu_t *cpup; 6897c478bd9Sstevel@tonic-gate disp_t *dp; 6907c478bd9Sstevel@tonic-gate kthread_t *tp; 6917c478bd9Sstevel@tonic-gate dispq_t *dq; 6927c478bd9Sstevel@tonic-gate int maxrunword; 6937c478bd9Sstevel@tonic-gate pri_t pri; 6947c478bd9Sstevel@tonic-gate disp_t *kpq; 6957c478bd9Sstevel@tonic-gate 6967c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_DISP_START, "disp_start"); 6977c478bd9Sstevel@tonic-gate 6987c478bd9Sstevel@tonic-gate cpup = CPU; 6997c478bd9Sstevel@tonic-gate /* 7007c478bd9Sstevel@tonic-gate * Find the highest priority loaded, runnable thread. 7017c478bd9Sstevel@tonic-gate */ 7027c478bd9Sstevel@tonic-gate dp = cpup->cpu_disp; 7037c478bd9Sstevel@tonic-gate 7047c478bd9Sstevel@tonic-gate reschedule: 7057c478bd9Sstevel@tonic-gate /* 7067c478bd9Sstevel@tonic-gate * If there is more important work on the global queue with a better 7077c478bd9Sstevel@tonic-gate * priority than the maximum on this CPU, take it now. 7087c478bd9Sstevel@tonic-gate */ 7097c478bd9Sstevel@tonic-gate kpq = &cpup->cpu_part->cp_kp_queue; 7107c478bd9Sstevel@tonic-gate while ((pri = kpq->disp_maxrunpri) >= 0 && 7117c478bd9Sstevel@tonic-gate pri >= dp->disp_maxrunpri && 7127c478bd9Sstevel@tonic-gate (cpup->cpu_flags & CPU_OFFLINE) == 0 && 7137c478bd9Sstevel@tonic-gate (tp = disp_getbest(kpq)) != NULL) { 7147c478bd9Sstevel@tonic-gate if (disp_ratify(tp, kpq) != NULL) { 7157c478bd9Sstevel@tonic-gate TRACE_1(TR_FAC_DISP, TR_DISP_END, 7167c478bd9Sstevel@tonic-gate "disp_end:tid %p", tp); 7177c478bd9Sstevel@tonic-gate return (tp); 7187c478bd9Sstevel@tonic-gate } 7197c478bd9Sstevel@tonic-gate } 7207c478bd9Sstevel@tonic-gate 7217c478bd9Sstevel@tonic-gate disp_lock_enter(&dp->disp_lock); 7227c478bd9Sstevel@tonic-gate pri = dp->disp_maxrunpri; 7237c478bd9Sstevel@tonic-gate 7247c478bd9Sstevel@tonic-gate /* 7257c478bd9Sstevel@tonic-gate * If there is nothing to run, look at what's runnable on other queues. 7267c478bd9Sstevel@tonic-gate * Choose the idle thread if the CPU is quiesced. 7277c478bd9Sstevel@tonic-gate * Note that CPUs that have the CPU_OFFLINE flag set can still run 7287c478bd9Sstevel@tonic-gate * interrupt threads, which will be the only threads on the CPU's own 7297c478bd9Sstevel@tonic-gate * queue, but cannot run threads from other queues. 7307c478bd9Sstevel@tonic-gate */ 7317c478bd9Sstevel@tonic-gate if (pri == -1) { 7327c478bd9Sstevel@tonic-gate if (!(cpup->cpu_flags & CPU_OFFLINE)) { 7337c478bd9Sstevel@tonic-gate disp_lock_exit(&dp->disp_lock); 734685679f7Sakolb if ((tp = disp_getwork(cpup)) == NULL || 735685679f7Sakolb tp == T_DONTSTEAL) { 7367c478bd9Sstevel@tonic-gate tp = cpup->cpu_idle_thread; 7377c478bd9Sstevel@tonic-gate (void) splhigh(); 7387c478bd9Sstevel@tonic-gate THREAD_ONPROC(tp, cpup); 7397c478bd9Sstevel@tonic-gate cpup->cpu_dispthread = tp; 7407c478bd9Sstevel@tonic-gate cpup->cpu_dispatch_pri = -1; 7417c478bd9Sstevel@tonic-gate cpup->cpu_runrun = cpup->cpu_kprunrun = 0; 7427c478bd9Sstevel@tonic-gate cpup->cpu_chosen_level = -1; 7437c478bd9Sstevel@tonic-gate } 7447c478bd9Sstevel@tonic-gate } else { 7457c478bd9Sstevel@tonic-gate disp_lock_exit_high(&dp->disp_lock); 7467c478bd9Sstevel@tonic-gate tp = cpup->cpu_idle_thread; 7477c478bd9Sstevel@tonic-gate THREAD_ONPROC(tp, cpup); 7487c478bd9Sstevel@tonic-gate cpup->cpu_dispthread = tp; 7497c478bd9Sstevel@tonic-gate cpup->cpu_dispatch_pri = -1; 7507c478bd9Sstevel@tonic-gate cpup->cpu_runrun = cpup->cpu_kprunrun = 0; 7517c478bd9Sstevel@tonic-gate cpup->cpu_chosen_level = -1; 7527c478bd9Sstevel@tonic-gate } 7537c478bd9Sstevel@tonic-gate TRACE_1(TR_FAC_DISP, TR_DISP_END, 7547c478bd9Sstevel@tonic-gate "disp_end:tid %p", tp); 7557c478bd9Sstevel@tonic-gate return (tp); 7567c478bd9Sstevel@tonic-gate } 7577c478bd9Sstevel@tonic-gate 7587c478bd9Sstevel@tonic-gate dq = &dp->disp_q[pri]; 7597c478bd9Sstevel@tonic-gate tp = dq->dq_first; 7607c478bd9Sstevel@tonic-gate 7617c478bd9Sstevel@tonic-gate ASSERT(tp != NULL); 7627c478bd9Sstevel@tonic-gate ASSERT(tp->t_schedflag & TS_LOAD); /* thread must be swapped in */ 7637c478bd9Sstevel@tonic-gate 7647c478bd9Sstevel@tonic-gate DTRACE_SCHED2(dequeue, kthread_t *, tp, disp_t *, dp); 7657c478bd9Sstevel@tonic-gate 7667c478bd9Sstevel@tonic-gate /* 7677c478bd9Sstevel@tonic-gate * Found it so remove it from queue. 7687c478bd9Sstevel@tonic-gate */ 7697c478bd9Sstevel@tonic-gate dp->disp_nrunnable--; 7707c478bd9Sstevel@tonic-gate dq->dq_sruncnt--; 7717c478bd9Sstevel@tonic-gate if ((dq->dq_first = tp->t_link) == NULL) { 7727c478bd9Sstevel@tonic-gate ulong_t *dqactmap = dp->disp_qactmap; 7737c478bd9Sstevel@tonic-gate 7747c478bd9Sstevel@tonic-gate ASSERT(dq->dq_sruncnt == 0); 7757c478bd9Sstevel@tonic-gate dq->dq_last = NULL; 7767c478bd9Sstevel@tonic-gate 7777c478bd9Sstevel@tonic-gate /* 7787c478bd9Sstevel@tonic-gate * The queue is empty, so the corresponding bit needs to be 7797c478bd9Sstevel@tonic-gate * turned off in dqactmap. If nrunnable != 0 just took the 7807c478bd9Sstevel@tonic-gate * last runnable thread off the 7817c478bd9Sstevel@tonic-gate * highest queue, so recompute disp_maxrunpri. 7827c478bd9Sstevel@tonic-gate */ 7837c478bd9Sstevel@tonic-gate maxrunword = pri >> BT_ULSHIFT; 7847c478bd9Sstevel@tonic-gate dqactmap[maxrunword] &= ~BT_BIW(pri); 7857c478bd9Sstevel@tonic-gate 7867c478bd9Sstevel@tonic-gate if (dp->disp_nrunnable == 0) { 7877c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = -1; 7887c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = -1; 7897c478bd9Sstevel@tonic-gate } else { 7907c478bd9Sstevel@tonic-gate int ipri; 7917c478bd9Sstevel@tonic-gate 7927c478bd9Sstevel@tonic-gate ipri = bt_gethighbit(dqactmap, maxrunword); 7937c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = ipri; 7947c478bd9Sstevel@tonic-gate if (ipri < dp->disp_max_unbound_pri) 7957c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = ipri; 7967c478bd9Sstevel@tonic-gate } 7977c478bd9Sstevel@tonic-gate } else { 7987c478bd9Sstevel@tonic-gate tp->t_link = NULL; 7997c478bd9Sstevel@tonic-gate } 8007c478bd9Sstevel@tonic-gate 8017c478bd9Sstevel@tonic-gate /* 8027c478bd9Sstevel@tonic-gate * Set TS_DONT_SWAP flag to prevent another processor from swapping 8037c478bd9Sstevel@tonic-gate * out this thread before we have a chance to run it. 8047c478bd9Sstevel@tonic-gate * While running, it is protected against swapping by t_lock. 8057c478bd9Sstevel@tonic-gate */ 8067c478bd9Sstevel@tonic-gate tp->t_schedflag |= TS_DONT_SWAP; 8077c478bd9Sstevel@tonic-gate cpup->cpu_dispthread = tp; /* protected by spl only */ 8087c478bd9Sstevel@tonic-gate cpup->cpu_dispatch_pri = pri; 8097c478bd9Sstevel@tonic-gate ASSERT(pri == DISP_PRIO(tp)); 8107c478bd9Sstevel@tonic-gate thread_onproc(tp, cpup); /* set t_state to TS_ONPROC */ 8117c478bd9Sstevel@tonic-gate disp_lock_exit_high(&dp->disp_lock); /* drop run queue lock */ 8127c478bd9Sstevel@tonic-gate 8137c478bd9Sstevel@tonic-gate ASSERT(tp != NULL); 8147c478bd9Sstevel@tonic-gate TRACE_1(TR_FAC_DISP, TR_DISP_END, 8157c478bd9Sstevel@tonic-gate "disp_end:tid %p", tp); 8167c478bd9Sstevel@tonic-gate 8177c478bd9Sstevel@tonic-gate if (disp_ratify(tp, kpq) == NULL) 8187c478bd9Sstevel@tonic-gate goto reschedule; 8197c478bd9Sstevel@tonic-gate 8207c478bd9Sstevel@tonic-gate return (tp); 8217c478bd9Sstevel@tonic-gate } 8227c478bd9Sstevel@tonic-gate 8237c478bd9Sstevel@tonic-gate /* 8247c478bd9Sstevel@tonic-gate * swtch() 8257c478bd9Sstevel@tonic-gate * Find best runnable thread and run it. 8267c478bd9Sstevel@tonic-gate * Called with the current thread already switched to a new state, 8277c478bd9Sstevel@tonic-gate * on a sleep queue, run queue, stopped, and not zombied. 8287c478bd9Sstevel@tonic-gate * May be called at any spl level less than or equal to LOCK_LEVEL. 8297c478bd9Sstevel@tonic-gate * Always drops spl to the base level (spl0()). 8307c478bd9Sstevel@tonic-gate */ 8317c478bd9Sstevel@tonic-gate void 8327c478bd9Sstevel@tonic-gate swtch() 8337c478bd9Sstevel@tonic-gate { 8347c478bd9Sstevel@tonic-gate kthread_t *t = curthread; 8357c478bd9Sstevel@tonic-gate kthread_t *next; 8367c478bd9Sstevel@tonic-gate cpu_t *cp; 8377c478bd9Sstevel@tonic-gate 8387c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_SWTCH_START, "swtch_start"); 8397c478bd9Sstevel@tonic-gate 8407c478bd9Sstevel@tonic-gate if (t->t_flag & T_INTR_THREAD) 8417c478bd9Sstevel@tonic-gate cpu_intr_swtch_enter(t); 8427c478bd9Sstevel@tonic-gate 8437c478bd9Sstevel@tonic-gate if (t->t_intr != NULL) { 8447c478bd9Sstevel@tonic-gate /* 8457c478bd9Sstevel@tonic-gate * We are an interrupt thread. Setup and return 8467c478bd9Sstevel@tonic-gate * the interrupted thread to be resumed. 8477c478bd9Sstevel@tonic-gate */ 8487c478bd9Sstevel@tonic-gate (void) splhigh(); /* block other scheduler action */ 8497c478bd9Sstevel@tonic-gate cp = CPU; /* now protected against migration */ 8507c478bd9Sstevel@tonic-gate ASSERT(CPU_ON_INTR(cp) == 0); /* not called with PIL > 10 */ 8517c478bd9Sstevel@tonic-gate CPU_STATS_ADDQ(cp, sys, pswitch, 1); 8527c478bd9Sstevel@tonic-gate CPU_STATS_ADDQ(cp, sys, intrblk, 1); 8537c478bd9Sstevel@tonic-gate next = thread_unpin(); 8547c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start"); 8557c478bd9Sstevel@tonic-gate resume_from_intr(next); 8567c478bd9Sstevel@tonic-gate } else { 8577c478bd9Sstevel@tonic-gate #ifdef DEBUG 8587c478bd9Sstevel@tonic-gate if (t->t_state == TS_ONPROC && 8597c478bd9Sstevel@tonic-gate t->t_disp_queue->disp_cpu == CPU && 8607c478bd9Sstevel@tonic-gate t->t_preempt == 0) { 8617c478bd9Sstevel@tonic-gate thread_lock(t); 8627c478bd9Sstevel@tonic-gate ASSERT(t->t_state != TS_ONPROC || 8637c478bd9Sstevel@tonic-gate t->t_disp_queue->disp_cpu != CPU || 8647c478bd9Sstevel@tonic-gate t->t_preempt != 0); /* cannot migrate */ 8657c478bd9Sstevel@tonic-gate thread_unlock_nopreempt(t); 8667c478bd9Sstevel@tonic-gate } 8677c478bd9Sstevel@tonic-gate #endif /* DEBUG */ 8687c478bd9Sstevel@tonic-gate cp = CPU; 8697c478bd9Sstevel@tonic-gate next = disp(); /* returns with spl high */ 8707c478bd9Sstevel@tonic-gate ASSERT(CPU_ON_INTR(cp) == 0); /* not called with PIL > 10 */ 8717c478bd9Sstevel@tonic-gate 8727c478bd9Sstevel@tonic-gate /* OK to steal anything left on run queue */ 8737c478bd9Sstevel@tonic-gate cp->cpu_disp_flags &= ~CPU_DISP_DONTSTEAL; 8747c478bd9Sstevel@tonic-gate 8757c478bd9Sstevel@tonic-gate if (next != t) { 8767c478bd9Sstevel@tonic-gate if (t == cp->cpu_idle_thread) { 8777c478bd9Sstevel@tonic-gate CHIP_NRUNNING(cp->cpu_chip, 1); 8787c478bd9Sstevel@tonic-gate } else if (next == cp->cpu_idle_thread) { 8797c478bd9Sstevel@tonic-gate CHIP_NRUNNING(cp->cpu_chip, -1); 8807c478bd9Sstevel@tonic-gate } 8817c478bd9Sstevel@tonic-gate 882*f2bd4627Sjohansen /* 883*f2bd4627Sjohansen * If t was previously in the TS_ONPROC state, 884*f2bd4627Sjohansen * setfrontdq and setbackdq won't have set its t_waitrq. 885*f2bd4627Sjohansen * Since we now finally know that we're switching away 886*f2bd4627Sjohansen * from this thread, set its t_waitrq if it is on a run 887*f2bd4627Sjohansen * queue. 888*f2bd4627Sjohansen */ 889*f2bd4627Sjohansen if ((t->t_state == TS_RUN) && (t->t_waitrq == 0)) { 890*f2bd4627Sjohansen t->t_waitrq = gethrtime_unscaled(); 891*f2bd4627Sjohansen } 892*f2bd4627Sjohansen 893*f2bd4627Sjohansen /* 894*f2bd4627Sjohansen * restore mstate of thread that we are switching to 895*f2bd4627Sjohansen */ 896*f2bd4627Sjohansen restore_mstate(next); 897*f2bd4627Sjohansen 8987c478bd9Sstevel@tonic-gate CPU_STATS_ADDQ(cp, sys, pswitch, 1); 8997c478bd9Sstevel@tonic-gate cp->cpu_last_swtch = t->t_disp_time = lbolt; 9007c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start"); 9017c478bd9Sstevel@tonic-gate 9027c478bd9Sstevel@tonic-gate if (dtrace_vtime_active) 9037c478bd9Sstevel@tonic-gate dtrace_vtime_switch(next); 9047c478bd9Sstevel@tonic-gate 9057c478bd9Sstevel@tonic-gate resume(next); 9067c478bd9Sstevel@tonic-gate /* 9077c478bd9Sstevel@tonic-gate * The TR_RESUME_END and TR_SWTCH_END trace points 9087c478bd9Sstevel@tonic-gate * appear at the end of resume(), because we may not 9097c478bd9Sstevel@tonic-gate * return here 9107c478bd9Sstevel@tonic-gate */ 9117c478bd9Sstevel@tonic-gate } else { 9127c478bd9Sstevel@tonic-gate if (t->t_flag & T_INTR_THREAD) 9137c478bd9Sstevel@tonic-gate cpu_intr_swtch_exit(t); 9147c478bd9Sstevel@tonic-gate 9157c478bd9Sstevel@tonic-gate DTRACE_SCHED(remain__cpu); 9167c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_SWTCH_END, "swtch_end"); 9177c478bd9Sstevel@tonic-gate (void) spl0(); 9187c478bd9Sstevel@tonic-gate } 9197c478bd9Sstevel@tonic-gate } 9207c478bd9Sstevel@tonic-gate } 9217c478bd9Sstevel@tonic-gate 9227c478bd9Sstevel@tonic-gate /* 9237c478bd9Sstevel@tonic-gate * swtch_from_zombie() 9247c478bd9Sstevel@tonic-gate * Special case of swtch(), which allows checks for TS_ZOMB to be 9257c478bd9Sstevel@tonic-gate * eliminated from normal resume. 9267c478bd9Sstevel@tonic-gate * Find best runnable thread and run it. 9277c478bd9Sstevel@tonic-gate * Called with the current thread zombied. 9287c478bd9Sstevel@tonic-gate * Zombies cannot migrate, so CPU references are safe. 9297c478bd9Sstevel@tonic-gate */ 9307c478bd9Sstevel@tonic-gate void 9317c478bd9Sstevel@tonic-gate swtch_from_zombie() 9327c478bd9Sstevel@tonic-gate { 9337c478bd9Sstevel@tonic-gate kthread_t *next; 9347c478bd9Sstevel@tonic-gate cpu_t *cpu = CPU; 9357c478bd9Sstevel@tonic-gate 9367c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_SWTCH_START, "swtch_start"); 9377c478bd9Sstevel@tonic-gate 9387c478bd9Sstevel@tonic-gate ASSERT(curthread->t_state == TS_ZOMB); 9397c478bd9Sstevel@tonic-gate 9407c478bd9Sstevel@tonic-gate next = disp(); /* returns with spl high */ 9417c478bd9Sstevel@tonic-gate ASSERT(CPU_ON_INTR(CPU) == 0); /* not called with PIL > 10 */ 9427c478bd9Sstevel@tonic-gate CPU_STATS_ADDQ(CPU, sys, pswitch, 1); 9437c478bd9Sstevel@tonic-gate ASSERT(next != curthread); 9447c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start"); 9457c478bd9Sstevel@tonic-gate 9467c478bd9Sstevel@tonic-gate if (next == cpu->cpu_idle_thread) 9477c478bd9Sstevel@tonic-gate CHIP_NRUNNING(cpu->cpu_chip, -1); 9487c478bd9Sstevel@tonic-gate 949*f2bd4627Sjohansen restore_mstate(next); 950*f2bd4627Sjohansen 9517c478bd9Sstevel@tonic-gate if (dtrace_vtime_active) 9527c478bd9Sstevel@tonic-gate dtrace_vtime_switch(next); 9537c478bd9Sstevel@tonic-gate 9547c478bd9Sstevel@tonic-gate resume_from_zombie(next); 9557c478bd9Sstevel@tonic-gate /* 9567c478bd9Sstevel@tonic-gate * The TR_RESUME_END and TR_SWTCH_END trace points 9577c478bd9Sstevel@tonic-gate * appear at the end of resume(), because we certainly will not 9587c478bd9Sstevel@tonic-gate * return here 9597c478bd9Sstevel@tonic-gate */ 9607c478bd9Sstevel@tonic-gate } 9617c478bd9Sstevel@tonic-gate 9627c478bd9Sstevel@tonic-gate #if defined(DEBUG) && (defined(DISP_DEBUG) || defined(lint)) 9637c478bd9Sstevel@tonic-gate static int 9647c478bd9Sstevel@tonic-gate thread_on_queue(kthread_t *tp) 9657c478bd9Sstevel@tonic-gate { 9667c478bd9Sstevel@tonic-gate cpu_t *cp; 9677c478bd9Sstevel@tonic-gate cpu_t *self; 9687c478bd9Sstevel@tonic-gate disp_t *dp; 9697c478bd9Sstevel@tonic-gate 9707c478bd9Sstevel@tonic-gate self = CPU; 9717c478bd9Sstevel@tonic-gate cp = self->cpu_next_onln; 9727c478bd9Sstevel@tonic-gate dp = cp->cpu_disp; 9737c478bd9Sstevel@tonic-gate for (;;) { 9747c478bd9Sstevel@tonic-gate dispq_t *dq; 9757c478bd9Sstevel@tonic-gate dispq_t *eq; 9767c478bd9Sstevel@tonic-gate 9777c478bd9Sstevel@tonic-gate disp_lock_enter_high(&dp->disp_lock); 9787c478bd9Sstevel@tonic-gate for (dq = dp->disp_q, eq = dp->disp_q_limit; dq < eq; ++dq) { 9797c478bd9Sstevel@tonic-gate kthread_t *rp; 9807c478bd9Sstevel@tonic-gate 9817c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last == NULL || 9827c478bd9Sstevel@tonic-gate dq->dq_last->t_link == NULL); 9837c478bd9Sstevel@tonic-gate for (rp = dq->dq_first; rp; rp = rp->t_link) 9847c478bd9Sstevel@tonic-gate if (tp == rp) { 9857c478bd9Sstevel@tonic-gate disp_lock_exit_high(&dp->disp_lock); 9867c478bd9Sstevel@tonic-gate return (1); 9877c478bd9Sstevel@tonic-gate } 9887c478bd9Sstevel@tonic-gate } 9897c478bd9Sstevel@tonic-gate disp_lock_exit_high(&dp->disp_lock); 9907c478bd9Sstevel@tonic-gate if (cp == NULL) 9917c478bd9Sstevel@tonic-gate break; 9927c478bd9Sstevel@tonic-gate if (cp == self) { 9937c478bd9Sstevel@tonic-gate cp = NULL; 9947c478bd9Sstevel@tonic-gate dp = &cp->cpu_part->cp_kp_queue; 9957c478bd9Sstevel@tonic-gate } else { 9967c478bd9Sstevel@tonic-gate cp = cp->cpu_next_onln; 9977c478bd9Sstevel@tonic-gate dp = cp->cpu_disp; 9987c478bd9Sstevel@tonic-gate } 9997c478bd9Sstevel@tonic-gate } 10007c478bd9Sstevel@tonic-gate return (0); 10017c478bd9Sstevel@tonic-gate } /* end of thread_on_queue */ 10027c478bd9Sstevel@tonic-gate #else 10037c478bd9Sstevel@tonic-gate 10047c478bd9Sstevel@tonic-gate #define thread_on_queue(tp) 0 /* ASSERT must be !thread_on_queue */ 10057c478bd9Sstevel@tonic-gate 10067c478bd9Sstevel@tonic-gate #endif /* DEBUG */ 10077c478bd9Sstevel@tonic-gate 10087c478bd9Sstevel@tonic-gate /* 10097c478bd9Sstevel@tonic-gate * like swtch(), but switch to a specified thread taken from another CPU. 10107c478bd9Sstevel@tonic-gate * called with spl high.. 10117c478bd9Sstevel@tonic-gate */ 10127c478bd9Sstevel@tonic-gate void 10137c478bd9Sstevel@tonic-gate swtch_to(kthread_t *next) 10147c478bd9Sstevel@tonic-gate { 10157c478bd9Sstevel@tonic-gate cpu_t *cp = CPU; 10167c478bd9Sstevel@tonic-gate 10177c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_SWTCH_START, "swtch_start"); 10187c478bd9Sstevel@tonic-gate 10197c478bd9Sstevel@tonic-gate /* 10207c478bd9Sstevel@tonic-gate * Update context switch statistics. 10217c478bd9Sstevel@tonic-gate */ 10227c478bd9Sstevel@tonic-gate CPU_STATS_ADDQ(cp, sys, pswitch, 1); 10237c478bd9Sstevel@tonic-gate 10247c478bd9Sstevel@tonic-gate TRACE_0(TR_FAC_DISP, TR_RESUME_START, "resume_start"); 10257c478bd9Sstevel@tonic-gate 10267c478bd9Sstevel@tonic-gate if (curthread == cp->cpu_idle_thread) 10277c478bd9Sstevel@tonic-gate CHIP_NRUNNING(cp->cpu_chip, 1); 10287c478bd9Sstevel@tonic-gate 10297c478bd9Sstevel@tonic-gate /* OK to steal anything left on run queue */ 10307c478bd9Sstevel@tonic-gate cp->cpu_disp_flags &= ~CPU_DISP_DONTSTEAL; 10317c478bd9Sstevel@tonic-gate 10327c478bd9Sstevel@tonic-gate /* record last execution time */ 10337c478bd9Sstevel@tonic-gate cp->cpu_last_swtch = curthread->t_disp_time = lbolt; 10347c478bd9Sstevel@tonic-gate 1035*f2bd4627Sjohansen /* 1036*f2bd4627Sjohansen * If t was previously in the TS_ONPROC state, setfrontdq and setbackdq 1037*f2bd4627Sjohansen * won't have set its t_waitrq. Since we now finally know that we're 1038*f2bd4627Sjohansen * switching away from this thread, set its t_waitrq if it is on a run 1039*f2bd4627Sjohansen * queue. 1040*f2bd4627Sjohansen */ 1041*f2bd4627Sjohansen if ((curthread->t_state == TS_RUN) && (curthread->t_waitrq == 0)) { 1042*f2bd4627Sjohansen curthread->t_waitrq = gethrtime_unscaled(); 1043*f2bd4627Sjohansen } 1044*f2bd4627Sjohansen 1045*f2bd4627Sjohansen /* restore next thread to previously running microstate */ 1046*f2bd4627Sjohansen restore_mstate(next); 1047*f2bd4627Sjohansen 10487c478bd9Sstevel@tonic-gate if (dtrace_vtime_active) 10497c478bd9Sstevel@tonic-gate dtrace_vtime_switch(next); 10507c478bd9Sstevel@tonic-gate 10517c478bd9Sstevel@tonic-gate resume(next); 10527c478bd9Sstevel@tonic-gate /* 10537c478bd9Sstevel@tonic-gate * The TR_RESUME_END and TR_SWTCH_END trace points 10547c478bd9Sstevel@tonic-gate * appear at the end of resume(), because we may not 10557c478bd9Sstevel@tonic-gate * return here 10567c478bd9Sstevel@tonic-gate */ 10577c478bd9Sstevel@tonic-gate } 10587c478bd9Sstevel@tonic-gate 10597c478bd9Sstevel@tonic-gate 10607c478bd9Sstevel@tonic-gate 10617c478bd9Sstevel@tonic-gate #define CPU_IDLING(pri) ((pri) == -1) 10627c478bd9Sstevel@tonic-gate 10637c478bd9Sstevel@tonic-gate static void 10647c478bd9Sstevel@tonic-gate cpu_resched(cpu_t *cp, pri_t tpri) 10657c478bd9Sstevel@tonic-gate { 10667c478bd9Sstevel@tonic-gate int call_poke_cpu = 0; 10677c478bd9Sstevel@tonic-gate pri_t cpupri = cp->cpu_dispatch_pri; 10687c478bd9Sstevel@tonic-gate 10697c478bd9Sstevel@tonic-gate if (!CPU_IDLING(cpupri) && (cpupri < tpri)) { 10707c478bd9Sstevel@tonic-gate TRACE_2(TR_FAC_DISP, TR_CPU_RESCHED, 10717c478bd9Sstevel@tonic-gate "CPU_RESCHED:Tpri %d Cpupri %d", tpri, cpupri); 10727c478bd9Sstevel@tonic-gate if (tpri >= upreemptpri && cp->cpu_runrun == 0) { 10737c478bd9Sstevel@tonic-gate cp->cpu_runrun = 1; 10747c478bd9Sstevel@tonic-gate aston(cp->cpu_dispthread); 10757c478bd9Sstevel@tonic-gate if (tpri < kpreemptpri && cp != CPU) 10767c478bd9Sstevel@tonic-gate call_poke_cpu = 1; 10777c478bd9Sstevel@tonic-gate } 10787c478bd9Sstevel@tonic-gate if (tpri >= kpreemptpri && cp->cpu_kprunrun == 0) { 10797c478bd9Sstevel@tonic-gate cp->cpu_kprunrun = 1; 10807c478bd9Sstevel@tonic-gate if (cp != CPU) 10817c478bd9Sstevel@tonic-gate call_poke_cpu = 1; 10827c478bd9Sstevel@tonic-gate } 10837c478bd9Sstevel@tonic-gate } 10847c478bd9Sstevel@tonic-gate 10857c478bd9Sstevel@tonic-gate /* 10867c478bd9Sstevel@tonic-gate * Propagate cpu_runrun, and cpu_kprunrun to global visibility. 10877c478bd9Sstevel@tonic-gate */ 10887c478bd9Sstevel@tonic-gate membar_enter(); 10897c478bd9Sstevel@tonic-gate 10907c478bd9Sstevel@tonic-gate if (call_poke_cpu) 10917c478bd9Sstevel@tonic-gate poke_cpu(cp->cpu_id); 10927c478bd9Sstevel@tonic-gate } 10937c478bd9Sstevel@tonic-gate 10947c478bd9Sstevel@tonic-gate /* 10957c478bd9Sstevel@tonic-gate * Routine used by setbackdq() to balance load across the physical 10967c478bd9Sstevel@tonic-gate * processors. Returns a CPU of a lesser loaded chip in the lgroup 10977c478bd9Sstevel@tonic-gate * if balancing is necessary, or the "hint" CPU if it's not. 10987c478bd9Sstevel@tonic-gate * 10997c478bd9Sstevel@tonic-gate * - tp is the thread being enqueued 11007c478bd9Sstevel@tonic-gate * - cp is a hint CPU (chosen by cpu_choose()). 11017c478bd9Sstevel@tonic-gate * - curchip (if not NULL) is the chip on which the current thread 11027c478bd9Sstevel@tonic-gate * is running. 11037c478bd9Sstevel@tonic-gate * 11047c478bd9Sstevel@tonic-gate * The thread lock for "tp" must be held while calling this routine. 11057c478bd9Sstevel@tonic-gate */ 11067c478bd9Sstevel@tonic-gate static cpu_t * 11077c478bd9Sstevel@tonic-gate chip_balance(kthread_t *tp, cpu_t *cp, chip_t *curchip) 11087c478bd9Sstevel@tonic-gate { 11097c478bd9Sstevel@tonic-gate int chp_nrun, ochp_nrun; 11107c478bd9Sstevel@tonic-gate chip_t *chp, *nchp; 11117c478bd9Sstevel@tonic-gate 11127c478bd9Sstevel@tonic-gate chp = cp->cpu_chip; 11137c478bd9Sstevel@tonic-gate chp_nrun = chp->chip_nrunning; 11147c478bd9Sstevel@tonic-gate 11157c478bd9Sstevel@tonic-gate if (chp == curchip) 11167c478bd9Sstevel@tonic-gate chp_nrun--; /* Ignore curthread */ 11177c478bd9Sstevel@tonic-gate 11187c478bd9Sstevel@tonic-gate /* 11197c478bd9Sstevel@tonic-gate * If this chip isn't at all idle, then let 11207c478bd9Sstevel@tonic-gate * run queue balancing do the work. 11217c478bd9Sstevel@tonic-gate */ 11227c478bd9Sstevel@tonic-gate if (chp_nrun == chp->chip_ncpu) 11237c478bd9Sstevel@tonic-gate return (cp); 11247c478bd9Sstevel@tonic-gate 11257c478bd9Sstevel@tonic-gate nchp = chp->chip_balance; 11267c478bd9Sstevel@tonic-gate do { 11277c478bd9Sstevel@tonic-gate if (nchp == chp || 11287c478bd9Sstevel@tonic-gate !CHIP_IN_CPUPART(nchp, tp->t_cpupart)) 11297c478bd9Sstevel@tonic-gate continue; 11307c478bd9Sstevel@tonic-gate 11317c478bd9Sstevel@tonic-gate ochp_nrun = nchp->chip_nrunning; 11327c478bd9Sstevel@tonic-gate 11337c478bd9Sstevel@tonic-gate /* 11347c478bd9Sstevel@tonic-gate * If the other chip is running less threads, 11357c478bd9Sstevel@tonic-gate * or if it's running the same number of threads, but 11367c478bd9Sstevel@tonic-gate * has more online logical CPUs, then choose to balance. 11377c478bd9Sstevel@tonic-gate */ 11387c478bd9Sstevel@tonic-gate if (chp_nrun > ochp_nrun || 11397c478bd9Sstevel@tonic-gate (chp_nrun == ochp_nrun && 11407c478bd9Sstevel@tonic-gate nchp->chip_ncpu > chp->chip_ncpu)) { 11417c478bd9Sstevel@tonic-gate cp = nchp->chip_cpus; 11427c478bd9Sstevel@tonic-gate nchp->chip_cpus = cp->cpu_next_chip; 11437c478bd9Sstevel@tonic-gate 11447c478bd9Sstevel@tonic-gate /* 11457c478bd9Sstevel@tonic-gate * Find a CPU on the chip in the correct 11467c478bd9Sstevel@tonic-gate * partition. We know at least one exists 11477c478bd9Sstevel@tonic-gate * because of the CHIP_IN_CPUPART() check above. 11487c478bd9Sstevel@tonic-gate */ 11497c478bd9Sstevel@tonic-gate while (cp->cpu_part != tp->t_cpupart) 11507c478bd9Sstevel@tonic-gate cp = cp->cpu_next_chip; 11517c478bd9Sstevel@tonic-gate } 11527c478bd9Sstevel@tonic-gate chp->chip_balance = nchp->chip_next_lgrp; 11537c478bd9Sstevel@tonic-gate break; 11547c478bd9Sstevel@tonic-gate } while ((nchp = nchp->chip_next_lgrp) != chp->chip_balance); 11557c478bd9Sstevel@tonic-gate 11567c478bd9Sstevel@tonic-gate ASSERT(CHIP_IN_CPUPART(cp->cpu_chip, tp->t_cpupart)); 11577c478bd9Sstevel@tonic-gate return (cp); 11587c478bd9Sstevel@tonic-gate } 11597c478bd9Sstevel@tonic-gate 11607c478bd9Sstevel@tonic-gate /* 11617c478bd9Sstevel@tonic-gate * setbackdq() keeps runqs balanced such that the difference in length 11627c478bd9Sstevel@tonic-gate * between the chosen runq and the next one is no more than RUNQ_MAX_DIFF. 11637c478bd9Sstevel@tonic-gate * For threads with priorities below RUNQ_MATCH_PRI levels, the runq's lengths 11647c478bd9Sstevel@tonic-gate * must match. When per-thread TS_RUNQMATCH flag is set, setbackdq() will 11657c478bd9Sstevel@tonic-gate * try to keep runqs perfectly balanced regardless of the thread priority. 11667c478bd9Sstevel@tonic-gate */ 11677c478bd9Sstevel@tonic-gate #define RUNQ_MATCH_PRI 16 /* pri below which queue lengths must match */ 11687c478bd9Sstevel@tonic-gate #define RUNQ_MAX_DIFF 2 /* maximum runq length difference */ 11697c478bd9Sstevel@tonic-gate #define RUNQ_LEN(cp, pri) ((cp)->cpu_disp->disp_q[pri].dq_sruncnt) 11707c478bd9Sstevel@tonic-gate 11717c478bd9Sstevel@tonic-gate /* 11727c478bd9Sstevel@tonic-gate * Put the specified thread on the back of the dispatcher 11737c478bd9Sstevel@tonic-gate * queue corresponding to its current priority. 11747c478bd9Sstevel@tonic-gate * 11757c478bd9Sstevel@tonic-gate * Called with the thread in transition, onproc or stopped state 11767c478bd9Sstevel@tonic-gate * and locked (transition implies locked) and at high spl. 11777c478bd9Sstevel@tonic-gate * Returns with the thread in TS_RUN state and still locked. 11787c478bd9Sstevel@tonic-gate */ 11797c478bd9Sstevel@tonic-gate void 11807c478bd9Sstevel@tonic-gate setbackdq(kthread_t *tp) 11817c478bd9Sstevel@tonic-gate { 11827c478bd9Sstevel@tonic-gate dispq_t *dq; 11837c478bd9Sstevel@tonic-gate disp_t *dp; 11847c478bd9Sstevel@tonic-gate chip_t *curchip = NULL; 11857c478bd9Sstevel@tonic-gate cpu_t *cp; 11867c478bd9Sstevel@tonic-gate pri_t tpri; 11877c478bd9Sstevel@tonic-gate int bound; 11887c478bd9Sstevel@tonic-gate 11897c478bd9Sstevel@tonic-gate ASSERT(THREAD_LOCK_HELD(tp)); 11907c478bd9Sstevel@tonic-gate ASSERT((tp->t_schedflag & TS_ALLSTART) == 0); 11917c478bd9Sstevel@tonic-gate ASSERT(!thread_on_queue(tp)); /* make sure tp isn't on a runq */ 11927c478bd9Sstevel@tonic-gate 11937c478bd9Sstevel@tonic-gate /* 11947c478bd9Sstevel@tonic-gate * If thread is "swapped" or on the swap queue don't 11957c478bd9Sstevel@tonic-gate * queue it, but wake sched. 11967c478bd9Sstevel@tonic-gate */ 11977c478bd9Sstevel@tonic-gate if ((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD) { 11987c478bd9Sstevel@tonic-gate disp_swapped_setrun(tp); 11997c478bd9Sstevel@tonic-gate return; 12007c478bd9Sstevel@tonic-gate } 12017c478bd9Sstevel@tonic-gate 12027c478bd9Sstevel@tonic-gate tpri = DISP_PRIO(tp); 12037c478bd9Sstevel@tonic-gate if (tp == curthread) { 12047c478bd9Sstevel@tonic-gate curchip = CPU->cpu_chip; 12057c478bd9Sstevel@tonic-gate } 12067c478bd9Sstevel@tonic-gate 12077c478bd9Sstevel@tonic-gate if (ncpus == 1) 12087c478bd9Sstevel@tonic-gate cp = tp->t_cpu; 12097c478bd9Sstevel@tonic-gate else if (!tp->t_bound_cpu && !tp->t_weakbound_cpu) { 12107c478bd9Sstevel@tonic-gate if (tpri >= kpqpri) { 12117c478bd9Sstevel@tonic-gate setkpdq(tp, SETKP_BACK); 12127c478bd9Sstevel@tonic-gate return; 12137c478bd9Sstevel@tonic-gate } 12147c478bd9Sstevel@tonic-gate /* 12157c478bd9Sstevel@tonic-gate * Let cpu_choose suggest a CPU. 12167c478bd9Sstevel@tonic-gate */ 12177c478bd9Sstevel@tonic-gate cp = cpu_choose(tp, tpri); 12187c478bd9Sstevel@tonic-gate 12197c478bd9Sstevel@tonic-gate if (tp->t_cpupart == cp->cpu_part) { 12207c478bd9Sstevel@tonic-gate int qlen; 12217c478bd9Sstevel@tonic-gate 12227c478bd9Sstevel@tonic-gate /* 12237c478bd9Sstevel@tonic-gate * Select another CPU if we need 12247c478bd9Sstevel@tonic-gate * to do some load balancing across the 12257c478bd9Sstevel@tonic-gate * physical processors. 12267c478bd9Sstevel@tonic-gate */ 12277c478bd9Sstevel@tonic-gate if (CHIP_SHOULD_BALANCE(cp->cpu_chip)) 12287c478bd9Sstevel@tonic-gate cp = chip_balance(tp, cp, curchip); 12297c478bd9Sstevel@tonic-gate 12307c478bd9Sstevel@tonic-gate /* 12317c478bd9Sstevel@tonic-gate * Balance across the run queues 12327c478bd9Sstevel@tonic-gate */ 12337c478bd9Sstevel@tonic-gate qlen = RUNQ_LEN(cp, tpri); 12347c478bd9Sstevel@tonic-gate if (tpri >= RUNQ_MATCH_PRI && 12357c478bd9Sstevel@tonic-gate !(tp->t_schedflag & TS_RUNQMATCH)) 12367c478bd9Sstevel@tonic-gate qlen -= RUNQ_MAX_DIFF; 12377c478bd9Sstevel@tonic-gate if (qlen > 0) { 1238685679f7Sakolb cpu_t *newcp; 12397c478bd9Sstevel@tonic-gate 1240685679f7Sakolb if (tp->t_lpl->lpl_lgrpid == LGRP_ROOTID) { 1241685679f7Sakolb newcp = cp->cpu_next_part; 1242685679f7Sakolb } else if ((newcp = cp->cpu_next_lpl) == cp) { 1243685679f7Sakolb newcp = cp->cpu_next_part; 1244685679f7Sakolb } 1245685679f7Sakolb 1246685679f7Sakolb if (RUNQ_LEN(newcp, tpri) < qlen) { 1247685679f7Sakolb DTRACE_PROBE3(runq__balance, 1248685679f7Sakolb kthread_t *, tp, 1249685679f7Sakolb cpu_t *, cp, cpu_t *, newcp); 1250685679f7Sakolb cp = newcp; 12517c478bd9Sstevel@tonic-gate } 12527c478bd9Sstevel@tonic-gate } 12537c478bd9Sstevel@tonic-gate } else { 12547c478bd9Sstevel@tonic-gate /* 12557c478bd9Sstevel@tonic-gate * Migrate to a cpu in the new partition. 12567c478bd9Sstevel@tonic-gate */ 12577c478bd9Sstevel@tonic-gate cp = disp_lowpri_cpu(tp->t_cpupart->cp_cpulist, 12587c478bd9Sstevel@tonic-gate tp->t_lpl, tp->t_pri, NULL); 12597c478bd9Sstevel@tonic-gate } 12607c478bd9Sstevel@tonic-gate bound = 0; 12617c478bd9Sstevel@tonic-gate ASSERT((cp->cpu_flags & CPU_QUIESCED) == 0); 12627c478bd9Sstevel@tonic-gate } else { 12637c478bd9Sstevel@tonic-gate /* 12647c478bd9Sstevel@tonic-gate * It is possible that t_weakbound_cpu != t_bound_cpu (for 12657c478bd9Sstevel@tonic-gate * a short time until weak binding that existed when the 12667c478bd9Sstevel@tonic-gate * strong binding was established has dropped) so we must 12677c478bd9Sstevel@tonic-gate * favour weak binding over strong. 12687c478bd9Sstevel@tonic-gate */ 12697c478bd9Sstevel@tonic-gate cp = tp->t_weakbound_cpu ? 12707c478bd9Sstevel@tonic-gate tp->t_weakbound_cpu : tp->t_bound_cpu; 12717c478bd9Sstevel@tonic-gate bound = 1; 12727c478bd9Sstevel@tonic-gate } 1273*f2bd4627Sjohansen /* 1274*f2bd4627Sjohansen * A thread that is ONPROC may be temporarily placed on the run queue 1275*f2bd4627Sjohansen * but then chosen to run again by disp. If the thread we're placing on 1276*f2bd4627Sjohansen * the queue is in TS_ONPROC state, don't set its t_waitrq until a 1277*f2bd4627Sjohansen * replacement process is actually scheduled in swtch(). In this 1278*f2bd4627Sjohansen * situation, curthread is the only thread that could be in the ONPROC 1279*f2bd4627Sjohansen * state. 1280*f2bd4627Sjohansen */ 1281*f2bd4627Sjohansen if ((tp != curthread) && (tp->t_waitrq == 0)) { 1282*f2bd4627Sjohansen hrtime_t curtime; 1283*f2bd4627Sjohansen 1284*f2bd4627Sjohansen curtime = gethrtime_unscaled(); 1285*f2bd4627Sjohansen (void) cpu_update_pct(tp, curtime); 1286*f2bd4627Sjohansen tp->t_waitrq = curtime; 1287*f2bd4627Sjohansen } else { 1288*f2bd4627Sjohansen (void) cpu_update_pct(tp, gethrtime_unscaled()); 1289*f2bd4627Sjohansen } 1290*f2bd4627Sjohansen 12917c478bd9Sstevel@tonic-gate dp = cp->cpu_disp; 12927c478bd9Sstevel@tonic-gate disp_lock_enter_high(&dp->disp_lock); 12937c478bd9Sstevel@tonic-gate 12947c478bd9Sstevel@tonic-gate DTRACE_SCHED3(enqueue, kthread_t *, tp, disp_t *, dp, int, 0); 12957c478bd9Sstevel@tonic-gate TRACE_3(TR_FAC_DISP, TR_BACKQ, "setbackdq:pri %d cpu %p tid %p", 12967c478bd9Sstevel@tonic-gate tpri, cp, tp); 12977c478bd9Sstevel@tonic-gate 12987c478bd9Sstevel@tonic-gate #ifndef NPROBE 12997c478bd9Sstevel@tonic-gate /* Kernel probe */ 13007c478bd9Sstevel@tonic-gate if (tnf_tracing_active) 13017c478bd9Sstevel@tonic-gate tnf_thread_queue(tp, cp, tpri); 13027c478bd9Sstevel@tonic-gate #endif /* NPROBE */ 13037c478bd9Sstevel@tonic-gate 13047c478bd9Sstevel@tonic-gate ASSERT(tpri >= 0 && tpri < dp->disp_npri); 13057c478bd9Sstevel@tonic-gate 13067c478bd9Sstevel@tonic-gate THREAD_RUN(tp, &dp->disp_lock); /* set t_state to TS_RUN */ 13077c478bd9Sstevel@tonic-gate tp->t_disp_queue = dp; 13087c478bd9Sstevel@tonic-gate tp->t_link = NULL; 13097c478bd9Sstevel@tonic-gate 13107c478bd9Sstevel@tonic-gate dq = &dp->disp_q[tpri]; 13117c478bd9Sstevel@tonic-gate dp->disp_nrunnable++; 1312685679f7Sakolb if (!bound) 1313685679f7Sakolb dp->disp_steal = 0; 13147c478bd9Sstevel@tonic-gate membar_enter(); 13157c478bd9Sstevel@tonic-gate 13167c478bd9Sstevel@tonic-gate if (dq->dq_sruncnt++ != 0) { 13177c478bd9Sstevel@tonic-gate ASSERT(dq->dq_first != NULL); 13187c478bd9Sstevel@tonic-gate dq->dq_last->t_link = tp; 13197c478bd9Sstevel@tonic-gate dq->dq_last = tp; 13207c478bd9Sstevel@tonic-gate } else { 13217c478bd9Sstevel@tonic-gate ASSERT(dq->dq_first == NULL); 13227c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last == NULL); 13237c478bd9Sstevel@tonic-gate dq->dq_first = dq->dq_last = tp; 13247c478bd9Sstevel@tonic-gate BT_SET(dp->disp_qactmap, tpri); 13257c478bd9Sstevel@tonic-gate if (tpri > dp->disp_maxrunpri) { 13267c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = tpri; 13277c478bd9Sstevel@tonic-gate membar_enter(); 13287c478bd9Sstevel@tonic-gate cpu_resched(cp, tpri); 13297c478bd9Sstevel@tonic-gate } 13307c478bd9Sstevel@tonic-gate } 13317c478bd9Sstevel@tonic-gate 13327c478bd9Sstevel@tonic-gate if (!bound && tpri > dp->disp_max_unbound_pri) { 13337c478bd9Sstevel@tonic-gate if (tp == curthread && dp->disp_max_unbound_pri == -1 && 13347c478bd9Sstevel@tonic-gate cp == CPU) { 13357c478bd9Sstevel@tonic-gate /* 13367c478bd9Sstevel@tonic-gate * If there are no other unbound threads on the 13377c478bd9Sstevel@tonic-gate * run queue, don't allow other CPUs to steal 13387c478bd9Sstevel@tonic-gate * this thread while we are in the middle of a 13397c478bd9Sstevel@tonic-gate * context switch. We may just switch to it 13407c478bd9Sstevel@tonic-gate * again right away. CPU_DISP_DONTSTEAL is cleared 13417c478bd9Sstevel@tonic-gate * in swtch and swtch_to. 13427c478bd9Sstevel@tonic-gate */ 13437c478bd9Sstevel@tonic-gate cp->cpu_disp_flags |= CPU_DISP_DONTSTEAL; 13447c478bd9Sstevel@tonic-gate } 13457c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = tpri; 13467c478bd9Sstevel@tonic-gate } 13477c478bd9Sstevel@tonic-gate (*disp_enq_thread)(cp, bound); 13487c478bd9Sstevel@tonic-gate } 13497c478bd9Sstevel@tonic-gate 13507c478bd9Sstevel@tonic-gate /* 13517c478bd9Sstevel@tonic-gate * Put the specified thread on the front of the dispatcher 13527c478bd9Sstevel@tonic-gate * queue corresponding to its current priority. 13537c478bd9Sstevel@tonic-gate * 13547c478bd9Sstevel@tonic-gate * Called with the thread in transition, onproc or stopped state 13557c478bd9Sstevel@tonic-gate * and locked (transition implies locked) and at high spl. 13567c478bd9Sstevel@tonic-gate * Returns with the thread in TS_RUN state and still locked. 13577c478bd9Sstevel@tonic-gate */ 13587c478bd9Sstevel@tonic-gate void 13597c478bd9Sstevel@tonic-gate setfrontdq(kthread_t *tp) 13607c478bd9Sstevel@tonic-gate { 13617c478bd9Sstevel@tonic-gate disp_t *dp; 13627c478bd9Sstevel@tonic-gate dispq_t *dq; 13637c478bd9Sstevel@tonic-gate cpu_t *cp; 13647c478bd9Sstevel@tonic-gate pri_t tpri; 13657c478bd9Sstevel@tonic-gate int bound; 13667c478bd9Sstevel@tonic-gate 13677c478bd9Sstevel@tonic-gate ASSERT(THREAD_LOCK_HELD(tp)); 13687c478bd9Sstevel@tonic-gate ASSERT((tp->t_schedflag & TS_ALLSTART) == 0); 13697c478bd9Sstevel@tonic-gate ASSERT(!thread_on_queue(tp)); /* make sure tp isn't on a runq */ 13707c478bd9Sstevel@tonic-gate 13717c478bd9Sstevel@tonic-gate /* 13727c478bd9Sstevel@tonic-gate * If thread is "swapped" or on the swap queue don't 13737c478bd9Sstevel@tonic-gate * queue it, but wake sched. 13747c478bd9Sstevel@tonic-gate */ 13757c478bd9Sstevel@tonic-gate if ((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD) { 13767c478bd9Sstevel@tonic-gate disp_swapped_setrun(tp); 13777c478bd9Sstevel@tonic-gate return; 13787c478bd9Sstevel@tonic-gate } 13797c478bd9Sstevel@tonic-gate 13807c478bd9Sstevel@tonic-gate tpri = DISP_PRIO(tp); 13817c478bd9Sstevel@tonic-gate if (ncpus == 1) 13827c478bd9Sstevel@tonic-gate cp = tp->t_cpu; 13837c478bd9Sstevel@tonic-gate else if (!tp->t_bound_cpu && !tp->t_weakbound_cpu) { 13847c478bd9Sstevel@tonic-gate if (tpri >= kpqpri) { 13857c478bd9Sstevel@tonic-gate setkpdq(tp, SETKP_FRONT); 13867c478bd9Sstevel@tonic-gate return; 13877c478bd9Sstevel@tonic-gate } 13887c478bd9Sstevel@tonic-gate cp = tp->t_cpu; 13897c478bd9Sstevel@tonic-gate if (tp->t_cpupart == cp->cpu_part) { 13907c478bd9Sstevel@tonic-gate /* 13917c478bd9Sstevel@tonic-gate * If we are of higher or equal priority than 13927c478bd9Sstevel@tonic-gate * the highest priority runnable thread of 13937c478bd9Sstevel@tonic-gate * the current CPU, just pick this CPU. Otherwise 13947c478bd9Sstevel@tonic-gate * Let cpu_choose() select the CPU. If this cpu 13957c478bd9Sstevel@tonic-gate * is the target of an offline request then do not 13967c478bd9Sstevel@tonic-gate * pick it - a thread_nomigrate() on the in motion 13977c478bd9Sstevel@tonic-gate * cpu relies on this when it forces a preempt. 13987c478bd9Sstevel@tonic-gate */ 13997c478bd9Sstevel@tonic-gate if (tpri < cp->cpu_disp->disp_maxrunpri || 14007c478bd9Sstevel@tonic-gate cp == cpu_inmotion) 14017c478bd9Sstevel@tonic-gate cp = cpu_choose(tp, tpri); 14027c478bd9Sstevel@tonic-gate } else { 14037c478bd9Sstevel@tonic-gate /* 14047c478bd9Sstevel@tonic-gate * Migrate to a cpu in the new partition. 14057c478bd9Sstevel@tonic-gate */ 14067c478bd9Sstevel@tonic-gate cp = disp_lowpri_cpu(tp->t_cpupart->cp_cpulist, 14077c478bd9Sstevel@tonic-gate tp->t_lpl, tp->t_pri, NULL); 14087c478bd9Sstevel@tonic-gate } 14097c478bd9Sstevel@tonic-gate bound = 0; 14107c478bd9Sstevel@tonic-gate ASSERT((cp->cpu_flags & CPU_QUIESCED) == 0); 14117c478bd9Sstevel@tonic-gate } else { 14127c478bd9Sstevel@tonic-gate /* 14137c478bd9Sstevel@tonic-gate * It is possible that t_weakbound_cpu != t_bound_cpu (for 14147c478bd9Sstevel@tonic-gate * a short time until weak binding that existed when the 14157c478bd9Sstevel@tonic-gate * strong binding was established has dropped) so we must 14167c478bd9Sstevel@tonic-gate * favour weak binding over strong. 14177c478bd9Sstevel@tonic-gate */ 14187c478bd9Sstevel@tonic-gate cp = tp->t_weakbound_cpu ? 14197c478bd9Sstevel@tonic-gate tp->t_weakbound_cpu : tp->t_bound_cpu; 14207c478bd9Sstevel@tonic-gate bound = 1; 14217c478bd9Sstevel@tonic-gate } 1422*f2bd4627Sjohansen 1423*f2bd4627Sjohansen /* 1424*f2bd4627Sjohansen * A thread that is ONPROC may be temporarily placed on the run queue 1425*f2bd4627Sjohansen * but then chosen to run again by disp. If the thread we're placing on 1426*f2bd4627Sjohansen * the queue is in TS_ONPROC state, don't set its t_waitrq until a 1427*f2bd4627Sjohansen * replacement process is actually scheduled in swtch(). In this 1428*f2bd4627Sjohansen * situation, curthread is the only thread that could be in the ONPROC 1429*f2bd4627Sjohansen * state. 1430*f2bd4627Sjohansen */ 1431*f2bd4627Sjohansen if ((tp != curthread) && (tp->t_waitrq == 0)) { 1432*f2bd4627Sjohansen hrtime_t curtime; 1433*f2bd4627Sjohansen 1434*f2bd4627Sjohansen curtime = gethrtime_unscaled(); 1435*f2bd4627Sjohansen (void) cpu_update_pct(tp, curtime); 1436*f2bd4627Sjohansen tp->t_waitrq = curtime; 1437*f2bd4627Sjohansen } else { 1438*f2bd4627Sjohansen (void) cpu_update_pct(tp, gethrtime_unscaled()); 1439*f2bd4627Sjohansen } 1440*f2bd4627Sjohansen 14417c478bd9Sstevel@tonic-gate dp = cp->cpu_disp; 14427c478bd9Sstevel@tonic-gate disp_lock_enter_high(&dp->disp_lock); 14437c478bd9Sstevel@tonic-gate 14447c478bd9Sstevel@tonic-gate TRACE_2(TR_FAC_DISP, TR_FRONTQ, "frontq:pri %d tid %p", tpri, tp); 14457c478bd9Sstevel@tonic-gate DTRACE_SCHED3(enqueue, kthread_t *, tp, disp_t *, dp, int, 1); 14467c478bd9Sstevel@tonic-gate 14477c478bd9Sstevel@tonic-gate #ifndef NPROBE 14487c478bd9Sstevel@tonic-gate /* Kernel probe */ 14497c478bd9Sstevel@tonic-gate if (tnf_tracing_active) 14507c478bd9Sstevel@tonic-gate tnf_thread_queue(tp, cp, tpri); 14517c478bd9Sstevel@tonic-gate #endif /* NPROBE */ 14527c478bd9Sstevel@tonic-gate 14537c478bd9Sstevel@tonic-gate ASSERT(tpri >= 0 && tpri < dp->disp_npri); 14547c478bd9Sstevel@tonic-gate 14557c478bd9Sstevel@tonic-gate THREAD_RUN(tp, &dp->disp_lock); /* set TS_RUN state and lock */ 14567c478bd9Sstevel@tonic-gate tp->t_disp_queue = dp; 14577c478bd9Sstevel@tonic-gate 14587c478bd9Sstevel@tonic-gate dq = &dp->disp_q[tpri]; 14597c478bd9Sstevel@tonic-gate dp->disp_nrunnable++; 1460685679f7Sakolb if (!bound) 1461685679f7Sakolb dp->disp_steal = 0; 14627c478bd9Sstevel@tonic-gate membar_enter(); 14637c478bd9Sstevel@tonic-gate 14647c478bd9Sstevel@tonic-gate if (dq->dq_sruncnt++ != 0) { 14657c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last != NULL); 14667c478bd9Sstevel@tonic-gate tp->t_link = dq->dq_first; 14677c478bd9Sstevel@tonic-gate dq->dq_first = tp; 14687c478bd9Sstevel@tonic-gate } else { 14697c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last == NULL); 14707c478bd9Sstevel@tonic-gate ASSERT(dq->dq_first == NULL); 14717c478bd9Sstevel@tonic-gate tp->t_link = NULL; 14727c478bd9Sstevel@tonic-gate dq->dq_first = dq->dq_last = tp; 14737c478bd9Sstevel@tonic-gate BT_SET(dp->disp_qactmap, tpri); 14747c478bd9Sstevel@tonic-gate if (tpri > dp->disp_maxrunpri) { 14757c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = tpri; 14767c478bd9Sstevel@tonic-gate membar_enter(); 14777c478bd9Sstevel@tonic-gate cpu_resched(cp, tpri); 14787c478bd9Sstevel@tonic-gate } 14797c478bd9Sstevel@tonic-gate } 14807c478bd9Sstevel@tonic-gate 14817c478bd9Sstevel@tonic-gate if (!bound && tpri > dp->disp_max_unbound_pri) { 14827c478bd9Sstevel@tonic-gate if (tp == curthread && dp->disp_max_unbound_pri == -1 && 14837c478bd9Sstevel@tonic-gate cp == CPU) { 14847c478bd9Sstevel@tonic-gate /* 14857c478bd9Sstevel@tonic-gate * If there are no other unbound threads on the 14867c478bd9Sstevel@tonic-gate * run queue, don't allow other CPUs to steal 14877c478bd9Sstevel@tonic-gate * this thread while we are in the middle of a 14887c478bd9Sstevel@tonic-gate * context switch. We may just switch to it 14897c478bd9Sstevel@tonic-gate * again right away. CPU_DISP_DONTSTEAL is cleared 14907c478bd9Sstevel@tonic-gate * in swtch and swtch_to. 14917c478bd9Sstevel@tonic-gate */ 14927c478bd9Sstevel@tonic-gate cp->cpu_disp_flags |= CPU_DISP_DONTSTEAL; 14937c478bd9Sstevel@tonic-gate } 14947c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = tpri; 14957c478bd9Sstevel@tonic-gate } 14967c478bd9Sstevel@tonic-gate (*disp_enq_thread)(cp, bound); 14977c478bd9Sstevel@tonic-gate } 14987c478bd9Sstevel@tonic-gate 14997c478bd9Sstevel@tonic-gate /* 15007c478bd9Sstevel@tonic-gate * Put a high-priority unbound thread on the kp queue 15017c478bd9Sstevel@tonic-gate */ 15027c478bd9Sstevel@tonic-gate static void 15037c478bd9Sstevel@tonic-gate setkpdq(kthread_t *tp, int borf) 15047c478bd9Sstevel@tonic-gate { 15057c478bd9Sstevel@tonic-gate dispq_t *dq; 15067c478bd9Sstevel@tonic-gate disp_t *dp; 15077c478bd9Sstevel@tonic-gate cpu_t *cp; 15087c478bd9Sstevel@tonic-gate pri_t tpri; 15097c478bd9Sstevel@tonic-gate 15107c478bd9Sstevel@tonic-gate tpri = DISP_PRIO(tp); 15117c478bd9Sstevel@tonic-gate 15127c478bd9Sstevel@tonic-gate dp = &tp->t_cpupart->cp_kp_queue; 15137c478bd9Sstevel@tonic-gate disp_lock_enter_high(&dp->disp_lock); 15147c478bd9Sstevel@tonic-gate 15157c478bd9Sstevel@tonic-gate TRACE_2(TR_FAC_DISP, TR_FRONTQ, "frontq:pri %d tid %p", tpri, tp); 15167c478bd9Sstevel@tonic-gate 15177c478bd9Sstevel@tonic-gate ASSERT(tpri >= 0 && tpri < dp->disp_npri); 15187c478bd9Sstevel@tonic-gate DTRACE_SCHED3(enqueue, kthread_t *, tp, disp_t *, dp, int, borf); 15197c478bd9Sstevel@tonic-gate THREAD_RUN(tp, &dp->disp_lock); /* set t_state to TS_RUN */ 15207c478bd9Sstevel@tonic-gate tp->t_disp_queue = dp; 15217c478bd9Sstevel@tonic-gate dp->disp_nrunnable++; 15227c478bd9Sstevel@tonic-gate dq = &dp->disp_q[tpri]; 15237c478bd9Sstevel@tonic-gate 15247c478bd9Sstevel@tonic-gate if (dq->dq_sruncnt++ != 0) { 15257c478bd9Sstevel@tonic-gate if (borf == SETKP_BACK) { 15267c478bd9Sstevel@tonic-gate ASSERT(dq->dq_first != NULL); 15277c478bd9Sstevel@tonic-gate tp->t_link = NULL; 15287c478bd9Sstevel@tonic-gate dq->dq_last->t_link = tp; 15297c478bd9Sstevel@tonic-gate dq->dq_last = tp; 15307c478bd9Sstevel@tonic-gate } else { 15317c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last != NULL); 15327c478bd9Sstevel@tonic-gate tp->t_link = dq->dq_first; 15337c478bd9Sstevel@tonic-gate dq->dq_first = tp; 15347c478bd9Sstevel@tonic-gate } 15357c478bd9Sstevel@tonic-gate } else { 15367c478bd9Sstevel@tonic-gate if (borf == SETKP_BACK) { 15377c478bd9Sstevel@tonic-gate ASSERT(dq->dq_first == NULL); 15387c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last == NULL); 15397c478bd9Sstevel@tonic-gate dq->dq_first = dq->dq_last = tp; 15407c478bd9Sstevel@tonic-gate } else { 15417c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last == NULL); 15427c478bd9Sstevel@tonic-gate ASSERT(dq->dq_first == NULL); 15437c478bd9Sstevel@tonic-gate tp->t_link = NULL; 15447c478bd9Sstevel@tonic-gate dq->dq_first = dq->dq_last = tp; 15457c478bd9Sstevel@tonic-gate } 15467c478bd9Sstevel@tonic-gate BT_SET(dp->disp_qactmap, tpri); 15477c478bd9Sstevel@tonic-gate if (tpri > dp->disp_max_unbound_pri) 15487c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = tpri; 15497c478bd9Sstevel@tonic-gate if (tpri > dp->disp_maxrunpri) { 15507c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = tpri; 15517c478bd9Sstevel@tonic-gate membar_enter(); 15527c478bd9Sstevel@tonic-gate } 15537c478bd9Sstevel@tonic-gate } 15547c478bd9Sstevel@tonic-gate 15557c478bd9Sstevel@tonic-gate cp = tp->t_cpu; 15567c478bd9Sstevel@tonic-gate if (tp->t_cpupart != cp->cpu_part) { 15577c478bd9Sstevel@tonic-gate /* migrate to a cpu in the new partition */ 15587c478bd9Sstevel@tonic-gate cp = tp->t_cpupart->cp_cpulist; 15597c478bd9Sstevel@tonic-gate } 15607c478bd9Sstevel@tonic-gate cp = disp_lowpri_cpu(cp, tp->t_lpl, tp->t_pri, NULL); 15617c478bd9Sstevel@tonic-gate disp_lock_enter_high(&cp->cpu_disp->disp_lock); 15627c478bd9Sstevel@tonic-gate ASSERT((cp->cpu_flags & CPU_QUIESCED) == 0); 15637c478bd9Sstevel@tonic-gate 15647c478bd9Sstevel@tonic-gate #ifndef NPROBE 15657c478bd9Sstevel@tonic-gate /* Kernel probe */ 15667c478bd9Sstevel@tonic-gate if (tnf_tracing_active) 15677c478bd9Sstevel@tonic-gate tnf_thread_queue(tp, cp, tpri); 15687c478bd9Sstevel@tonic-gate #endif /* NPROBE */ 15697c478bd9Sstevel@tonic-gate 15707c478bd9Sstevel@tonic-gate if (cp->cpu_chosen_level < tpri) 15717c478bd9Sstevel@tonic-gate cp->cpu_chosen_level = tpri; 15727c478bd9Sstevel@tonic-gate cpu_resched(cp, tpri); 15737c478bd9Sstevel@tonic-gate disp_lock_exit_high(&cp->cpu_disp->disp_lock); 15747c478bd9Sstevel@tonic-gate (*disp_enq_thread)(cp, 0); 15757c478bd9Sstevel@tonic-gate } 15767c478bd9Sstevel@tonic-gate 15777c478bd9Sstevel@tonic-gate /* 15787c478bd9Sstevel@tonic-gate * Remove a thread from the dispatcher queue if it is on it. 15797c478bd9Sstevel@tonic-gate * It is not an error if it is not found but we return whether 15807c478bd9Sstevel@tonic-gate * or not it was found in case the caller wants to check. 15817c478bd9Sstevel@tonic-gate */ 15827c478bd9Sstevel@tonic-gate int 15837c478bd9Sstevel@tonic-gate dispdeq(kthread_t *tp) 15847c478bd9Sstevel@tonic-gate { 15857c478bd9Sstevel@tonic-gate disp_t *dp; 15867c478bd9Sstevel@tonic-gate dispq_t *dq; 15877c478bd9Sstevel@tonic-gate kthread_t *rp; 15887c478bd9Sstevel@tonic-gate kthread_t *trp; 15897c478bd9Sstevel@tonic-gate kthread_t **ptp; 15907c478bd9Sstevel@tonic-gate int tpri; 15917c478bd9Sstevel@tonic-gate 15927c478bd9Sstevel@tonic-gate ASSERT(THREAD_LOCK_HELD(tp)); 15937c478bd9Sstevel@tonic-gate 15947c478bd9Sstevel@tonic-gate if (tp->t_state != TS_RUN) 15957c478bd9Sstevel@tonic-gate return (0); 15967c478bd9Sstevel@tonic-gate 15977c478bd9Sstevel@tonic-gate /* 15987c478bd9Sstevel@tonic-gate * The thread is "swapped" or is on the swap queue and 15997c478bd9Sstevel@tonic-gate * hence no longer on the run queue, so return true. 16007c478bd9Sstevel@tonic-gate */ 16017c478bd9Sstevel@tonic-gate if ((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD) 16027c478bd9Sstevel@tonic-gate return (1); 16037c478bd9Sstevel@tonic-gate 16047c478bd9Sstevel@tonic-gate tpri = DISP_PRIO(tp); 16057c478bd9Sstevel@tonic-gate dp = tp->t_disp_queue; 16067c478bd9Sstevel@tonic-gate ASSERT(tpri < dp->disp_npri); 16077c478bd9Sstevel@tonic-gate dq = &dp->disp_q[tpri]; 16087c478bd9Sstevel@tonic-gate ptp = &dq->dq_first; 16097c478bd9Sstevel@tonic-gate rp = *ptp; 16107c478bd9Sstevel@tonic-gate trp = NULL; 16117c478bd9Sstevel@tonic-gate 16127c478bd9Sstevel@tonic-gate ASSERT(dq->dq_last == NULL || dq->dq_last->t_link == NULL); 16137c478bd9Sstevel@tonic-gate 16147c478bd9Sstevel@tonic-gate /* 16157c478bd9Sstevel@tonic-gate * Search for thread in queue. 16167c478bd9Sstevel@tonic-gate * Double links would simplify this at the expense of disp/setrun. 16177c478bd9Sstevel@tonic-gate */ 16187c478bd9Sstevel@tonic-gate while (rp != tp && rp != NULL) { 16197c478bd9Sstevel@tonic-gate trp = rp; 16207c478bd9Sstevel@tonic-gate ptp = &trp->t_link; 16217c478bd9Sstevel@tonic-gate rp = trp->t_link; 16227c478bd9Sstevel@tonic-gate } 16237c478bd9Sstevel@tonic-gate 16247c478bd9Sstevel@tonic-gate if (rp == NULL) { 16257c478bd9Sstevel@tonic-gate panic("dispdeq: thread not on queue"); 16267c478bd9Sstevel@tonic-gate } 16277c478bd9Sstevel@tonic-gate 16287c478bd9Sstevel@tonic-gate DTRACE_SCHED2(dequeue, kthread_t *, tp, disp_t *, dp); 16297c478bd9Sstevel@tonic-gate 16307c478bd9Sstevel@tonic-gate /* 16317c478bd9Sstevel@tonic-gate * Found it so remove it from queue. 16327c478bd9Sstevel@tonic-gate */ 16337c478bd9Sstevel@tonic-gate if ((*ptp = rp->t_link) == NULL) 16347c478bd9Sstevel@tonic-gate dq->dq_last = trp; 16357c478bd9Sstevel@tonic-gate 16367c478bd9Sstevel@tonic-gate dp->disp_nrunnable--; 16377c478bd9Sstevel@tonic-gate if (--dq->dq_sruncnt == 0) { 16387c478bd9Sstevel@tonic-gate dp->disp_qactmap[tpri >> BT_ULSHIFT] &= ~BT_BIW(tpri); 16397c478bd9Sstevel@tonic-gate if (dp->disp_nrunnable == 0) { 16407c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = -1; 16417c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = -1; 16427c478bd9Sstevel@tonic-gate } else if (tpri == dp->disp_maxrunpri) { 16437c478bd9Sstevel@tonic-gate int ipri; 16447c478bd9Sstevel@tonic-gate 16457c478bd9Sstevel@tonic-gate ipri = bt_gethighbit(dp->disp_qactmap, 16467c478bd9Sstevel@tonic-gate dp->disp_maxrunpri >> BT_ULSHIFT); 16477c478bd9Sstevel@tonic-gate if (ipri < dp->disp_max_unbound_pri) 16487c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = ipri; 16497c478bd9Sstevel@tonic-gate dp->disp_maxrunpri = ipri; 16507c478bd9Sstevel@tonic-gate } 16517c478bd9Sstevel@tonic-gate } 16527c478bd9Sstevel@tonic-gate tp->t_link = NULL; 16537c478bd9Sstevel@tonic-gate THREAD_TRANSITION(tp); /* put in intermediate state */ 16547c478bd9Sstevel@tonic-gate return (1); 16557c478bd9Sstevel@tonic-gate } 16567c478bd9Sstevel@tonic-gate 16577c478bd9Sstevel@tonic-gate 16587c478bd9Sstevel@tonic-gate /* 16597c478bd9Sstevel@tonic-gate * dq_sruninc and dq_srundec are public functions for 16607c478bd9Sstevel@tonic-gate * incrementing/decrementing the sruncnts when a thread on 16617c478bd9Sstevel@tonic-gate * a dispatcher queue is made schedulable/unschedulable by 16627c478bd9Sstevel@tonic-gate * resetting the TS_LOAD flag. 16637c478bd9Sstevel@tonic-gate * 16647c478bd9Sstevel@tonic-gate * The caller MUST have the thread lock and therefore the dispatcher 16657c478bd9Sstevel@tonic-gate * queue lock so that the operation which changes 16667c478bd9Sstevel@tonic-gate * the flag, the operation that checks the status of the thread to 16677c478bd9Sstevel@tonic-gate * determine if it's on a disp queue AND the call to this function 16687c478bd9Sstevel@tonic-gate * are one atomic operation with respect to interrupts. 16697c478bd9Sstevel@tonic-gate */ 16707c478bd9Sstevel@tonic-gate 16717c478bd9Sstevel@tonic-gate /* 16727c478bd9Sstevel@tonic-gate * Called by sched AFTER TS_LOAD flag is set on a swapped, runnable thread. 16737c478bd9Sstevel@tonic-gate */ 16747c478bd9Sstevel@tonic-gate void 16757c478bd9Sstevel@tonic-gate dq_sruninc(kthread_t *t) 16767c478bd9Sstevel@tonic-gate { 16777c478bd9Sstevel@tonic-gate ASSERT(t->t_state == TS_RUN); 16787c478bd9Sstevel@tonic-gate ASSERT(t->t_schedflag & TS_LOAD); 16797c478bd9Sstevel@tonic-gate 16807c478bd9Sstevel@tonic-gate THREAD_TRANSITION(t); 16817c478bd9Sstevel@tonic-gate setfrontdq(t); 16827c478bd9Sstevel@tonic-gate } 16837c478bd9Sstevel@tonic-gate 16847c478bd9Sstevel@tonic-gate /* 16857c478bd9Sstevel@tonic-gate * See comment on calling conventions above. 16867c478bd9Sstevel@tonic-gate * Called by sched BEFORE TS_LOAD flag is cleared on a runnable thread. 16877c478bd9Sstevel@tonic-gate */ 16887c478bd9Sstevel@tonic-gate void 16897c478bd9Sstevel@tonic-gate dq_srundec(kthread_t *t) 16907c478bd9Sstevel@tonic-gate { 16917c478bd9Sstevel@tonic-gate ASSERT(t->t_schedflag & TS_LOAD); 16927c478bd9Sstevel@tonic-gate 16937c478bd9Sstevel@tonic-gate (void) dispdeq(t); 16947c478bd9Sstevel@tonic-gate disp_swapped_enq(t); 16957c478bd9Sstevel@tonic-gate } 16967c478bd9Sstevel@tonic-gate 16977c478bd9Sstevel@tonic-gate /* 16987c478bd9Sstevel@tonic-gate * Change the dispatcher lock of thread to the "swapped_lock" 16997c478bd9Sstevel@tonic-gate * and return with thread lock still held. 17007c478bd9Sstevel@tonic-gate * 17017c478bd9Sstevel@tonic-gate * Called with thread_lock held, in transition state, and at high spl. 17027c478bd9Sstevel@tonic-gate */ 17037c478bd9Sstevel@tonic-gate void 17047c478bd9Sstevel@tonic-gate disp_swapped_enq(kthread_t *tp) 17057c478bd9Sstevel@tonic-gate { 17067c478bd9Sstevel@tonic-gate ASSERT(THREAD_LOCK_HELD(tp)); 17077c478bd9Sstevel@tonic-gate ASSERT(tp->t_schedflag & TS_LOAD); 17087c478bd9Sstevel@tonic-gate 17097c478bd9Sstevel@tonic-gate switch (tp->t_state) { 17107c478bd9Sstevel@tonic-gate case TS_RUN: 17117c478bd9Sstevel@tonic-gate disp_lock_enter_high(&swapped_lock); 17127c478bd9Sstevel@tonic-gate THREAD_SWAP(tp, &swapped_lock); /* set TS_RUN state and lock */ 17137c478bd9Sstevel@tonic-gate break; 17147c478bd9Sstevel@tonic-gate case TS_ONPROC: 17157c478bd9Sstevel@tonic-gate disp_lock_enter_high(&swapped_lock); 17167c478bd9Sstevel@tonic-gate THREAD_TRANSITION(tp); 17177c478bd9Sstevel@tonic-gate wake_sched_sec = 1; /* tell clock to wake sched */ 17187c478bd9Sstevel@tonic-gate THREAD_SWAP(tp, &swapped_lock); /* set TS_RUN state and lock */ 17197c478bd9Sstevel@tonic-gate break; 17207c478bd9Sstevel@tonic-gate default: 17217c478bd9Sstevel@tonic-gate panic("disp_swapped: tp: %p bad t_state", (void *)tp); 17227c478bd9Sstevel@tonic-gate } 17237c478bd9Sstevel@tonic-gate } 17247c478bd9Sstevel@tonic-gate 17257c478bd9Sstevel@tonic-gate /* 17267c478bd9Sstevel@tonic-gate * This routine is called by setbackdq/setfrontdq if the thread is 17277c478bd9Sstevel@tonic-gate * not loaded or loaded and on the swap queue. 17287c478bd9Sstevel@tonic-gate * 17297c478bd9Sstevel@tonic-gate * Thread state TS_SLEEP implies that a swapped thread 17307c478bd9Sstevel@tonic-gate * has been woken up and needs to be swapped in by the swapper. 17317c478bd9Sstevel@tonic-gate * 17327c478bd9Sstevel@tonic-gate * Thread state TS_RUN, it implies that the priority of a swapped 17337c478bd9Sstevel@tonic-gate * thread is being increased by scheduling class (e.g. ts_update). 17347c478bd9Sstevel@tonic-gate */ 17357c478bd9Sstevel@tonic-gate static void 17367c478bd9Sstevel@tonic-gate disp_swapped_setrun(kthread_t *tp) 17377c478bd9Sstevel@tonic-gate { 17387c478bd9Sstevel@tonic-gate ASSERT(THREAD_LOCK_HELD(tp)); 17397c478bd9Sstevel@tonic-gate ASSERT((tp->t_schedflag & (TS_LOAD | TS_ON_SWAPQ)) != TS_LOAD); 17407c478bd9Sstevel@tonic-gate 17417c478bd9Sstevel@tonic-gate switch (tp->t_state) { 17427c478bd9Sstevel@tonic-gate case TS_SLEEP: 17437c478bd9Sstevel@tonic-gate disp_lock_enter_high(&swapped_lock); 17447c478bd9Sstevel@tonic-gate /* 17457c478bd9Sstevel@tonic-gate * Wakeup sched immediately (i.e., next tick) if the 17467c478bd9Sstevel@tonic-gate * thread priority is above maxclsyspri. 17477c478bd9Sstevel@tonic-gate */ 17487c478bd9Sstevel@tonic-gate if (DISP_PRIO(tp) > maxclsyspri) 17497c478bd9Sstevel@tonic-gate wake_sched = 1; 17507c478bd9Sstevel@tonic-gate else 17517c478bd9Sstevel@tonic-gate wake_sched_sec = 1; 17527c478bd9Sstevel@tonic-gate THREAD_RUN(tp, &swapped_lock); /* set TS_RUN state and lock */ 17537c478bd9Sstevel@tonic-gate break; 17547c478bd9Sstevel@tonic-gate case TS_RUN: /* called from ts_update */ 17557c478bd9Sstevel@tonic-gate break; 17567c478bd9Sstevel@tonic-gate default: 17577c478bd9Sstevel@tonic-gate panic("disp_swapped_setrun: tp: %p bad t_state", tp); 17587c478bd9Sstevel@tonic-gate } 17597c478bd9Sstevel@tonic-gate } 17607c478bd9Sstevel@tonic-gate 17617c478bd9Sstevel@tonic-gate 17627c478bd9Sstevel@tonic-gate /* 17637c478bd9Sstevel@tonic-gate * Make a thread give up its processor. Find the processor on 17647c478bd9Sstevel@tonic-gate * which this thread is executing, and have that processor 17657c478bd9Sstevel@tonic-gate * preempt. 17667c478bd9Sstevel@tonic-gate */ 17677c478bd9Sstevel@tonic-gate void 17687c478bd9Sstevel@tonic-gate cpu_surrender(kthread_t *tp) 17697c478bd9Sstevel@tonic-gate { 17707c478bd9Sstevel@tonic-gate cpu_t *cpup; 17717c478bd9Sstevel@tonic-gate int max_pri; 17727c478bd9Sstevel@tonic-gate int max_run_pri; 17737c478bd9Sstevel@tonic-gate klwp_t *lwp; 17747c478bd9Sstevel@tonic-gate 17757c478bd9Sstevel@tonic-gate ASSERT(THREAD_LOCK_HELD(tp)); 17767c478bd9Sstevel@tonic-gate 17777c478bd9Sstevel@tonic-gate if (tp->t_state != TS_ONPROC) 17787c478bd9Sstevel@tonic-gate return; 17797c478bd9Sstevel@tonic-gate cpup = tp->t_disp_queue->disp_cpu; /* CPU thread dispatched to */ 17807c478bd9Sstevel@tonic-gate max_pri = cpup->cpu_disp->disp_maxrunpri; /* best pri of that CPU */ 17817c478bd9Sstevel@tonic-gate max_run_pri = CP_MAXRUNPRI(cpup->cpu_part); 17827c478bd9Sstevel@tonic-gate if (max_pri < max_run_pri) 17837c478bd9Sstevel@tonic-gate max_pri = max_run_pri; 17847c478bd9Sstevel@tonic-gate 17857c478bd9Sstevel@tonic-gate cpup->cpu_runrun = 1; 17867c478bd9Sstevel@tonic-gate if (max_pri >= kpreemptpri && cpup->cpu_kprunrun == 0) { 17877c478bd9Sstevel@tonic-gate cpup->cpu_kprunrun = 1; 17887c478bd9Sstevel@tonic-gate } 17897c478bd9Sstevel@tonic-gate 17907c478bd9Sstevel@tonic-gate /* 17917c478bd9Sstevel@tonic-gate * Propagate cpu_runrun, and cpu_kprunrun to global visibility. 17927c478bd9Sstevel@tonic-gate */ 17937c478bd9Sstevel@tonic-gate membar_enter(); 17947c478bd9Sstevel@tonic-gate 17957c478bd9Sstevel@tonic-gate DTRACE_SCHED1(surrender, kthread_t *, tp); 17967c478bd9Sstevel@tonic-gate 17977c478bd9Sstevel@tonic-gate /* 17987c478bd9Sstevel@tonic-gate * Make the target thread take an excursion through trap() 17997c478bd9Sstevel@tonic-gate * to do preempt() (unless we're already in trap or post_syscall, 18007c478bd9Sstevel@tonic-gate * calling cpu_surrender via CL_TRAPRET). 18017c478bd9Sstevel@tonic-gate */ 18027c478bd9Sstevel@tonic-gate if (tp != curthread || (lwp = tp->t_lwp) == NULL || 18037c478bd9Sstevel@tonic-gate lwp->lwp_state != LWP_USER) { 18047c478bd9Sstevel@tonic-gate aston(tp); 18057c478bd9Sstevel@tonic-gate if (cpup != CPU) 18067c478bd9Sstevel@tonic-gate poke_cpu(cpup->cpu_id); 18077c478bd9Sstevel@tonic-gate } 18087c478bd9Sstevel@tonic-gate TRACE_2(TR_FAC_DISP, TR_CPU_SURRENDER, 18097c478bd9Sstevel@tonic-gate "cpu_surrender:tid %p cpu %p", tp, cpup); 18107c478bd9Sstevel@tonic-gate } 18117c478bd9Sstevel@tonic-gate 18127c478bd9Sstevel@tonic-gate 18137c478bd9Sstevel@tonic-gate /* 18147c478bd9Sstevel@tonic-gate * Commit to and ratify a scheduling decision 18157c478bd9Sstevel@tonic-gate */ 18167c478bd9Sstevel@tonic-gate /*ARGSUSED*/ 18177c478bd9Sstevel@tonic-gate static kthread_t * 18187c478bd9Sstevel@tonic-gate disp_ratify(kthread_t *tp, disp_t *kpq) 18197c478bd9Sstevel@tonic-gate { 18207c478bd9Sstevel@tonic-gate pri_t tpri, maxpri; 18217c478bd9Sstevel@tonic-gate pri_t maxkpri; 18227c478bd9Sstevel@tonic-gate cpu_t *cpup; 18237c478bd9Sstevel@tonic-gate 18247c478bd9Sstevel@tonic-gate ASSERT(tp != NULL); 18257c478bd9Sstevel@tonic-gate /* 18267c478bd9Sstevel@tonic-gate * Commit to, then ratify scheduling decision 18277c478bd9Sstevel@tonic-gate */ 18287c478bd9Sstevel@tonic-gate cpup = CPU; 18297c478bd9Sstevel@tonic-gate if (cpup->cpu_runrun != 0) 18307c478bd9Sstevel@tonic-gate cpup->cpu_runrun = 0; 18317c478bd9Sstevel@tonic-gate if (cpup->cpu_kprunrun != 0) 18327c478bd9Sstevel@tonic-gate cpup->cpu_kprunrun = 0; 18337c478bd9Sstevel@tonic-gate if (cpup->cpu_chosen_level != -1) 18347c478bd9Sstevel@tonic-gate cpup->cpu_chosen_level = -1; 18357c478bd9Sstevel@tonic-gate membar_enter(); 18367c478bd9Sstevel@tonic-gate tpri = DISP_PRIO(tp); 18377c478bd9Sstevel@tonic-gate maxpri = cpup->cpu_disp->disp_maxrunpri; 18387c478bd9Sstevel@tonic-gate maxkpri = kpq->disp_maxrunpri; 18397c478bd9Sstevel@tonic-gate if (maxpri < maxkpri) 18407c478bd9Sstevel@tonic-gate maxpri = maxkpri; 18417c478bd9Sstevel@tonic-gate if (tpri < maxpri) { 18427c478bd9Sstevel@tonic-gate /* 18437c478bd9Sstevel@tonic-gate * should have done better 18447c478bd9Sstevel@tonic-gate * put this one back and indicate to try again 18457c478bd9Sstevel@tonic-gate */ 18467c478bd9Sstevel@tonic-gate cpup->cpu_dispthread = curthread; /* fixup dispthread */ 18477c478bd9Sstevel@tonic-gate cpup->cpu_dispatch_pri = DISP_PRIO(curthread); 18487c478bd9Sstevel@tonic-gate thread_lock_high(tp); 18497c478bd9Sstevel@tonic-gate THREAD_TRANSITION(tp); 18507c478bd9Sstevel@tonic-gate setfrontdq(tp); 18517c478bd9Sstevel@tonic-gate thread_unlock_nopreempt(tp); 18527c478bd9Sstevel@tonic-gate 18537c478bd9Sstevel@tonic-gate tp = NULL; 18547c478bd9Sstevel@tonic-gate } 18557c478bd9Sstevel@tonic-gate return (tp); 18567c478bd9Sstevel@tonic-gate } 18577c478bd9Sstevel@tonic-gate 18587c478bd9Sstevel@tonic-gate /* 18597c478bd9Sstevel@tonic-gate * See if there is any work on the dispatcher queue for other CPUs. 18607c478bd9Sstevel@tonic-gate * If there is, dequeue the best thread and return. 18617c478bd9Sstevel@tonic-gate */ 18627c478bd9Sstevel@tonic-gate static kthread_t * 18637c478bd9Sstevel@tonic-gate disp_getwork(cpu_t *cp) 18647c478bd9Sstevel@tonic-gate { 18657c478bd9Sstevel@tonic-gate cpu_t *ocp; /* other CPU */ 18667c478bd9Sstevel@tonic-gate cpu_t *ocp_start; 18677c478bd9Sstevel@tonic-gate cpu_t *tcp; /* target local CPU */ 18687c478bd9Sstevel@tonic-gate kthread_t *tp; 1869685679f7Sakolb kthread_t *retval = NULL; 18707c478bd9Sstevel@tonic-gate pri_t maxpri; 18717c478bd9Sstevel@tonic-gate disp_t *kpq; /* kp queue for this partition */ 18727c478bd9Sstevel@tonic-gate lpl_t *lpl, *lpl_leaf; 18737c478bd9Sstevel@tonic-gate int hint, leafidx; 1874685679f7Sakolb hrtime_t stealtime; 18757c478bd9Sstevel@tonic-gate 18767c478bd9Sstevel@tonic-gate maxpri = -1; 18777c478bd9Sstevel@tonic-gate tcp = NULL; 18787c478bd9Sstevel@tonic-gate 18797c478bd9Sstevel@tonic-gate kpq = &cp->cpu_part->cp_kp_queue; 18807c478bd9Sstevel@tonic-gate while (kpq->disp_maxrunpri >= 0) { 18817c478bd9Sstevel@tonic-gate /* 18827c478bd9Sstevel@tonic-gate * Try to take a thread from the kp_queue. 18837c478bd9Sstevel@tonic-gate */ 18847c478bd9Sstevel@tonic-gate tp = (disp_getbest(kpq)); 18857c478bd9Sstevel@tonic-gate if (tp) 18867c478bd9Sstevel@tonic-gate return (disp_ratify(tp, kpq)); 18877c478bd9Sstevel@tonic-gate } 18887c478bd9Sstevel@tonic-gate 1889ab761399Sesaxe kpreempt_disable(); /* protect the cpu_active list */ 18907c478bd9Sstevel@tonic-gate 18917c478bd9Sstevel@tonic-gate /* 18927c478bd9Sstevel@tonic-gate * Try to find something to do on another CPU's run queue. 18937c478bd9Sstevel@tonic-gate * Loop through all other CPUs looking for the one with the highest 18947c478bd9Sstevel@tonic-gate * priority unbound thread. 18957c478bd9Sstevel@tonic-gate * 18967c478bd9Sstevel@tonic-gate * On NUMA machines, the partition's CPUs are consulted in order of 18977c478bd9Sstevel@tonic-gate * distance from the current CPU. This way, the first available 18987c478bd9Sstevel@tonic-gate * work found is also the closest, and will suffer the least 18997c478bd9Sstevel@tonic-gate * from being migrated. 19007c478bd9Sstevel@tonic-gate */ 19017c478bd9Sstevel@tonic-gate lpl = lpl_leaf = cp->cpu_lpl; 19027c478bd9Sstevel@tonic-gate hint = leafidx = 0; 19037c478bd9Sstevel@tonic-gate 19047c478bd9Sstevel@tonic-gate /* 19057c478bd9Sstevel@tonic-gate * This loop traverses the lpl hierarchy. Higher level lpls represent 19067c478bd9Sstevel@tonic-gate * broader levels of locality 19077c478bd9Sstevel@tonic-gate */ 19087c478bd9Sstevel@tonic-gate do { 19097c478bd9Sstevel@tonic-gate /* This loop iterates over the lpl's leaves */ 19107c478bd9Sstevel@tonic-gate do { 19117c478bd9Sstevel@tonic-gate if (lpl_leaf != cp->cpu_lpl) 19127c478bd9Sstevel@tonic-gate ocp = lpl_leaf->lpl_cpus; 19137c478bd9Sstevel@tonic-gate else 19147c478bd9Sstevel@tonic-gate ocp = cp->cpu_next_lpl; 19157c478bd9Sstevel@tonic-gate 19167c478bd9Sstevel@tonic-gate /* This loop iterates over the CPUs in the leaf */ 19177c478bd9Sstevel@tonic-gate ocp_start = ocp; 19187c478bd9Sstevel@tonic-gate do { 19197c478bd9Sstevel@tonic-gate pri_t pri; 19207c478bd9Sstevel@tonic-gate 19217c478bd9Sstevel@tonic-gate ASSERT(CPU_ACTIVE(ocp)); 19227c478bd9Sstevel@tonic-gate 19237c478bd9Sstevel@tonic-gate /* 19247c478bd9Sstevel@tonic-gate * End our stroll around the partition if: 19257c478bd9Sstevel@tonic-gate * 19267c478bd9Sstevel@tonic-gate * - Something became runnable on the local 19277c478bd9Sstevel@tonic-gate * queue 19287c478bd9Sstevel@tonic-gate * 19297c478bd9Sstevel@tonic-gate * - We're at the broadest level of locality and 19307c478bd9Sstevel@tonic-gate * we happen across another idle CPU. At the 19317c478bd9Sstevel@tonic-gate * highest level of locality, all CPUs will 19327c478bd9Sstevel@tonic-gate * walk the partition's CPUs in the same 19337c478bd9Sstevel@tonic-gate * order, so we can end our stroll taking 19347c478bd9Sstevel@tonic-gate * comfort in knowing the other idle CPU is 19357c478bd9Sstevel@tonic-gate * already covering the next portion of the 19367c478bd9Sstevel@tonic-gate * list. 19377c478bd9Sstevel@tonic-gate */ 19387c478bd9Sstevel@tonic-gate if (cp->cpu_disp->disp_nrunnable != 0) 19397c478bd9Sstevel@tonic-gate break; 19407c478bd9Sstevel@tonic-gate if (ocp->cpu_dispatch_pri == -1) { 19417c478bd9Sstevel@tonic-gate if (ocp->cpu_disp_flags & 19427c478bd9Sstevel@tonic-gate CPU_DISP_HALTED) 19437c478bd9Sstevel@tonic-gate continue; 19447c478bd9Sstevel@tonic-gate else if (lpl->lpl_parent == NULL) 19457c478bd9Sstevel@tonic-gate break; 19467c478bd9Sstevel@tonic-gate } 19477c478bd9Sstevel@tonic-gate 19487c478bd9Sstevel@tonic-gate /* 19497c478bd9Sstevel@tonic-gate * If there's only one thread and the CPU 19507c478bd9Sstevel@tonic-gate * is in the middle of a context switch, 19517c478bd9Sstevel@tonic-gate * or it's currently running the idle thread, 19527c478bd9Sstevel@tonic-gate * don't steal it. 19537c478bd9Sstevel@tonic-gate */ 19547c478bd9Sstevel@tonic-gate if ((ocp->cpu_disp_flags & 19557c478bd9Sstevel@tonic-gate CPU_DISP_DONTSTEAL) && 19567c478bd9Sstevel@tonic-gate ocp->cpu_disp->disp_nrunnable == 1) 19577c478bd9Sstevel@tonic-gate continue; 19587c478bd9Sstevel@tonic-gate 19597c478bd9Sstevel@tonic-gate pri = ocp->cpu_disp->disp_max_unbound_pri; 19607c478bd9Sstevel@tonic-gate if (pri > maxpri) { 1961685679f7Sakolb /* 1962685679f7Sakolb * Don't steal threads that we attempted 1963685679f7Sakolb * to be stolen very recently until 1964685679f7Sakolb * they're ready to be stolen again. 1965685679f7Sakolb */ 1966685679f7Sakolb stealtime = ocp->cpu_disp->disp_steal; 1967685679f7Sakolb if (stealtime == 0 || 1968685679f7Sakolb stealtime - gethrtime() <= 0) { 1969685679f7Sakolb maxpri = pri; 1970685679f7Sakolb tcp = ocp; 1971685679f7Sakolb } else { 1972685679f7Sakolb /* 1973685679f7Sakolb * Don't update tcp, just set 1974685679f7Sakolb * the retval to T_DONTSTEAL, so 1975685679f7Sakolb * that if no acceptable CPUs 1976685679f7Sakolb * are found the return value 1977685679f7Sakolb * will be T_DONTSTEAL rather 1978685679f7Sakolb * then NULL. 1979685679f7Sakolb */ 1980685679f7Sakolb retval = T_DONTSTEAL; 1981685679f7Sakolb } 19827c478bd9Sstevel@tonic-gate } 19837c478bd9Sstevel@tonic-gate } while ((ocp = ocp->cpu_next_lpl) != ocp_start); 19847c478bd9Sstevel@tonic-gate 19857c478bd9Sstevel@tonic-gate if ((lpl_leaf = lpl->lpl_rset[++leafidx]) == NULL) { 19867c478bd9Sstevel@tonic-gate leafidx = 0; 19877c478bd9Sstevel@tonic-gate lpl_leaf = lpl->lpl_rset[leafidx]; 19887c478bd9Sstevel@tonic-gate } 19897c478bd9Sstevel@tonic-gate } while (leafidx != hint); 19907c478bd9Sstevel@tonic-gate 19917c478bd9Sstevel@tonic-gate hint = leafidx = lpl->lpl_hint; 19927c478bd9Sstevel@tonic-gate if ((lpl = lpl->lpl_parent) != NULL) 19937c478bd9Sstevel@tonic-gate lpl_leaf = lpl->lpl_rset[hint]; 19947c478bd9Sstevel@tonic-gate } while (!tcp && lpl); 19957c478bd9Sstevel@tonic-gate 1996ab761399Sesaxe kpreempt_enable(); 19977c478bd9Sstevel@tonic-gate 19987c478bd9Sstevel@tonic-gate /* 19997c478bd9Sstevel@tonic-gate * If another queue looks good, and there is still nothing on 20007c478bd9Sstevel@tonic-gate * the local queue, try to transfer one or more threads 20017c478bd9Sstevel@tonic-gate * from it to our queue. 20027c478bd9Sstevel@tonic-gate */ 20037c478bd9Sstevel@tonic-gate if (tcp && cp->cpu_disp->disp_nrunnable == 0) { 2004685679f7Sakolb tp = disp_getbest(tcp->cpu_disp); 2005685679f7Sakolb if (tp == NULL || tp == T_DONTSTEAL) 2006685679f7Sakolb return (tp); 2007685679f7Sakolb return (disp_ratify(tp, kpq)); 20087c478bd9Sstevel@tonic-gate } 2009685679f7Sakolb return (retval); 20107c478bd9Sstevel@tonic-gate } 20117c478bd9Sstevel@tonic-gate 20127c478bd9Sstevel@tonic-gate 20137c478bd9Sstevel@tonic-gate /* 20147c478bd9Sstevel@tonic-gate * disp_fix_unbound_pri() 20157c478bd9Sstevel@tonic-gate * Determines the maximum priority of unbound threads on the queue. 20167c478bd9Sstevel@tonic-gate * The priority is kept for the queue, but is only increased, never 20177c478bd9Sstevel@tonic-gate * reduced unless some CPU is looking for something on that queue. 20187c478bd9Sstevel@tonic-gate * 20197c478bd9Sstevel@tonic-gate * The priority argument is the known upper limit. 20207c478bd9Sstevel@tonic-gate * 20217c478bd9Sstevel@tonic-gate * Perhaps this should be kept accurately, but that probably means 20227c478bd9Sstevel@tonic-gate * separate bitmaps for bound and unbound threads. Since only idled 20237c478bd9Sstevel@tonic-gate * CPUs will have to do this recalculation, it seems better this way. 20247c478bd9Sstevel@tonic-gate */ 20257c478bd9Sstevel@tonic-gate static void 20267c478bd9Sstevel@tonic-gate disp_fix_unbound_pri(disp_t *dp, pri_t pri) 20277c478bd9Sstevel@tonic-gate { 20287c478bd9Sstevel@tonic-gate kthread_t *tp; 20297c478bd9Sstevel@tonic-gate dispq_t *dq; 20307c478bd9Sstevel@tonic-gate ulong_t *dqactmap = dp->disp_qactmap; 20317c478bd9Sstevel@tonic-gate ulong_t mapword; 20327c478bd9Sstevel@tonic-gate int wx; 20337c478bd9Sstevel@tonic-gate 20347c478bd9Sstevel@tonic-gate ASSERT(DISP_LOCK_HELD(&dp->disp_lock)); 20357c478bd9Sstevel@tonic-gate 20367c478bd9Sstevel@tonic-gate ASSERT(pri >= 0); /* checked by caller */ 20377c478bd9Sstevel@tonic-gate 20387c478bd9Sstevel@tonic-gate /* 20397c478bd9Sstevel@tonic-gate * Start the search at the next lowest priority below the supplied 20407c478bd9Sstevel@tonic-gate * priority. This depends on the bitmap implementation. 20417c478bd9Sstevel@tonic-gate */ 20427c478bd9Sstevel@tonic-gate do { 20437c478bd9Sstevel@tonic-gate wx = pri >> BT_ULSHIFT; /* index of word in map */ 20447c478bd9Sstevel@tonic-gate 20457c478bd9Sstevel@tonic-gate /* 20467c478bd9Sstevel@tonic-gate * Form mask for all lower priorities in the word. 20477c478bd9Sstevel@tonic-gate */ 20487c478bd9Sstevel@tonic-gate mapword = dqactmap[wx] & (BT_BIW(pri) - 1); 20497c478bd9Sstevel@tonic-gate 20507c478bd9Sstevel@tonic-gate /* 20517c478bd9Sstevel@tonic-gate * Get next lower active priority. 20527c478bd9Sstevel@tonic-gate */ 20537c478bd9Sstevel@tonic-gate if (mapword != 0) { 20547c478bd9Sstevel@tonic-gate pri = (wx << BT_ULSHIFT) + highbit(mapword) - 1; 20557c478bd9Sstevel@tonic-gate } else if (wx > 0) { 20567c478bd9Sstevel@tonic-gate pri = bt_gethighbit(dqactmap, wx - 1); /* sign extend */ 20577c478bd9Sstevel@tonic-gate if (pri < 0) 20587c478bd9Sstevel@tonic-gate break; 20597c478bd9Sstevel@tonic-gate } else { 20607c478bd9Sstevel@tonic-gate pri = -1; 20617c478bd9Sstevel@tonic-gate break; 20627c478bd9Sstevel@tonic-gate } 20637c478bd9Sstevel@tonic-gate 20647c478bd9Sstevel@tonic-gate /* 20657c478bd9Sstevel@tonic-gate * Search the queue for unbound, runnable threads. 20667c478bd9Sstevel@tonic-gate */ 20677c478bd9Sstevel@tonic-gate dq = &dp->disp_q[pri]; 20687c478bd9Sstevel@tonic-gate tp = dq->dq_first; 20697c478bd9Sstevel@tonic-gate 20707c478bd9Sstevel@tonic-gate while (tp && (tp->t_bound_cpu || tp->t_weakbound_cpu)) { 20717c478bd9Sstevel@tonic-gate tp = tp->t_link; 20727c478bd9Sstevel@tonic-gate } 20737c478bd9Sstevel@tonic-gate 20747c478bd9Sstevel@tonic-gate /* 20757c478bd9Sstevel@tonic-gate * If a thread was found, set the priority and return. 20767c478bd9Sstevel@tonic-gate */ 20777c478bd9Sstevel@tonic-gate } while (tp == NULL); 20787c478bd9Sstevel@tonic-gate 20797c478bd9Sstevel@tonic-gate /* 20807c478bd9Sstevel@tonic-gate * pri holds the maximum unbound thread priority or -1. 20817c478bd9Sstevel@tonic-gate */ 20827c478bd9Sstevel@tonic-gate if (dp->disp_max_unbound_pri != pri) 20837c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = pri; 20847c478bd9Sstevel@tonic-gate } 20857c478bd9Sstevel@tonic-gate 20867c478bd9Sstevel@tonic-gate /* 20877c478bd9Sstevel@tonic-gate * disp_adjust_unbound_pri() - thread is becoming unbound, so we should 20887c478bd9Sstevel@tonic-gate * check if the CPU to which is was previously bound should have 20897c478bd9Sstevel@tonic-gate * its disp_max_unbound_pri increased. 20907c478bd9Sstevel@tonic-gate */ 20917c478bd9Sstevel@tonic-gate void 20927c478bd9Sstevel@tonic-gate disp_adjust_unbound_pri(kthread_t *tp) 20937c478bd9Sstevel@tonic-gate { 20947c478bd9Sstevel@tonic-gate disp_t *dp; 20957c478bd9Sstevel@tonic-gate pri_t tpri; 20967c478bd9Sstevel@tonic-gate 20977c478bd9Sstevel@tonic-gate ASSERT(THREAD_LOCK_HELD(tp)); 20987c478bd9Sstevel@tonic-gate 20997c478bd9Sstevel@tonic-gate /* 21007c478bd9Sstevel@tonic-gate * Don't do anything if the thread is not bound, or 21017c478bd9Sstevel@tonic-gate * currently not runnable or swapped out. 21027c478bd9Sstevel@tonic-gate */ 21037c478bd9Sstevel@tonic-gate if (tp->t_bound_cpu == NULL || 21047c478bd9Sstevel@tonic-gate tp->t_state != TS_RUN || 21057c478bd9Sstevel@tonic-gate tp->t_schedflag & TS_ON_SWAPQ) 21067c478bd9Sstevel@tonic-gate return; 21077c478bd9Sstevel@tonic-gate 21087c478bd9Sstevel@tonic-gate tpri = DISP_PRIO(tp); 21097c478bd9Sstevel@tonic-gate dp = tp->t_bound_cpu->cpu_disp; 21107c478bd9Sstevel@tonic-gate ASSERT(tpri >= 0 && tpri < dp->disp_npri); 21117c478bd9Sstevel@tonic-gate if (tpri > dp->disp_max_unbound_pri) 21127c478bd9Sstevel@tonic-gate dp->disp_max_unbound_pri = tpri; 21137c478bd9Sstevel@tonic-gate } 21147c478bd9Sstevel@tonic-gate 21157c478bd9Sstevel@tonic-gate /* 2116685679f7Sakolb * disp_getbest() 2117685679f7Sakolb * De-queue the highest priority unbound runnable thread. 2118685679f7Sakolb * Returns with the thread unlocked and onproc but at splhigh (like disp()). 2119685679f7Sakolb * Returns NULL if nothing found. 2120685679f7Sakolb * Returns T_DONTSTEAL if the thread was not stealable. 2121685679f7Sakolb * so that the caller will try again later. 21227c478bd9Sstevel@tonic-gate * 2123685679f7Sakolb * Passed a pointer to a dispatch queue not associated with this CPU, and 2124685679f7Sakolb * its type. 21257c478bd9Sstevel@tonic-gate */ 21267c478bd9Sstevel@tonic-gate static kthread_t * 21277c478bd9Sstevel@tonic-gate disp_getbest(disp_t *dp) 21287c478bd9Sstevel@tonic-gate { 21297c478bd9Sstevel@tonic-gate kthread_t *tp; 21307c478bd9Sstevel@tonic-gate dispq_t *dq; 21317c478bd9Sstevel@tonic-gate pri_t pri; 2132685679f7Sakolb cpu_t *cp, *tcp; 2133685679f7Sakolb boolean_t allbound; 21347c478bd9Sstevel@tonic-gate 21357c478bd9Sstevel@tonic-gate disp_lock_enter(&dp->disp_lock); 21367c478bd9Sstevel@tonic-gate 21377c478bd9Sstevel@tonic-gate /* 21387c478bd9Sstevel@tonic-gate * If there is nothing to run, or the CPU is in the middle of a 21397c478bd9Sstevel@tonic-gate * context switch of the only thread, return NULL. 21407c478bd9Sstevel@tonic-gate */ 2141685679f7Sakolb tcp = dp->disp_cpu; 2142685679f7Sakolb cp = CPU; 21437c478bd9Sstevel@tonic-gate pri = dp->disp_max_unbound_pri; 21447c478bd9Sstevel@tonic-gate if (pri == -1 || 2145685679f7Sakolb (tcp != NULL && (tcp->cpu_disp_flags & CPU_DISP_DONTSTEAL) && 2146685679f7Sakolb tcp->cpu_disp->disp_nrunnable == 1)) { 21477c478bd9Sstevel@tonic-gate disp_lock_exit_nopreempt(&dp->disp_lock); 21487c478bd9Sstevel@tonic-gate return (NULL); 21497c478bd9Sstevel@tonic-gate } 21507c478bd9Sstevel@tonic-gate 21517c478bd9Sstevel@tonic-gate dq = &dp->disp_q[pri]; 2152685679f7Sakolb 21537c478bd9Sstevel@tonic-gate 21547c478bd9Sstevel@tonic-gate /* 2155685679f7Sakolb * Assume that all threads are bound on this queue, and change it 2156685679f7Sakolb * later when we find out that it is not the case. 21577c478bd9Sstevel@tonic-gate */ 2158685679f7Sakolb allbound = B_TRUE; 2159685679f7Sakolb for (tp = dq->dq_first; tp != NULL; tp = tp->t_link) { 2160685679f7Sakolb hrtime_t now, nosteal, rqtime; 2161685679f7Sakolb chip_type_t chtype; 2162685679f7Sakolb chip_t *chip; 2163685679f7Sakolb 2164685679f7Sakolb /* 2165685679f7Sakolb * Skip over bound threads which could be here even 2166685679f7Sakolb * though disp_max_unbound_pri indicated this level. 2167685679f7Sakolb */ 2168685679f7Sakolb if (tp->t_bound_cpu || tp->t_weakbound_cpu) 2169685679f7Sakolb continue; 2170685679f7Sakolb 2171685679f7Sakolb /* 2172685679f7Sakolb * We've got some unbound threads on this queue, so turn 2173685679f7Sakolb * the allbound flag off now. 2174685679f7Sakolb */ 2175685679f7Sakolb allbound = B_FALSE; 2176685679f7Sakolb 2177685679f7Sakolb /* 2178685679f7Sakolb * The thread is a candidate for stealing from its run queue. We 2179685679f7Sakolb * don't want to steal threads that became runnable just a 2180685679f7Sakolb * moment ago. This improves CPU affinity for threads that get 2181685679f7Sakolb * preempted for short periods of time and go back on the run 2182685679f7Sakolb * queue. 2183685679f7Sakolb * 2184685679f7Sakolb * We want to let it stay on its run queue if it was only placed 2185685679f7Sakolb * there recently and it was running on the same CPU before that 2186685679f7Sakolb * to preserve its cache investment. For the thread to remain on 2187685679f7Sakolb * its run queue, ALL of the following conditions must be 2188685679f7Sakolb * satisfied: 2189685679f7Sakolb * 2190685679f7Sakolb * - the disp queue should not be the kernel preemption queue 2191685679f7Sakolb * - delayed idle stealing should not be disabled 2192685679f7Sakolb * - nosteal_nsec should be non-zero 2193685679f7Sakolb * - it should run with user priority 2194685679f7Sakolb * - it should be on the run queue of the CPU where it was 2195685679f7Sakolb * running before being placed on the run queue 2196685679f7Sakolb * - it should be the only thread on the run queue (to prevent 2197685679f7Sakolb * extra scheduling latency for other threads) 2198685679f7Sakolb * - it should sit on the run queue for less than per-chip 2199685679f7Sakolb * nosteal interval or global nosteal interval 2200685679f7Sakolb * - in case of CPUs with shared cache it should sit in a run 2201685679f7Sakolb * queue of a CPU from a different chip 2202685679f7Sakolb * 2203685679f7Sakolb * The checks are arranged so that the ones that are faster are 2204685679f7Sakolb * placed earlier. 2205685679f7Sakolb */ 2206685679f7Sakolb if (tcp == NULL || 2207685679f7Sakolb pri >= minclsyspri || 2208685679f7Sakolb tp->t_cpu != tcp) 2209685679f7Sakolb break; 2210685679f7Sakolb 2211685679f7Sakolb /* 2212685679f7Sakolb * Steal immediately if the chip has shared cache and we are 2213685679f7Sakolb * sharing the chip with the target thread's CPU. 2214685679f7Sakolb */ 2215685679f7Sakolb chip = tcp->cpu_chip; 2216685679f7Sakolb chtype = chip->chip_type; 2217685679f7Sakolb if ((chtype == CHIP_SMT || chtype == CHIP_CMP_SHARED_CACHE) && 2218685679f7Sakolb chip == cp->cpu_chip) 2219685679f7Sakolb break; 2220685679f7Sakolb 2221685679f7Sakolb /* 2222685679f7Sakolb * Get the value of nosteal interval either from nosteal_nsec 2223685679f7Sakolb * global variable or from a value specified by a chip 2224685679f7Sakolb */ 2225685679f7Sakolb nosteal = nosteal_nsec ? nosteal_nsec : chip->chip_nosteal; 2226685679f7Sakolb if (nosteal == 0 || nosteal == NOSTEAL_DISABLED) 2227685679f7Sakolb break; 2228685679f7Sakolb 2229685679f7Sakolb /* 2230685679f7Sakolb * Calculate time spent sitting on run queue 2231685679f7Sakolb */ 2232685679f7Sakolb now = gethrtime_unscaled(); 2233685679f7Sakolb rqtime = now - tp->t_waitrq; 2234685679f7Sakolb scalehrtime(&rqtime); 2235685679f7Sakolb 2236685679f7Sakolb /* 2237685679f7Sakolb * Steal immediately if the time spent on this run queue is more 2238685679f7Sakolb * than allowed nosteal delay. 2239685679f7Sakolb * 2240685679f7Sakolb * Negative rqtime check is needed here to avoid infinite 2241685679f7Sakolb * stealing delays caused by unlikely but not impossible 2242685679f7Sakolb * drifts between CPU times on different CPUs. 2243685679f7Sakolb */ 2244685679f7Sakolb if (rqtime > nosteal || rqtime < 0) 2245685679f7Sakolb break; 2246685679f7Sakolb 2247685679f7Sakolb DTRACE_PROBE4(nosteal, kthread_t *, tp, 2248685679f7Sakolb cpu_t *, tcp, cpu_t *, cp, hrtime_t, rqtime); 2249685679f7Sakolb scalehrtime(&now); 2250685679f7Sakolb /* 2251685679f7Sakolb * Calculate when this thread becomes stealable 2252685679f7Sakolb */ 2253685679f7Sakolb now += (nosteal - rqtime); 2254685679f7Sakolb 2255685679f7Sakolb /* 2256685679f7Sakolb * Calculate time when some thread becomes stealable 2257685679f7Sakolb */ 2258685679f7Sakolb if (now < dp->disp_steal) 2259685679f7Sakolb dp->disp_steal = now; 22607c478bd9Sstevel@tonic-gate } 22617c478bd9Sstevel@tonic-gate 22627c478bd9Sstevel@tonic-gate /* 22637c478bd9Sstevel@tonic-gate * If there were no unbound threads on this queue, find the queue 2264685679f7Sakolb * where they are and then return later. The value of 2265685679f7Sakolb * disp_max_unbound_pri is not always accurate because it isn't 2266685679f7Sakolb * reduced until another idle CPU looks for work. 22677c478bd9Sstevel@tonic-gate */ 2268685679f7Sakolb if (allbound) 22697c478bd9Sstevel@tonic-gate disp_fix_unbound_pri(dp, pri); 2270685679f7Sakolb 2271685679f7Sakolb /* 2272685679f7Sakolb * If we reached the end of the queue and found no unbound threads 2273685679f7Sakolb * then return NULL so that other CPUs will be considered. If there 2274685679f7Sakolb * are unbound threads but they cannot yet be stolen, then 2275685679f7Sakolb * return T_DONTSTEAL and try again later. 2276685679f7Sakolb */ 2277685679f7Sakolb if (tp == NULL) { 22787c478bd9Sstevel@tonic-gate disp_lock_exit_nopreempt(&dp->disp_lock); 2279685679f7Sakolb return (allbound ? NULL : T_DONTSTEAL); 22807c478bd9Sstevel@tonic-gate } 22817c478bd9Sstevel@tonic-gate 22827c478bd9Sstevel@tonic-gate /* 22837c478bd9Sstevel@tonic-gate * Found a runnable, unbound thread, so remove it from queue. 22847c478bd9Sstevel@tonic-gate * dispdeq() requires that we have the thread locked, and we do, 22857c478bd9Sstevel@tonic-gate * by virtue of holding the dispatch queue lock. dispdeq() will 22867c478bd9Sstevel@tonic-gate * put the thread in transition state, thereby dropping the dispq 22877c478bd9Sstevel@tonic-gate * lock. 22887c478bd9Sstevel@tonic-gate */ 2289685679f7Sakolb 22907c478bd9Sstevel@tonic-gate #ifdef DEBUG 22917c478bd9Sstevel@tonic-gate { 22927c478bd9Sstevel@tonic-gate int thread_was_on_queue; 22937c478bd9Sstevel@tonic-gate 22947c478bd9Sstevel@tonic-gate thread_was_on_queue = dispdeq(tp); /* drops disp_lock */ 22957c478bd9Sstevel@tonic-gate ASSERT(thread_was_on_queue); 22967c478bd9Sstevel@tonic-gate } 2297685679f7Sakolb 22987c478bd9Sstevel@tonic-gate #else /* DEBUG */ 22997c478bd9Sstevel@tonic-gate (void) dispdeq(tp); /* drops disp_lock */ 23007c478bd9Sstevel@tonic-gate #endif /* DEBUG */ 23017c478bd9Sstevel@tonic-gate 2302685679f7Sakolb /* 2303685679f7Sakolb * Reset the disp_queue steal time - we do not know what is the smallest 2304685679f7Sakolb * value across the queue is. 2305685679f7Sakolb */ 2306685679f7Sakolb dp->disp_steal = 0; 2307685679f7Sakolb 23087c478bd9Sstevel@tonic-gate tp->t_schedflag |= TS_DONT_SWAP; 23097c478bd9Sstevel@tonic-gate 23107c478bd9Sstevel@tonic-gate /* 23117c478bd9Sstevel@tonic-gate * Setup thread to run on the current CPU. 23127c478bd9Sstevel@tonic-gate */ 23137c478bd9Sstevel@tonic-gate tp->t_disp_queue = cp->cpu_disp; 23147c478bd9Sstevel@tonic-gate 23157c478bd9Sstevel@tonic-gate cp->cpu_dispthread = tp; /* protected by spl only */ 23167c478bd9Sstevel@tonic-gate cp->cpu_dispatch_pri = pri; 23177c478bd9Sstevel@tonic-gate ASSERT(pri == DISP_PRIO(tp)); 23187c478bd9Sstevel@tonic-gate 2319685679f7Sakolb DTRACE_PROBE3(steal, kthread_t *, tp, cpu_t *, tcp, cpu_t *, cp); 2320685679f7Sakolb 23217c478bd9Sstevel@tonic-gate thread_onproc(tp, cp); /* set t_state to TS_ONPROC */ 23227c478bd9Sstevel@tonic-gate 23237c478bd9Sstevel@tonic-gate /* 23247c478bd9Sstevel@tonic-gate * Return with spl high so that swtch() won't need to raise it. 23257c478bd9Sstevel@tonic-gate * The disp_lock was dropped by dispdeq(). 23267c478bd9Sstevel@tonic-gate */ 23277c478bd9Sstevel@tonic-gate 23287c478bd9Sstevel@tonic-gate return (tp); 23297c478bd9Sstevel@tonic-gate } 23307c478bd9Sstevel@tonic-gate 23317c478bd9Sstevel@tonic-gate /* 23327c478bd9Sstevel@tonic-gate * disp_bound_common() - common routine for higher level functions 23337c478bd9Sstevel@tonic-gate * that check for bound threads under certain conditions. 23347c478bd9Sstevel@tonic-gate * If 'threadlistsafe' is set then there is no need to acquire 23357c478bd9Sstevel@tonic-gate * pidlock to stop the thread list from changing (eg, if 23367c478bd9Sstevel@tonic-gate * disp_bound_* is called with cpus paused). 23377c478bd9Sstevel@tonic-gate */ 23387c478bd9Sstevel@tonic-gate static int 23397c478bd9Sstevel@tonic-gate disp_bound_common(cpu_t *cp, int threadlistsafe, int flag) 23407c478bd9Sstevel@tonic-gate { 23417c478bd9Sstevel@tonic-gate int found = 0; 23427c478bd9Sstevel@tonic-gate kthread_t *tp; 23437c478bd9Sstevel@tonic-gate 23447c478bd9Sstevel@tonic-gate ASSERT(flag); 23457c478bd9Sstevel@tonic-gate 23467c478bd9Sstevel@tonic-gate if (!threadlistsafe) 23477c478bd9Sstevel@tonic-gate mutex_enter(&pidlock); 23487c478bd9Sstevel@tonic-gate tp = curthread; /* faster than allthreads */ 23497c478bd9Sstevel@tonic-gate do { 23507c478bd9Sstevel@tonic-gate if (tp->t_state != TS_FREE) { 23517c478bd9Sstevel@tonic-gate /* 23527c478bd9Sstevel@tonic-gate * If an interrupt thread is busy, but the 23537c478bd9Sstevel@tonic-gate * caller doesn't care (i.e. BOUND_INTR is off), 23547c478bd9Sstevel@tonic-gate * then just ignore it and continue through. 23557c478bd9Sstevel@tonic-gate */ 23567c478bd9Sstevel@tonic-gate if ((tp->t_flag & T_INTR_THREAD) && 23577c478bd9Sstevel@tonic-gate !(flag & BOUND_INTR)) 23587c478bd9Sstevel@tonic-gate continue; 23597c478bd9Sstevel@tonic-gate 23607c478bd9Sstevel@tonic-gate /* 23617c478bd9Sstevel@tonic-gate * Skip the idle thread for the CPU 23627c478bd9Sstevel@tonic-gate * we're about to set offline. 23637c478bd9Sstevel@tonic-gate */ 23647c478bd9Sstevel@tonic-gate if (tp == cp->cpu_idle_thread) 23657c478bd9Sstevel@tonic-gate continue; 23667c478bd9Sstevel@tonic-gate 23677c478bd9Sstevel@tonic-gate /* 23687c478bd9Sstevel@tonic-gate * Skip the pause thread for the CPU 23697c478bd9Sstevel@tonic-gate * we're about to set offline. 23707c478bd9Sstevel@tonic-gate */ 23717c478bd9Sstevel@tonic-gate if (tp == cp->cpu_pause_thread) 23727c478bd9Sstevel@tonic-gate continue; 23737c478bd9Sstevel@tonic-gate 23747c478bd9Sstevel@tonic-gate if ((flag & BOUND_CPU) && 23757c478bd9Sstevel@tonic-gate (tp->t_bound_cpu == cp || 23767c478bd9Sstevel@tonic-gate tp->t_bind_cpu == cp->cpu_id || 23777c478bd9Sstevel@tonic-gate tp->t_weakbound_cpu == cp)) { 23787c478bd9Sstevel@tonic-gate found = 1; 23797c478bd9Sstevel@tonic-gate break; 23807c478bd9Sstevel@tonic-gate } 23817c478bd9Sstevel@tonic-gate 23827c478bd9Sstevel@tonic-gate if ((flag & BOUND_PARTITION) && 23837c478bd9Sstevel@tonic-gate (tp->t_cpupart == cp->cpu_part)) { 23847c478bd9Sstevel@tonic-gate found = 1; 23857c478bd9Sstevel@tonic-gate break; 23867c478bd9Sstevel@tonic-gate } 23877c478bd9Sstevel@tonic-gate } 23887c478bd9Sstevel@tonic-gate } while ((tp = tp->t_next) != curthread && found == 0); 23897c478bd9Sstevel@tonic-gate if (!threadlistsafe) 23907c478bd9Sstevel@tonic-gate mutex_exit(&pidlock); 23917c478bd9Sstevel@tonic-gate return (found); 23927c478bd9Sstevel@tonic-gate } 23937c478bd9Sstevel@tonic-gate 23947c478bd9Sstevel@tonic-gate /* 23957c478bd9Sstevel@tonic-gate * disp_bound_threads - return nonzero if threads are bound to the processor. 23967c478bd9Sstevel@tonic-gate * Called infrequently. Keep this simple. 23977c478bd9Sstevel@tonic-gate * Includes threads that are asleep or stopped but not onproc. 23987c478bd9Sstevel@tonic-gate */ 23997c478bd9Sstevel@tonic-gate int 24007c478bd9Sstevel@tonic-gate disp_bound_threads(cpu_t *cp, int threadlistsafe) 24017c478bd9Sstevel@tonic-gate { 24027c478bd9Sstevel@tonic-gate return (disp_bound_common(cp, threadlistsafe, BOUND_CPU)); 24037c478bd9Sstevel@tonic-gate } 24047c478bd9Sstevel@tonic-gate 24057c478bd9Sstevel@tonic-gate /* 24067c478bd9Sstevel@tonic-gate * disp_bound_anythreads - return nonzero if _any_ threads are bound 24077c478bd9Sstevel@tonic-gate * to the given processor, including interrupt threads. 24087c478bd9Sstevel@tonic-gate */ 24097c478bd9Sstevel@tonic-gate int 24107c478bd9Sstevel@tonic-gate disp_bound_anythreads(cpu_t *cp, int threadlistsafe) 24117c478bd9Sstevel@tonic-gate { 24127c478bd9Sstevel@tonic-gate return (disp_bound_common(cp, threadlistsafe, BOUND_CPU | BOUND_INTR)); 24137c478bd9Sstevel@tonic-gate } 24147c478bd9Sstevel@tonic-gate 24157c478bd9Sstevel@tonic-gate /* 24167c478bd9Sstevel@tonic-gate * disp_bound_partition - return nonzero if threads are bound to the same 24177c478bd9Sstevel@tonic-gate * partition as the processor. 24187c478bd9Sstevel@tonic-gate * Called infrequently. Keep this simple. 24197c478bd9Sstevel@tonic-gate * Includes threads that are asleep or stopped but not onproc. 24207c478bd9Sstevel@tonic-gate */ 24217c478bd9Sstevel@tonic-gate int 24227c478bd9Sstevel@tonic-gate disp_bound_partition(cpu_t *cp, int threadlistsafe) 24237c478bd9Sstevel@tonic-gate { 24247c478bd9Sstevel@tonic-gate return (disp_bound_common(cp, threadlistsafe, BOUND_PARTITION)); 24257c478bd9Sstevel@tonic-gate } 24267c478bd9Sstevel@tonic-gate 24277c478bd9Sstevel@tonic-gate /* 24287c478bd9Sstevel@tonic-gate * disp_cpu_inactive - make a CPU inactive by moving all of its unbound 24297c478bd9Sstevel@tonic-gate * threads to other CPUs. 24307c478bd9Sstevel@tonic-gate */ 24317c478bd9Sstevel@tonic-gate void 24327c478bd9Sstevel@tonic-gate disp_cpu_inactive(cpu_t *cp) 24337c478bd9Sstevel@tonic-gate { 24347c478bd9Sstevel@tonic-gate kthread_t *tp; 24357c478bd9Sstevel@tonic-gate disp_t *dp = cp->cpu_disp; 24367c478bd9Sstevel@tonic-gate dispq_t *dq; 24377c478bd9Sstevel@tonic-gate pri_t pri; 24387c478bd9Sstevel@tonic-gate int wasonq; 24397c478bd9Sstevel@tonic-gate 24407c478bd9Sstevel@tonic-gate disp_lock_enter(&dp->disp_lock); 24417c478bd9Sstevel@tonic-gate while ((pri = dp->disp_max_unbound_pri) != -1) { 24427c478bd9Sstevel@tonic-gate dq = &dp->disp_q[pri]; 24437c478bd9Sstevel@tonic-gate tp = dq->dq_first; 24447c478bd9Sstevel@tonic-gate 24457c478bd9Sstevel@tonic-gate /* 24467c478bd9Sstevel@tonic-gate * Skip over bound threads. 24477c478bd9Sstevel@tonic-gate */ 24487c478bd9Sstevel@tonic-gate while (tp != NULL && tp->t_bound_cpu != NULL) { 24497c478bd9Sstevel@tonic-gate tp = tp->t_link; 24507c478bd9Sstevel@tonic-gate } 24517c478bd9Sstevel@tonic-gate 24527c478bd9Sstevel@tonic-gate if (tp == NULL) { 24537c478bd9Sstevel@tonic-gate /* disp_max_unbound_pri must be inaccurate, so fix it */ 24547c478bd9Sstevel@tonic-gate disp_fix_unbound_pri(dp, pri); 24557c478bd9Sstevel@tonic-gate continue; 24567c478bd9Sstevel@tonic-gate } 24577c478bd9Sstevel@tonic-gate 24587c478bd9Sstevel@tonic-gate wasonq = dispdeq(tp); /* drops disp_lock */ 24597c478bd9Sstevel@tonic-gate ASSERT(wasonq); 24607c478bd9Sstevel@tonic-gate ASSERT(tp->t_weakbound_cpu == NULL); 24617c478bd9Sstevel@tonic-gate 24627c478bd9Sstevel@tonic-gate setbackdq(tp); 24637c478bd9Sstevel@tonic-gate /* 24647c478bd9Sstevel@tonic-gate * Called from cpu_offline: 24657c478bd9Sstevel@tonic-gate * 24667c478bd9Sstevel@tonic-gate * cp has already been removed from the list of active cpus 24677c478bd9Sstevel@tonic-gate * and tp->t_cpu has been changed so there is no risk of 24687c478bd9Sstevel@tonic-gate * tp ending up back on cp. 24697c478bd9Sstevel@tonic-gate * 24707c478bd9Sstevel@tonic-gate * Called from cpupart_move_cpu: 24717c478bd9Sstevel@tonic-gate * 24727c478bd9Sstevel@tonic-gate * The cpu has moved to a new cpupart. Any threads that 24737c478bd9Sstevel@tonic-gate * were on it's dispatch queues before the move remain 24747c478bd9Sstevel@tonic-gate * in the old partition and can't run in the new partition. 24757c478bd9Sstevel@tonic-gate */ 24767c478bd9Sstevel@tonic-gate ASSERT(tp->t_cpu != cp); 24777c478bd9Sstevel@tonic-gate thread_unlock(tp); 24787c478bd9Sstevel@tonic-gate 24797c478bd9Sstevel@tonic-gate disp_lock_enter(&dp->disp_lock); 24807c478bd9Sstevel@tonic-gate } 24817c478bd9Sstevel@tonic-gate disp_lock_exit(&dp->disp_lock); 24827c478bd9Sstevel@tonic-gate } 24837c478bd9Sstevel@tonic-gate 24847c478bd9Sstevel@tonic-gate /* 24857c478bd9Sstevel@tonic-gate * disp_lowpri_cpu - find CPU running the lowest priority thread. 24867c478bd9Sstevel@tonic-gate * The hint passed in is used as a starting point so we don't favor 24877c478bd9Sstevel@tonic-gate * CPU 0 or any other CPU. The caller should pass in the most recently 24887c478bd9Sstevel@tonic-gate * used CPU for the thread. 24897c478bd9Sstevel@tonic-gate * 24907c478bd9Sstevel@tonic-gate * The lgroup and priority are used to determine the best CPU to run on 24917c478bd9Sstevel@tonic-gate * in a NUMA machine. The lgroup specifies which CPUs are closest while 24927c478bd9Sstevel@tonic-gate * the thread priority will indicate whether the thread will actually run 24937c478bd9Sstevel@tonic-gate * there. To pick the best CPU, the CPUs inside and outside of the given 24947c478bd9Sstevel@tonic-gate * lgroup which are running the lowest priority threads are found. The 24957c478bd9Sstevel@tonic-gate * remote CPU is chosen only if the thread will not run locally on a CPU 24967c478bd9Sstevel@tonic-gate * within the lgroup, but will run on the remote CPU. If the thread 24977c478bd9Sstevel@tonic-gate * cannot immediately run on any CPU, the best local CPU will be chosen. 24987c478bd9Sstevel@tonic-gate * 24997c478bd9Sstevel@tonic-gate * The lpl specified also identifies the cpu partition from which 25007c478bd9Sstevel@tonic-gate * disp_lowpri_cpu should select a CPU. 25017c478bd9Sstevel@tonic-gate * 25027c478bd9Sstevel@tonic-gate * curcpu is used to indicate that disp_lowpri_cpu is being called on 25037c478bd9Sstevel@tonic-gate * behalf of the current thread. (curthread is looking for a new cpu) 25047c478bd9Sstevel@tonic-gate * In this case, cpu_dispatch_pri for this thread's cpu should be 25057c478bd9Sstevel@tonic-gate * ignored. 25067c478bd9Sstevel@tonic-gate * 25077c478bd9Sstevel@tonic-gate * If a cpu is the target of an offline request then try to avoid it. 25087c478bd9Sstevel@tonic-gate * 25097c478bd9Sstevel@tonic-gate * This function must be called at either high SPL, or with preemption 25107c478bd9Sstevel@tonic-gate * disabled, so that the "hint" CPU cannot be removed from the online 25117c478bd9Sstevel@tonic-gate * CPU list while we are traversing it. 25127c478bd9Sstevel@tonic-gate */ 25137c478bd9Sstevel@tonic-gate cpu_t * 25147c478bd9Sstevel@tonic-gate disp_lowpri_cpu(cpu_t *hint, lpl_t *lpl, pri_t tpri, cpu_t *curcpu) 25157c478bd9Sstevel@tonic-gate { 25167c478bd9Sstevel@tonic-gate cpu_t *bestcpu; 25177c478bd9Sstevel@tonic-gate cpu_t *besthomecpu; 25187c478bd9Sstevel@tonic-gate cpu_t *cp, *cpstart; 25197c478bd9Sstevel@tonic-gate 25207c478bd9Sstevel@tonic-gate pri_t bestpri; 25217c478bd9Sstevel@tonic-gate pri_t cpupri; 25227c478bd9Sstevel@tonic-gate 25237c478bd9Sstevel@tonic-gate klgrpset_t done; 25247c478bd9Sstevel@tonic-gate klgrpset_t cur_set; 25257c478bd9Sstevel@tonic-gate 25267c478bd9Sstevel@tonic-gate lpl_t *lpl_iter, *lpl_leaf; 25277c478bd9Sstevel@tonic-gate int i; 25287c478bd9Sstevel@tonic-gate 25297c478bd9Sstevel@tonic-gate /* 25307c478bd9Sstevel@tonic-gate * Scan for a CPU currently running the lowest priority thread. 25317c478bd9Sstevel@tonic-gate * Cannot get cpu_lock here because it is adaptive. 25327c478bd9Sstevel@tonic-gate * We do not require lock on CPU list. 25337c478bd9Sstevel@tonic-gate */ 25347c478bd9Sstevel@tonic-gate ASSERT(hint != NULL); 25357c478bd9Sstevel@tonic-gate ASSERT(lpl != NULL); 25367c478bd9Sstevel@tonic-gate ASSERT(lpl->lpl_ncpu > 0); 25377c478bd9Sstevel@tonic-gate 25387c478bd9Sstevel@tonic-gate /* 25397c478bd9Sstevel@tonic-gate * First examine local CPUs. Note that it's possible the hint CPU 25407c478bd9Sstevel@tonic-gate * passed in in remote to the specified home lgroup. If our priority 25417c478bd9Sstevel@tonic-gate * isn't sufficient enough such that we can run immediately at home, 25427c478bd9Sstevel@tonic-gate * then examine CPUs remote to our home lgroup. 25437c478bd9Sstevel@tonic-gate * We would like to give preference to CPUs closest to "home". 25447c478bd9Sstevel@tonic-gate * If we can't find a CPU where we'll run at a given level 25457c478bd9Sstevel@tonic-gate * of locality, we expand our search to include the next level. 25467c478bd9Sstevel@tonic-gate */ 25477c478bd9Sstevel@tonic-gate bestcpu = besthomecpu = NULL; 25487c478bd9Sstevel@tonic-gate klgrpset_clear(done); 25497c478bd9Sstevel@tonic-gate /* start with lpl we were passed */ 25507c478bd9Sstevel@tonic-gate 25517c478bd9Sstevel@tonic-gate lpl_iter = lpl; 25527c478bd9Sstevel@tonic-gate 25537c478bd9Sstevel@tonic-gate do { 25547c478bd9Sstevel@tonic-gate 25557c478bd9Sstevel@tonic-gate bestpri = SHRT_MAX; 25567c478bd9Sstevel@tonic-gate klgrpset_clear(cur_set); 25577c478bd9Sstevel@tonic-gate 25587c478bd9Sstevel@tonic-gate for (i = 0; i < lpl_iter->lpl_nrset; i++) { 25597c478bd9Sstevel@tonic-gate lpl_leaf = lpl_iter->lpl_rset[i]; 25607c478bd9Sstevel@tonic-gate if (klgrpset_ismember(done, lpl_leaf->lpl_lgrpid)) 25617c478bd9Sstevel@tonic-gate continue; 25627c478bd9Sstevel@tonic-gate 25637c478bd9Sstevel@tonic-gate klgrpset_add(cur_set, lpl_leaf->lpl_lgrpid); 25647c478bd9Sstevel@tonic-gate 25657c478bd9Sstevel@tonic-gate if (hint->cpu_lpl == lpl_leaf) 25667c478bd9Sstevel@tonic-gate cp = cpstart = hint; 25677c478bd9Sstevel@tonic-gate else 25687c478bd9Sstevel@tonic-gate cp = cpstart = lpl_leaf->lpl_cpus; 25697c478bd9Sstevel@tonic-gate 25707c478bd9Sstevel@tonic-gate do { 25717c478bd9Sstevel@tonic-gate if (cp == curcpu) 25727c478bd9Sstevel@tonic-gate cpupri = -1; 25737c478bd9Sstevel@tonic-gate else if (cp == cpu_inmotion) 25747c478bd9Sstevel@tonic-gate cpupri = SHRT_MAX; 25757c478bd9Sstevel@tonic-gate else 25767c478bd9Sstevel@tonic-gate cpupri = cp->cpu_dispatch_pri; 25777c478bd9Sstevel@tonic-gate if (cp->cpu_disp->disp_maxrunpri > cpupri) 25787c478bd9Sstevel@tonic-gate cpupri = cp->cpu_disp->disp_maxrunpri; 25797c478bd9Sstevel@tonic-gate if (cp->cpu_chosen_level > cpupri) 25807c478bd9Sstevel@tonic-gate cpupri = cp->cpu_chosen_level; 25817c478bd9Sstevel@tonic-gate if (cpupri < bestpri) { 25827c478bd9Sstevel@tonic-gate if (CPU_IDLING(cpupri)) { 25837c478bd9Sstevel@tonic-gate ASSERT((cp->cpu_flags & 25847c478bd9Sstevel@tonic-gate CPU_QUIESCED) == 0); 25857c478bd9Sstevel@tonic-gate return (cp); 25867c478bd9Sstevel@tonic-gate } 25877c478bd9Sstevel@tonic-gate bestcpu = cp; 25887c478bd9Sstevel@tonic-gate bestpri = cpupri; 25897c478bd9Sstevel@tonic-gate } 25907c478bd9Sstevel@tonic-gate } while ((cp = cp->cpu_next_lpl) != cpstart); 25917c478bd9Sstevel@tonic-gate } 25927c478bd9Sstevel@tonic-gate 25937c478bd9Sstevel@tonic-gate if (bestcpu && (tpri > bestpri)) { 25947c478bd9Sstevel@tonic-gate ASSERT((bestcpu->cpu_flags & CPU_QUIESCED) == 0); 25957c478bd9Sstevel@tonic-gate return (bestcpu); 25967c478bd9Sstevel@tonic-gate } 25977c478bd9Sstevel@tonic-gate if (besthomecpu == NULL) 25987c478bd9Sstevel@tonic-gate besthomecpu = bestcpu; 25997c478bd9Sstevel@tonic-gate /* 26007c478bd9Sstevel@tonic-gate * Add the lgrps we just considered to the "done" set 26017c478bd9Sstevel@tonic-gate */ 26027c478bd9Sstevel@tonic-gate klgrpset_or(done, cur_set); 26037c478bd9Sstevel@tonic-gate 26047c478bd9Sstevel@tonic-gate } while ((lpl_iter = lpl_iter->lpl_parent) != NULL); 26057c478bd9Sstevel@tonic-gate 26067c478bd9Sstevel@tonic-gate /* 26077c478bd9Sstevel@tonic-gate * The specified priority isn't high enough to run immediately 26087c478bd9Sstevel@tonic-gate * anywhere, so just return the best CPU from the home lgroup. 26097c478bd9Sstevel@tonic-gate */ 26107c478bd9Sstevel@tonic-gate ASSERT((besthomecpu->cpu_flags & CPU_QUIESCED) == 0); 26117c478bd9Sstevel@tonic-gate return (besthomecpu); 26127c478bd9Sstevel@tonic-gate } 26137c478bd9Sstevel@tonic-gate 26147c478bd9Sstevel@tonic-gate /* 26157c478bd9Sstevel@tonic-gate * This routine provides the generic idle cpu function for all processors. 26167c478bd9Sstevel@tonic-gate * If a processor has some specific code to execute when idle (say, to stop 26177c478bd9Sstevel@tonic-gate * the pipeline and save power) then that routine should be defined in the 26187c478bd9Sstevel@tonic-gate * processors specific code (module_xx.c) and the global variable idle_cpu 26197c478bd9Sstevel@tonic-gate * set to that function. 26207c478bd9Sstevel@tonic-gate */ 26217c478bd9Sstevel@tonic-gate static void 26227c478bd9Sstevel@tonic-gate generic_idle_cpu(void) 26237c478bd9Sstevel@tonic-gate { 26247c478bd9Sstevel@tonic-gate } 26257c478bd9Sstevel@tonic-gate 26267c478bd9Sstevel@tonic-gate /*ARGSUSED*/ 26277c478bd9Sstevel@tonic-gate static void 26287c478bd9Sstevel@tonic-gate generic_enq_thread(cpu_t *cpu, int bound) 26297c478bd9Sstevel@tonic-gate { 26307c478bd9Sstevel@tonic-gate } 26317c478bd9Sstevel@tonic-gate 26327c478bd9Sstevel@tonic-gate /* 26337c478bd9Sstevel@tonic-gate * Select a CPU for this thread to run on. Choose t->t_cpu unless: 26347c478bd9Sstevel@tonic-gate * - t->t_cpu is not in this thread's assigned lgrp 26357c478bd9Sstevel@tonic-gate * - the time since the thread last came off t->t_cpu exceeds the 26367c478bd9Sstevel@tonic-gate * rechoose time for this cpu (ignore this if t is curthread in 26377c478bd9Sstevel@tonic-gate * which case it's on CPU and t->t_disp_time is inaccurate) 26387c478bd9Sstevel@tonic-gate * - t->t_cpu is presently the target of an offline or partition move 26397c478bd9Sstevel@tonic-gate * request 26407c478bd9Sstevel@tonic-gate */ 26417c478bd9Sstevel@tonic-gate static cpu_t * 26427c478bd9Sstevel@tonic-gate cpu_choose(kthread_t *t, pri_t tpri) 26437c478bd9Sstevel@tonic-gate { 26447c478bd9Sstevel@tonic-gate ASSERT(tpri < kpqpri); 26457c478bd9Sstevel@tonic-gate 26467c478bd9Sstevel@tonic-gate if ((((lbolt - t->t_disp_time) > t->t_cpu->cpu_rechoose) && 26477c478bd9Sstevel@tonic-gate t != curthread) || t->t_cpu == cpu_inmotion) { 26487c478bd9Sstevel@tonic-gate return (disp_lowpri_cpu(t->t_cpu, t->t_lpl, tpri, NULL)); 26497c478bd9Sstevel@tonic-gate } 26507c478bd9Sstevel@tonic-gate 26517c478bd9Sstevel@tonic-gate /* 26527c478bd9Sstevel@tonic-gate * Take a trip through disp_lowpri_cpu() if the thread was 26537c478bd9Sstevel@tonic-gate * running outside it's home lgroup 26547c478bd9Sstevel@tonic-gate */ 26557c478bd9Sstevel@tonic-gate if (!klgrpset_ismember(t->t_lpl->lpl_lgrp->lgrp_set[LGRP_RSRC_CPU], 26567c478bd9Sstevel@tonic-gate t->t_cpu->cpu_lpl->lpl_lgrpid)) { 26577c478bd9Sstevel@tonic-gate return (disp_lowpri_cpu(t->t_cpu, t->t_lpl, tpri, 26587c478bd9Sstevel@tonic-gate (t == curthread) ? t->t_cpu : NULL)); 26597c478bd9Sstevel@tonic-gate } 26607c478bd9Sstevel@tonic-gate return (t->t_cpu); 26617c478bd9Sstevel@tonic-gate } 2662