kern_thread.c revision 103838
199026Sjulian/* 299026Sjulian * Copyright (C) 2001 Julian Elischer <julian@freebsd.org>. 399026Sjulian * All rights reserved. 499026Sjulian * 599026Sjulian * Redistribution and use in source and binary forms, with or without 699026Sjulian * modification, are permitted provided that the following conditions 799026Sjulian * are met: 899026Sjulian * 1. Redistributions of source code must retain the above copyright 999026Sjulian * notice(s), this list of conditions and the following disclaimer as 1099026Sjulian * the first lines of this file unmodified other than the possible 1199026Sjulian * addition of one or more copyright notices. 1299026Sjulian * 2. Redistributions in binary form must reproduce the above copyright 1399026Sjulian * notice(s), this list of conditions and the following disclaimer in the 1499026Sjulian * documentation and/or other materials provided with the distribution. 1599026Sjulian * 1699026Sjulian * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY 1799026Sjulian * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 1899026Sjulian * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 1999026Sjulian * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY 2099026Sjulian * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 2199026Sjulian * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 2299026Sjulian * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 2399026Sjulian * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2499026Sjulian * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2599026Sjulian * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH 2699026Sjulian * DAMAGE. 2799026Sjulian * 2899026Sjulian * $FreeBSD: head/sys/kern/kern_thread.c 103838 2002-09-23 06:14:30Z julian $ 2999026Sjulian */ 3099026Sjulian 3199026Sjulian#include <sys/param.h> 3299026Sjulian#include <sys/systm.h> 3399026Sjulian#include <sys/kernel.h> 3499026Sjulian#include <sys/lock.h> 3599026Sjulian#include <sys/malloc.h> 3699026Sjulian#include <sys/mutex.h> 3799026Sjulian#include <sys/proc.h> 3899026Sjulian#include <sys/sysctl.h> 3999026Sjulian#include <sys/filedesc.h> 4099026Sjulian#include <sys/tty.h> 4199026Sjulian#include <sys/signalvar.h> 4299026Sjulian#include <sys/sx.h> 4399026Sjulian#include <sys/user.h> 4499026Sjulian#include <sys/jail.h> 4599026Sjulian#include <sys/kse.h> 4699026Sjulian#include <sys/ktr.h> 47103410Smini#include <sys/ucontext.h> 4899026Sjulian 4999026Sjulian#include <vm/vm.h> 5099026Sjulian#include <vm/vm_object.h> 5199026Sjulian#include <vm/pmap.h> 5299026Sjulian#include <vm/uma.h> 5399026Sjulian#include <vm/vm_map.h> 5499026Sjulian 55100273Speter#include <machine/frame.h> 56100273Speter 5799026Sjulian/* 58103367Sjulian * KSEGRP related storage. 5999026Sjulian */ 60103367Sjulianstatic uma_zone_t ksegrp_zone; 61103367Sjulianstatic uma_zone_t kse_zone; 6299026Sjulianstatic uma_zone_t thread_zone; 6399026Sjulian 64103367Sjulian/* DEBUG ONLY */ 6599026SjulianSYSCTL_NODE(_kern, OID_AUTO, threads, CTLFLAG_RW, 0, "thread allocation"); 6699026Sjulianstatic int oiks_debug = 1; /* 0 disable, 1 printf, 2 enter debugger */ 6799026SjulianSYSCTL_INT(_kern_threads, OID_AUTO, oiks, CTLFLAG_RW, 6899026Sjulian &oiks_debug, 0, "OIKS thread debug"); 6999026Sjulian 70103838Sjulianstatic int max_threads_per_proc = 6; 71103367SjulianSYSCTL_INT(_kern_threads, OID_AUTO, max_per_proc, CTLFLAG_RW, 72103367Sjulian &max_threads_per_proc, 0, "Limit on threads per proc"); 73103367Sjulian 7499026Sjulian#define RANGEOF(type, start, end) (offsetof(type, end) - offsetof(type, start)) 7599026Sjulian 7699026Sjulianstruct threadqueue zombie_threads = TAILQ_HEAD_INITIALIZER(zombie_threads); 7799026Sjulianstruct mtx zombie_thread_lock; 7899026SjulianMTX_SYSINIT(zombie_thread_lock, &zombie_thread_lock, 7999026Sjulian "zombie_thread_lock", MTX_SPIN); 8099026Sjulian 8199026Sjulian/* 8299026Sjulian * Pepare a thread for use. 8399026Sjulian */ 8499026Sjulianstatic void 8599026Sjulianthread_ctor(void *mem, int size, void *arg) 8699026Sjulian{ 8799026Sjulian struct thread *td; 8899026Sjulian 8999026Sjulian KASSERT((size == sizeof(struct thread)), 9099552Speter ("size mismatch: %d != %d\n", size, (int)sizeof(struct thread))); 9199026Sjulian 9299026Sjulian td = (struct thread *)mem; 93103216Sjulian td->td_state = TDS_INACTIVE; 9499026Sjulian td->td_flags |= TDF_UNBOUND; 9599026Sjulian} 9699026Sjulian 9799026Sjulian/* 9899026Sjulian * Reclaim a thread after use. 9999026Sjulian */ 10099026Sjulianstatic void 10199026Sjulianthread_dtor(void *mem, int size, void *arg) 10299026Sjulian{ 10399026Sjulian struct thread *td; 10499026Sjulian 10599026Sjulian KASSERT((size == sizeof(struct thread)), 10699552Speter ("size mismatch: %d != %d\n", size, (int)sizeof(struct thread))); 10799026Sjulian 10899026Sjulian td = (struct thread *)mem; 10999026Sjulian 11099026Sjulian#ifdef INVARIANTS 11199026Sjulian /* Verify that this thread is in a safe state to free. */ 11299026Sjulian switch (td->td_state) { 113103216Sjulian case TDS_INHIBITED: 114103216Sjulian case TDS_RUNNING: 115103216Sjulian case TDS_CAN_RUN: 11699026Sjulian case TDS_RUNQ: 11799026Sjulian /* 11899026Sjulian * We must never unlink a thread that is in one of 11999026Sjulian * these states, because it is currently active. 12099026Sjulian */ 12199026Sjulian panic("bad state for thread unlinking"); 12299026Sjulian /* NOTREACHED */ 123103216Sjulian case TDS_INACTIVE: 12499026Sjulian break; 12599026Sjulian default: 12699026Sjulian panic("bad thread state"); 12799026Sjulian /* NOTREACHED */ 12899026Sjulian } 12999026Sjulian#endif 13099026Sjulian} 13199026Sjulian 13299026Sjulian/* 13399026Sjulian * Initialize type-stable parts of a thread (when newly created). 13499026Sjulian */ 13599026Sjulianstatic void 13699026Sjulianthread_init(void *mem, int size) 13799026Sjulian{ 13899026Sjulian struct thread *td; 13999026Sjulian 14099026Sjulian KASSERT((size == sizeof(struct thread)), 14199552Speter ("size mismatch: %d != %d\n", size, (int)sizeof(struct thread))); 14299026Sjulian 14399026Sjulian td = (struct thread *)mem; 144103312Sjulian mtx_lock(&Giant); 14599026Sjulian pmap_new_thread(td); 146103312Sjulian mtx_unlock(&Giant); 14799026Sjulian cpu_thread_setup(td); 14899026Sjulian} 14999026Sjulian 15099026Sjulian/* 15199026Sjulian * Tear down type-stable parts of a thread (just before being discarded). 15299026Sjulian */ 15399026Sjulianstatic void 15499026Sjulianthread_fini(void *mem, int size) 15599026Sjulian{ 15699026Sjulian struct thread *td; 15799026Sjulian 15899026Sjulian KASSERT((size == sizeof(struct thread)), 15999552Speter ("size mismatch: %d != %d\n", size, (int)sizeof(struct thread))); 16099026Sjulian 16199026Sjulian td = (struct thread *)mem; 16299026Sjulian pmap_dispose_thread(td); 16399026Sjulian} 16499026Sjulian 16599026Sjulian/* 166103410Smini * Fill a ucontext_t with a thread's context information. 167103410Smini * 168103410Smini * This is an analogue to getcontext(3). 169103410Smini */ 170103410Sminivoid 171103410Sminithread_getcontext(struct thread *td, ucontext_t *uc) 172103410Smini{ 173103410Smini 174103464Speter/* 175103464Speter * XXX this is declared in a MD include file, i386/include/ucontext.h but 176103464Speter * is used in MI code. 177103464Speter */ 178103463Speter#ifdef __i386__ 179103410Smini get_mcontext(td, &uc->uc_mcontext); 180103463Speter#endif 181103410Smini uc->uc_sigmask = td->td_proc->p_sigmask; 182103410Smini} 183103410Smini 184103410Smini/* 185103410Smini * Set a thread's context from a ucontext_t. 186103410Smini * 187103410Smini * This is an analogue to setcontext(3). 188103410Smini */ 189103410Sminiint 190103410Sminithread_setcontext(struct thread *td, ucontext_t *uc) 191103410Smini{ 192103410Smini int ret; 193103410Smini 194103464Speter/* 195103464Speter * XXX this is declared in a MD include file, i386/include/ucontext.h but 196103464Speter * is used in MI code. 197103464Speter */ 198103463Speter#ifdef __i386__ 199103410Smini ret = set_mcontext(td, &uc->uc_mcontext); 200103463Speter#else 201103463Speter ret = ENOSYS; 202103463Speter#endif 203103410Smini if (ret == 0) { 204103410Smini SIG_CANTMASK(uc->uc_sigmask); 205103410Smini PROC_LOCK(td->td_proc); 206103410Smini td->td_proc->p_sigmask = uc->uc_sigmask; 207103410Smini PROC_UNLOCK(td->td_proc); 208103410Smini } 209103410Smini return (ret); 210103410Smini} 211103410Smini 212103410Smini/* 21399026Sjulian * Initialize global thread allocation resources. 21499026Sjulian */ 21599026Sjulianvoid 21699026Sjulianthreadinit(void) 21799026Sjulian{ 21899026Sjulian 21999026Sjulian thread_zone = uma_zcreate("THREAD", sizeof (struct thread), 22099026Sjulian thread_ctor, thread_dtor, thread_init, thread_fini, 22199026Sjulian UMA_ALIGN_CACHE, 0); 222103367Sjulian ksegrp_zone = uma_zcreate("KSEGRP", sizeof (struct ksegrp), 223103367Sjulian NULL, NULL, NULL, NULL, 224103367Sjulian UMA_ALIGN_CACHE, 0); 225103367Sjulian kse_zone = uma_zcreate("KSE", sizeof (struct kse), 226103367Sjulian NULL, NULL, NULL, NULL, 227103367Sjulian UMA_ALIGN_CACHE, 0); 22899026Sjulian} 22999026Sjulian 23099026Sjulian/* 231103002Sjulian * Stash an embarasingly extra thread into the zombie thread queue. 23299026Sjulian */ 23399026Sjulianvoid 23499026Sjulianthread_stash(struct thread *td) 23599026Sjulian{ 23699026Sjulian mtx_lock_spin(&zombie_thread_lock); 23799026Sjulian TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq); 23899026Sjulian mtx_unlock_spin(&zombie_thread_lock); 23999026Sjulian} 24099026Sjulian 241103410Smini/* 242103410Smini * Reap zombie threads. 24399026Sjulian */ 24499026Sjulianvoid 24599026Sjulianthread_reap(void) 24699026Sjulian{ 24799026Sjulian struct thread *td_reaped; 24899026Sjulian 24999026Sjulian /* 25099026Sjulian * don't even bother to lock if none at this instant 25199026Sjulian * We really don't care about the next instant.. 25299026Sjulian */ 25399026Sjulian if (!TAILQ_EMPTY(&zombie_threads)) { 25499026Sjulian mtx_lock_spin(&zombie_thread_lock); 25599026Sjulian while (!TAILQ_EMPTY(&zombie_threads)) { 25699026Sjulian td_reaped = TAILQ_FIRST(&zombie_threads); 25799026Sjulian TAILQ_REMOVE(&zombie_threads, td_reaped, td_runq); 25899026Sjulian mtx_unlock_spin(&zombie_thread_lock); 25999026Sjulian thread_free(td_reaped); 26099026Sjulian mtx_lock_spin(&zombie_thread_lock); 26199026Sjulian } 26299026Sjulian mtx_unlock_spin(&zombie_thread_lock); 26399026Sjulian } 26499026Sjulian} 26599026Sjulian 26699026Sjulian/* 267103367Sjulian * Allocate a ksegrp. 268103367Sjulian */ 269103367Sjulianstruct ksegrp * 270103367Sjulianksegrp_alloc(void) 271103367Sjulian{ 272103367Sjulian return (uma_zalloc(ksegrp_zone, M_WAITOK)); 273103367Sjulian} 274103367Sjulian 275103367Sjulian/* 276103367Sjulian * Allocate a kse. 277103367Sjulian */ 278103367Sjulianstruct kse * 279103367Sjuliankse_alloc(void) 280103367Sjulian{ 281103367Sjulian return (uma_zalloc(kse_zone, M_WAITOK)); 282103367Sjulian} 283103367Sjulian 284103367Sjulian/* 28599026Sjulian * Allocate a thread. 28699026Sjulian */ 28799026Sjulianstruct thread * 28899026Sjulianthread_alloc(void) 28999026Sjulian{ 29099026Sjulian thread_reap(); /* check if any zombies to get */ 29199026Sjulian return (uma_zalloc(thread_zone, M_WAITOK)); 29299026Sjulian} 29399026Sjulian 29499026Sjulian/* 295103367Sjulian * Deallocate a ksegrp. 296103367Sjulian */ 297103367Sjulianvoid 298103367Sjulianksegrp_free(struct ksegrp *td) 299103367Sjulian{ 300103367Sjulian uma_zfree(ksegrp_zone, td); 301103367Sjulian} 302103367Sjulian 303103367Sjulian/* 304103367Sjulian * Deallocate a kse. 305103367Sjulian */ 306103367Sjulianvoid 307103367Sjuliankse_free(struct kse *td) 308103367Sjulian{ 309103367Sjulian uma_zfree(kse_zone, td); 310103367Sjulian} 311103367Sjulian 312103367Sjulian/* 31399026Sjulian * Deallocate a thread. 31499026Sjulian */ 31599026Sjulianvoid 31699026Sjulianthread_free(struct thread *td) 31799026Sjulian{ 31899026Sjulian uma_zfree(thread_zone, td); 31999026Sjulian} 32099026Sjulian 32199026Sjulian/* 32299026Sjulian * Store the thread context in the UTS's mailbox. 32399026Sjulian */ 32499026Sjulianint 32599026Sjulianthread_export_context(struct thread *td) 32699026Sjulian{ 32799026Sjulian struct kse *ke; 32899026Sjulian uintptr_t td2_mbx; 32999026Sjulian void *addr1; 33099026Sjulian void *addr2; 33199026Sjulian int error; 332103410Smini ucontext_t uc; 333103838Sjulian int unbound; 33499026Sjulian 335103838Sjulian unbound = (td->td_flags & TDF_UNBOUND); 336103838Sjulian td->td_flags &= ~TDF_UNBOUND; 337100271Speter#ifdef __ia64__ 338100271Speter td2_mbx = 0; /* pacify gcc (!) */ 339100271Speter#endif 340103410Smini /* Export the user/machine context. */ 341103410Smini error = copyin((caddr_t)td->td_mailbox + 342103410Smini offsetof(struct thread_mailbox, tm_context), 343103410Smini &uc, 344103410Smini sizeof(ucontext_t)); 345103410Smini if (error == 0) { 346103410Smini thread_getcontext(td, &uc); 347103410Smini error = copyout(&uc, (caddr_t)td->td_mailbox + 348103410Smini offsetof(struct thread_mailbox, tm_context), 349103410Smini sizeof(ucontext_t)); 350103410Smini } 35199026Sjulian 35299026Sjulian ke = td->td_kse; 35399026Sjulian addr1 = (caddr_t)ke->ke_mailbox 354103410Smini + offsetof(struct kse_mailbox, km_completed); 35599026Sjulian addr2 = (caddr_t)td->td_mailbox 356103410Smini + offsetof(struct thread_mailbox , tm_next); 35799026Sjulian /* Then link it into it's KSE's list of completed threads. */ 35899026Sjulian if (!error) { 35999026Sjulian error = td2_mbx = fuword(addr1); 36099026Sjulian if (error == -1) 36199026Sjulian error = EFAULT; 36299026Sjulian else 36399026Sjulian error = 0; 36499026Sjulian } 36599026Sjulian if (!error) 36699026Sjulian error = suword(addr2, td2_mbx); 36799026Sjulian if (!error) 36899026Sjulian error = suword(addr1, (u_long)td->td_mailbox); 36999026Sjulian if (error == -1) 37099026Sjulian error = EFAULT; 371103838Sjulian td->td_flags |= unbound; 37299026Sjulian return (error); 37399026Sjulian} 37499026Sjulian 37599026Sjulian 37699026Sjulian/* 37799026Sjulian * Discard the current thread and exit from its context. 37899026Sjulian * 37999026Sjulian * Because we can't free a thread while we're operating under its context, 38099026Sjulian * push the current thread into our KSE's ke_tdspare slot, freeing the 38199026Sjulian * thread that might be there currently. Because we know that only this 38299026Sjulian * processor will run our KSE, we needn't worry about someone else grabbing 38399026Sjulian * our context before we do a cpu_throw. 38499026Sjulian */ 38599026Sjulianvoid 38699026Sjulianthread_exit(void) 38799026Sjulian{ 38899026Sjulian struct thread *td; 38999026Sjulian struct kse *ke; 39099026Sjulian struct proc *p; 39199026Sjulian struct ksegrp *kg; 39299026Sjulian 39399026Sjulian td = curthread; 39499026Sjulian kg = td->td_ksegrp; 39599026Sjulian p = td->td_proc; 39699026Sjulian ke = td->td_kse; 39799026Sjulian 39899026Sjulian mtx_assert(&sched_lock, MA_OWNED); 399102581Sjulian KASSERT(p != NULL, ("thread exiting without a process")); 400102581Sjulian KASSERT(ke != NULL, ("thread exiting without a kse")); 401102581Sjulian KASSERT(kg != NULL, ("thread exiting without a kse group")); 40299026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 40399026Sjulian CTR1(KTR_PROC, "thread_exit: thread %p", td); 40499026Sjulian KASSERT(!mtx_owned(&Giant), ("dying thread owns giant")); 40599026Sjulian 40699026Sjulian if (ke->ke_tdspare != NULL) { 407103216Sjulian thread_stash(ke->ke_tdspare); 40899026Sjulian ke->ke_tdspare = NULL; 40999026Sjulian } 41099026Sjulian cpu_thread_exit(td); /* XXXSMP */ 41199026Sjulian 412102581Sjulian /* 413103002Sjulian * The last thread is left attached to the process 414103002Sjulian * So that the whole bundle gets recycled. Skip 415103002Sjulian * all this stuff. 416102581Sjulian */ 417103002Sjulian if (p->p_numthreads > 1) { 418103002Sjulian /* Reassign this thread's KSE. */ 419103002Sjulian ke->ke_thread = NULL; 420103002Sjulian td->td_kse = NULL; 421103002Sjulian ke->ke_state = KES_UNQUEUED; 422103002Sjulian kse_reassign(ke); 423103002Sjulian 424103002Sjulian /* Unlink this thread from its proc. and the kseg */ 425103002Sjulian TAILQ_REMOVE(&p->p_threads, td, td_plist); 426103002Sjulian p->p_numthreads--; 427103002Sjulian TAILQ_REMOVE(&kg->kg_threads, td, td_kglist); 428103002Sjulian kg->kg_numthreads--; 429103002Sjulian /* 430103002Sjulian * The test below is NOT true if we are the 431103002Sjulian * sole exiting thread. P_STOPPED_SNGL is unset 432103002Sjulian * in exit1() after it is the only survivor. 433103002Sjulian */ 434103002Sjulian if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 435103002Sjulian if (p->p_numthreads == p->p_suspcount) { 436103216Sjulian thread_unsuspend_one(p->p_singlethread); 437103002Sjulian } 43899026Sjulian } 439103002Sjulian PROC_UNLOCK(p); 440103216Sjulian td->td_state = TDS_INACTIVE; 441103002Sjulian td->td_proc = NULL; 442103002Sjulian td->td_ksegrp = NULL; 443103002Sjulian td->td_last_kse = NULL; 444103002Sjulian ke->ke_tdspare = td; 445103002Sjulian } else { 446103002Sjulian PROC_UNLOCK(p); 44799026Sjulian } 448103002Sjulian 44999026Sjulian cpu_throw(); 45099026Sjulian /* NOTREACHED */ 45199026Sjulian} 45299026Sjulian 45399026Sjulian/* 45499026Sjulian * Link a thread to a process. 455103002Sjulian * set up anything that needs to be initialized for it to 456103002Sjulian * be used by the process. 45799026Sjulian * 45899026Sjulian * Note that we do not link to the proc's ucred here. 45999026Sjulian * The thread is linked as if running but no KSE assigned. 46099026Sjulian */ 46199026Sjulianvoid 46299026Sjulianthread_link(struct thread *td, struct ksegrp *kg) 46399026Sjulian{ 46499026Sjulian struct proc *p; 46599026Sjulian 46699026Sjulian p = kg->kg_proc; 467103216Sjulian td->td_state = TDS_INACTIVE; 46899026Sjulian td->td_proc = p; 46999026Sjulian td->td_ksegrp = kg; 47099026Sjulian td->td_last_kse = NULL; 47199026Sjulian 472103002Sjulian LIST_INIT(&td->td_contested); 473103002Sjulian callout_init(&td->td_slpcallout, 1); 47499026Sjulian TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist); 47599026Sjulian TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist); 47699026Sjulian p->p_numthreads++; 47799026Sjulian kg->kg_numthreads++; 478103367Sjulian if (oiks_debug && p->p_numthreads > max_threads_per_proc) { 47999026Sjulian printf("OIKS %d\n", p->p_numthreads); 48099026Sjulian if (oiks_debug > 1) 48199026Sjulian Debugger("OIKS"); 48299026Sjulian } 48399026Sjulian td->td_kse = NULL; 48499026Sjulian} 48599026Sjulian 48699026Sjulian/* 487103410Smini * Create a thread and schedule it for upcall on the KSE given. 48899026Sjulian */ 48999026Sjulianstruct thread * 49099026Sjulianthread_schedule_upcall(struct thread *td, struct kse *ke) 49199026Sjulian{ 49299026Sjulian struct thread *td2; 49399026Sjulian 49499026Sjulian mtx_assert(&sched_lock, MA_OWNED); 49599026Sjulian if (ke->ke_tdspare != NULL) { 49699026Sjulian td2 = ke->ke_tdspare; 49799026Sjulian ke->ke_tdspare = NULL; 49899026Sjulian } else { 49999026Sjulian mtx_unlock_spin(&sched_lock); 50099026Sjulian td2 = thread_alloc(); 50199026Sjulian mtx_lock_spin(&sched_lock); 50299026Sjulian } 50399026Sjulian CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)", 50499026Sjulian td, td->td_proc->p_pid, td->td_proc->p_comm); 505103072Sjulian bzero(&td2->td_startzero, 506103002Sjulian (unsigned)RANGEOF(struct thread, td_startzero, td_endzero)); 507103002Sjulian bcopy(&td->td_startcopy, &td2->td_startcopy, 508103002Sjulian (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy)); 50999026Sjulian thread_link(td2, ke->ke_ksegrp); 510103410Smini cpu_set_upcall(td2, td->td_pcb); 511103410Smini bcopy(td->td_frame, td2->td_frame, sizeof(struct trapframe)); 512103410Smini /* 513103410Smini * The user context for this thread is selected when we choose 514103410Smini * a KSE and return to userland on it. All we need do here is 515103410Smini * note that the thread exists in order to perform an upcall. 516103410Smini * 517103410Smini * Since selecting a KSE to perform the upcall involves locking 518103410Smini * that KSE's context to our upcall, its best to wait until the 519103410Smini * last possible moment before grabbing a KSE. We do this in 520103410Smini * userret(). 521103410Smini */ 52299026Sjulian td2->td_ucred = crhold(td->td_ucred); 52399026Sjulian td2->td_flags = TDF_UNBOUND|TDF_UPCALLING; 524103216Sjulian TD_SET_CAN_RUN(td2); 52599026Sjulian setrunqueue(td2); 52699026Sjulian return (td2); 52799026Sjulian} 52899026Sjulian 52999026Sjulian/* 530103410Smini * Schedule an upcall to notify a KSE process recieved signals. 53199026Sjulian * 532103410Smini * XXX - Modifying a sigset_t like this is totally bogus. 533103410Smini */ 534103410Sministruct thread * 535103410Sminisignal_upcall(struct proc *p, int sig) 536103410Smini{ 537103410Smini struct thread *td, *td2; 538103410Smini struct kse *ke; 539103410Smini sigset_t ss; 540103410Smini int error; 541103410Smini 542103410Smini PROC_LOCK_ASSERT(p, MA_OWNED); 543103410Smini 544103410Smini td = FIRST_THREAD_IN_PROC(p); 545103410Smini ke = td->td_kse; 546103410Smini PROC_UNLOCK(p); 547103410Smini error = copyin(&ke->ke_mailbox->km_sigscaught, &ss, sizeof(sigset_t)); 548103410Smini PROC_LOCK(p); 549103410Smini if (error) 550103410Smini return (NULL); 551103410Smini SIGADDSET(ss, sig); 552103410Smini PROC_UNLOCK(p); 553103410Smini error = copyout(&ss, &ke->ke_mailbox->km_sigscaught, sizeof(sigset_t)); 554103410Smini PROC_LOCK(p); 555103410Smini if (error) 556103410Smini return (NULL); 557103410Smini mtx_lock_spin(&sched_lock); 558103410Smini td2 = thread_schedule_upcall(td, ke); 559103410Smini mtx_unlock_spin(&sched_lock); 560103410Smini return (td2); 561103410Smini} 562103410Smini 563103410Smini/* 564103410Smini * Consider whether or not an upcall should be made, and update the 565103410Smini * TDF_UPCALLING flag appropriately. 566103410Smini * 567103410Smini * This function is called when the current thread had been bound to a user 568103410Smini * thread that performed a syscall that blocked, and is now returning. 569103410Smini * Got that? syscall -> msleep -> wakeup -> syscall_return -> us. 570103410Smini * 571103410Smini * This thread will be returned to the UTS in its mailbox as a completed 572103410Smini * thread. We need to decide whether or not to perform an upcall now, 573103410Smini * or simply queue the thread for later. 574103410Smini * 575103410Smini * XXXKSE Future enhancement: We could also return back to 576103410Smini * the thread if we haven't had to do an upcall since then. 577103410Smini * If the KSE's copy is == the thread's copy, and there are 578103410Smini * no other completed threads. 579103410Smini */ 580103410Sministatic int 581103838Sjulianthread_consider_upcalling(struct thread *td) 582103410Smini{ 583103838Sjulian struct proc *p; 584103838Sjulian struct ksegrp *kg; 585103410Smini int error; 586103410Smini 587103410Smini /* 588103410Smini * Save the thread's context, and link it 589103410Smini * into the KSE's list of completed threads. 590103410Smini */ 591103410Smini error = thread_export_context(td); 592103410Smini td->td_mailbox = NULL; 593103410Smini if (error) 594103410Smini /* 595103410Smini * Failing to do the KSE operation just defaults 596103410Smini * back to synchonous operation, so just return from 597103410Smini * the syscall. 598103410Smini */ 599103410Smini return (error); 600103410Smini 601103410Smini /* 602103410Smini * Decide whether to perfom an upcall now. 603103410Smini */ 604103410Smini /* Make sure there are no other threads waiting to run. */ 605103838Sjulian p = td->td_proc; 606103838Sjulian kg = td->td_ksegrp; 607103838Sjulian PROC_LOCK(p); 608103838Sjulian mtx_lock_spin(&sched_lock); 609103838Sjulian /* bogus test, ok for testing though */ 610103838Sjulian if (TAILQ_FIRST(&kg->kg_runq) && 611103838Sjulian (TAILQ_LAST(&kg->kg_runq, threadqueue) 612103838Sjulian != kg->kg_last_assigned)) { 613103410Smini /* 614103410Smini * Another thread in this KSEG needs to run. 615103410Smini * Switch to it instead of performing an upcall, 616103410Smini * abondoning this thread. Perform the upcall 617103410Smini * later; discard this thread for now. 618103410Smini * 619103410Smini * XXXKSE - As for the other threads to run; 620103410Smini * we COULD rush through all the threads 621103410Smini * in this KSEG at this priority, or we 622103410Smini * could throw the ball back into the court 623103410Smini * and just run the highest prio kse available. 624103410Smini * What is OUR priority? The priority of the highest 625103410Smini * sycall waiting to be returned? 626103410Smini * For now, just let another KSE run (easiest). 627103410Smini * 628103410Smini * XXXKSE Future enhancement: Shove threads in this 629103410Smini * state onto a list of completed threads hanging 630103410Smini * off the KSEG. Then, collect them before performing 631103410Smini * an upcall. This way, we don't commit to an upcall 632103410Smini * on a particular KSE, but report completed threads on 633103410Smini * the next upcall to any KSE in this KSEG. 634103410Smini * 635103410Smini */ 636103410Smini thread_exit(); /* Abandon current thread. */ 637103410Smini /* NOTREACHED */ 638103410Smini } else 639103410Smini /* 640103410Smini * Perform an upcall now. 641103410Smini * 642103410Smini * XXXKSE - Assumes we are going to userland, and not 643103410Smini * nested in the kernel. 644103410Smini */ 645103410Smini td->td_flags |= TDF_UPCALLING; 646103838Sjulian mtx_unlock_spin(&sched_lock); 647103838Sjulian PROC_UNLOCK(p); 648103410Smini return (0); 649103410Smini} 650103410Smini 651103410Smini/* 652103410Smini * The extra work we go through if we are a threaded process when we 653103410Smini * return to userland. 654103410Smini * 65599026Sjulian * If we are a KSE process and returning to user mode, check for 65699026Sjulian * extra work to do before we return (e.g. for more syscalls 65799026Sjulian * to complete first). If we were in a critical section, we should 65899026Sjulian * just return to let it finish. Same if we were in the UTS (in 659103410Smini * which case the mailbox's context's busy indicator will be set). 660103410Smini * The only traps we suport will have set the mailbox. 661103410Smini * We will clear it here. 66299026Sjulian */ 66399026Sjulianint 664103838Sjulianthread_userret(struct thread *td, struct trapframe *frame) 66599026Sjulian{ 666103410Smini int error; 66799026Sjulian 668103838Sjulian#if 0 669103410Smini /* 670103410Smini * Ensure that we have a spare thread available. 671103410Smini */ 67299026Sjulian if (ke->ke_tdspare == NULL) { 673103410Smini mtx_lock(&Giant); 67499026Sjulian ke->ke_tdspare = thread_alloc(); 675103410Smini mtx_unlock(&Giant); 67699026Sjulian } 677103838Sjulian#endif 678103410Smini /* 679103410Smini * Bound threads need no additional work. 680103410Smini */ 681103410Smini if ((td->td_flags & TDF_UNBOUND) == 0) 682103410Smini return (0); 683103410Smini error = 0; 684103410Smini /* 685103410Smini * Decide whether or not we should perform an upcall now. 686103410Smini */ 687103410Smini if (((td->td_flags & TDF_UPCALLING) == 0) && td->td_mailbox) { 688103838Sjulian error = thread_consider_upcalling(td); 689103410Smini if (error != 0) 690103410Smini /* 691103410Smini * Failing to do the KSE operation just defaults 692103410Smini * back to synchonous operation, so just return from 693103410Smini * the syscall. 694103410Smini */ 695103410Smini goto cont; 696103410Smini } 697103410Smini if (td->td_flags & TDF_UPCALLING) { 69899026Sjulian /* 699103410Smini * There is no more work to do and we are going to ride 700103410Smini * this thead/KSE up to userland. 70199026Sjulian */ 702103410Smini CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)", 703103410Smini td, p->p_pid, p->p_comm); 70499026Sjulian 70599026Sjulian /* 706103410Smini * Set user context to the UTS. 70799026Sjulian */ 708103838Sjulian td->td_flags &= ~TDF_UNBOUND; 709103838Sjulian cpu_set_upcall_kse(td, td->td_kse); 710103410Smini if (error) 71199026Sjulian /* 712103410Smini * Failing to do the KSE operation just defaults 713103410Smini * back to synchonous operation, so just return from 714103410Smini * the syscall. 71599026Sjulian */ 716103410Smini goto cont; 71799026Sjulian 71899026Sjulian /* 719103410Smini * Set state and mailbox. 72099026Sjulian */ 721103410Smini td->td_flags &= ~TDF_UPCALLING; 722103410Smini error = suword((caddr_t)td->td_kse->ke_mailbox + 723103410Smini offsetof(struct kse_mailbox, km_curthread), 724103410Smini 0); 725103410Smini } 72699026Sjuliancont: 727103410Smini /* 728103410Smini * Stop any chance that we may be separated from 729103410Smini * the KSE we are currently on. This is "biting the bullet", 730103410Smini * we are committing to go to user space as as this KSE here. 731103410Smini */ 732103410Smini td->td_flags &= ~TDF_UNBOUND; /* Bind to this user thread. */ 73399026Sjulian return (error); 73499026Sjulian} 73599026Sjulian 73699026Sjulian/* 73799026Sjulian * Enforce single-threading. 73899026Sjulian * 73999026Sjulian * Returns 1 if the caller must abort (another thread is waiting to 74099026Sjulian * exit the process or similar). Process is locked! 74199026Sjulian * Returns 0 when you are successfully the only thread running. 74299026Sjulian * A process has successfully single threaded in the suspend mode when 74399026Sjulian * There are no threads in user mode. Threads in the kernel must be 74499026Sjulian * allowed to continue until they get to the user boundary. They may even 74599026Sjulian * copy out their return values and data before suspending. They may however be 74699026Sjulian * accellerated in reaching the user boundary as we will wake up 74799026Sjulian * any sleeping threads that are interruptable. (PCATCH). 74899026Sjulian */ 74999026Sjulianint 75099026Sjulianthread_single(int force_exit) 75199026Sjulian{ 75299026Sjulian struct thread *td; 75399026Sjulian struct thread *td2; 75499026Sjulian struct proc *p; 75599026Sjulian 75699026Sjulian td = curthread; 75799026Sjulian p = td->td_proc; 75899026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 75999026Sjulian KASSERT((td != NULL), ("curthread is NULL")); 76099026Sjulian 76199026Sjulian if ((p->p_flag & P_KSES) == 0) 76299026Sjulian return (0); 76399026Sjulian 764100648Sjulian /* Is someone already single threading? */ 765100648Sjulian if (p->p_singlethread) 76699026Sjulian return (1); 76799026Sjulian 768102950Sdavidxu if (force_exit == SINGLE_EXIT) 76999026Sjulian p->p_flag |= P_SINGLE_EXIT; 77099026Sjulian else 77199026Sjulian p->p_flag &= ~P_SINGLE_EXIT; 772102950Sdavidxu p->p_flag |= P_STOPPED_SINGLE; 77399026Sjulian p->p_singlethread = td; 77499026Sjulian while ((p->p_numthreads - p->p_suspcount) != 1) { 775103216Sjulian mtx_lock_spin(&sched_lock); 77699026Sjulian FOREACH_THREAD_IN_PROC(p, td2) { 77799026Sjulian if (td2 == td) 77899026Sjulian continue; 779103216Sjulian if (TD_IS_INHIBITED(td2)) { 780103216Sjulian if (TD_IS_SUSPENDED(td2)) { 781103216Sjulian if (force_exit == SINGLE_EXIT) { 782103216Sjulian thread_unsuspend_one(td2); 783103216Sjulian } 78499026Sjulian } 785103216Sjulian if ( TD_IS_SLEEPING(td2)) { 786103216Sjulian if (td2->td_flags & TDF_CVWAITQ) 787103216Sjulian cv_waitq_remove(td2); 788103216Sjulian else 789103216Sjulian unsleep(td2); 790103216Sjulian break; 791103216Sjulian } 792103216Sjulian if (TD_CAN_RUN(td2)) 793103216Sjulian setrunqueue(td2); 79499026Sjulian } 79599026Sjulian } 79699026Sjulian /* 79799026Sjulian * Wake us up when everyone else has suspended. 798100648Sjulian * In the mean time we suspend as well. 79999026Sjulian */ 800103216Sjulian thread_suspend_one(td); 80199026Sjulian mtx_unlock(&Giant); 80299026Sjulian PROC_UNLOCK(p); 80399026Sjulian mi_switch(); 80499026Sjulian mtx_unlock_spin(&sched_lock); 80599026Sjulian mtx_lock(&Giant); 80699026Sjulian PROC_LOCK(p); 80799026Sjulian } 80899026Sjulian return (0); 80999026Sjulian} 81099026Sjulian 81199026Sjulian/* 81299026Sjulian * Called in from locations that can safely check to see 81399026Sjulian * whether we have to suspend or at least throttle for a 81499026Sjulian * single-thread event (e.g. fork). 81599026Sjulian * 81699026Sjulian * Such locations include userret(). 81799026Sjulian * If the "return_instead" argument is non zero, the thread must be able to 81899026Sjulian * accept 0 (caller may continue), or 1 (caller must abort) as a result. 81999026Sjulian * 82099026Sjulian * The 'return_instead' argument tells the function if it may do a 82199026Sjulian * thread_exit() or suspend, or whether the caller must abort and back 82299026Sjulian * out instead. 82399026Sjulian * 82499026Sjulian * If the thread that set the single_threading request has set the 82599026Sjulian * P_SINGLE_EXIT bit in the process flags then this call will never return 82699026Sjulian * if 'return_instead' is false, but will exit. 82799026Sjulian * 82899026Sjulian * P_SINGLE_EXIT | return_instead == 0| return_instead != 0 82999026Sjulian *---------------+--------------------+--------------------- 83099026Sjulian * 0 | returns 0 | returns 0 or 1 83199026Sjulian * | when ST ends | immediatly 83299026Sjulian *---------------+--------------------+--------------------- 83399026Sjulian * 1 | thread exits | returns 1 83499026Sjulian * | | immediatly 83599026Sjulian * 0 = thread_exit() or suspension ok, 83699026Sjulian * other = return error instead of stopping the thread. 83799026Sjulian * 83899026Sjulian * While a full suspension is under effect, even a single threading 83999026Sjulian * thread would be suspended if it made this call (but it shouldn't). 84099026Sjulian * This call should only be made from places where 84199026Sjulian * thread_exit() would be safe as that may be the outcome unless 84299026Sjulian * return_instead is set. 84399026Sjulian */ 84499026Sjulianint 84599026Sjulianthread_suspend_check(int return_instead) 84699026Sjulian{ 84799026Sjulian struct thread *td = curthread; 84899026Sjulian struct proc *p = td->td_proc; 84999026Sjulian 85099026Sjulian td = curthread; 85199026Sjulian p = td->td_proc; 85299026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 85399026Sjulian while (P_SHOULDSTOP(p)) { 854102950Sdavidxu if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 85599026Sjulian KASSERT(p->p_singlethread != NULL, 85699026Sjulian ("singlethread not set")); 85799026Sjulian /* 858100648Sjulian * The only suspension in action is a 859100648Sjulian * single-threading. Single threader need not stop. 860100646Sjulian * XXX Should be safe to access unlocked 861100646Sjulian * as it can only be set to be true by us. 86299026Sjulian */ 863100648Sjulian if (p->p_singlethread == td) 86499026Sjulian return (0); /* Exempt from stopping. */ 86599026Sjulian } 866100648Sjulian if (return_instead) 86799026Sjulian return (1); 86899026Sjulian 86999026Sjulian /* 87099026Sjulian * If the process is waiting for us to exit, 87199026Sjulian * this thread should just suicide. 872102950Sdavidxu * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE. 87399026Sjulian */ 87499026Sjulian if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { 87599026Sjulian mtx_lock_spin(&sched_lock); 87699026Sjulian while (mtx_owned(&Giant)) 87799026Sjulian mtx_unlock(&Giant); 87899026Sjulian thread_exit(); 87999026Sjulian } 88099026Sjulian 88199026Sjulian /* 88299026Sjulian * When a thread suspends, it just 88399026Sjulian * moves to the processes's suspend queue 88499026Sjulian * and stays there. 88599026Sjulian * 88699026Sjulian * XXXKSE if TDF_BOUND is true 88799026Sjulian * it will not release it's KSE which might 88899026Sjulian * lead to deadlock if there are not enough KSEs 88999026Sjulian * to complete all waiting threads. 89099026Sjulian * Maybe be able to 'lend' it out again. 89199026Sjulian * (lent kse's can not go back to userland?) 89299026Sjulian * and can only be lent in STOPPED state. 89399026Sjulian */ 894102238Sjulian mtx_lock_spin(&sched_lock); 895102950Sdavidxu if ((p->p_flag & P_STOPPED_SIG) && 896102238Sjulian (p->p_suspcount+1 == p->p_numthreads)) { 897102238Sjulian mtx_unlock_spin(&sched_lock); 898102238Sjulian PROC_LOCK(p->p_pptr); 899102238Sjulian if ((p->p_pptr->p_procsig->ps_flag & 900102238Sjulian PS_NOCLDSTOP) == 0) { 901102238Sjulian psignal(p->p_pptr, SIGCHLD); 902102238Sjulian } 903102238Sjulian PROC_UNLOCK(p->p_pptr); 904103055Sjulian mtx_lock_spin(&sched_lock); 905102238Sjulian } 90699026Sjulian mtx_assert(&Giant, MA_NOTOWNED); 907103216Sjulian thread_suspend_one(td); 90899026Sjulian PROC_UNLOCK(p); 909102950Sdavidxu if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 910100632Sjulian if (p->p_numthreads == p->p_suspcount) { 911103216Sjulian thread_unsuspend_one(p->p_singlethread); 912100632Sjulian } 913100632Sjulian } 914100594Sjulian p->p_stats->p_ru.ru_nivcsw++; 91599026Sjulian mi_switch(); 91699026Sjulian mtx_unlock_spin(&sched_lock); 91799026Sjulian PROC_LOCK(p); 91899026Sjulian } 91999026Sjulian return (0); 92099026Sjulian} 92199026Sjulian 922102898Sdavidxuvoid 923102898Sdavidxuthread_suspend_one(struct thread *td) 924102898Sdavidxu{ 925102898Sdavidxu struct proc *p = td->td_proc; 926102898Sdavidxu 927102898Sdavidxu mtx_assert(&sched_lock, MA_OWNED); 928102898Sdavidxu p->p_suspcount++; 929103216Sjulian TD_SET_SUSPENDED(td); 930102898Sdavidxu TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq); 931103216Sjulian /* 932103216Sjulian * Hack: If we are suspending but are on the sleep queue 933103216Sjulian * then we are in msleep or the cv equivalent. We 934103216Sjulian * want to look like we have two Inhibitors. 935103216Sjulian */ 936103216Sjulian if (TD_ON_SLEEPQ(td)) 937103216Sjulian TD_SET_SLEEPING(td); 938102898Sdavidxu} 939102898Sdavidxu 940102898Sdavidxuvoid 941102898Sdavidxuthread_unsuspend_one(struct thread *td) 942102898Sdavidxu{ 943102898Sdavidxu struct proc *p = td->td_proc; 944102898Sdavidxu 945102898Sdavidxu mtx_assert(&sched_lock, MA_OWNED); 946102898Sdavidxu TAILQ_REMOVE(&p->p_suspended, td, td_runq); 947103216Sjulian TD_CLR_SUSPENDED(td); 948102898Sdavidxu p->p_suspcount--; 949103216Sjulian setrunnable(td); 950102898Sdavidxu} 951102898Sdavidxu 95299026Sjulian/* 95399026Sjulian * Allow all threads blocked by single threading to continue running. 95499026Sjulian */ 95599026Sjulianvoid 95699026Sjulianthread_unsuspend(struct proc *p) 95799026Sjulian{ 95899026Sjulian struct thread *td; 95999026Sjulian 960100646Sjulian mtx_assert(&sched_lock, MA_OWNED); 96199026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 96299026Sjulian if (!P_SHOULDSTOP(p)) { 96399026Sjulian while (( td = TAILQ_FIRST(&p->p_suspended))) { 964102898Sdavidxu thread_unsuspend_one(td); 96599026Sjulian } 966102950Sdavidxu } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) && 96799026Sjulian (p->p_numthreads == p->p_suspcount)) { 96899026Sjulian /* 96999026Sjulian * Stopping everything also did the job for the single 97099026Sjulian * threading request. Now we've downgraded to single-threaded, 97199026Sjulian * let it continue. 97299026Sjulian */ 973102898Sdavidxu thread_unsuspend_one(p->p_singlethread); 97499026Sjulian } 97599026Sjulian} 97699026Sjulian 97799026Sjulianvoid 97899026Sjulianthread_single_end(void) 97999026Sjulian{ 98099026Sjulian struct thread *td; 98199026Sjulian struct proc *p; 98299026Sjulian 98399026Sjulian td = curthread; 98499026Sjulian p = td->td_proc; 98599026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 986102950Sdavidxu p->p_flag &= ~P_STOPPED_SINGLE; 98799026Sjulian p->p_singlethread = NULL; 988102292Sjulian /* 989102292Sjulian * If there are other threads they mey now run, 990102292Sjulian * unless of course there is a blanket 'stop order' 991102292Sjulian * on the process. The single threader must be allowed 992102292Sjulian * to continue however as this is a bad place to stop. 993102292Sjulian */ 994102292Sjulian if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) { 995102292Sjulian mtx_lock_spin(&sched_lock); 996102292Sjulian while (( td = TAILQ_FIRST(&p->p_suspended))) { 997103216Sjulian thread_unsuspend_one(td); 998102292Sjulian } 999102292Sjulian mtx_unlock_spin(&sched_lock); 1000102292Sjulian } 100199026Sjulian} 100299026Sjulian 1003102292Sjulian 1004