kern_thread.c revision 103410
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 103410 2002-09-16 19:26:48Z mini $ 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 70103367Sjulianstatic int max_threads_per_proc = 4; 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 174103410Smini get_mcontext(td, &uc->uc_mcontext); 175103410Smini uc->uc_sigmask = td->td_proc->p_sigmask; 176103410Smini} 177103410Smini 178103410Smini/* 179103410Smini * Set a thread's context from a ucontext_t. 180103410Smini * 181103410Smini * This is an analogue to setcontext(3). 182103410Smini */ 183103410Sminiint 184103410Sminithread_setcontext(struct thread *td, ucontext_t *uc) 185103410Smini{ 186103410Smini int ret; 187103410Smini 188103410Smini ret = set_mcontext(td, &uc->uc_mcontext); 189103410Smini if (ret == 0) { 190103410Smini SIG_CANTMASK(uc->uc_sigmask); 191103410Smini PROC_LOCK(td->td_proc); 192103410Smini td->td_proc->p_sigmask = uc->uc_sigmask; 193103410Smini PROC_UNLOCK(td->td_proc); 194103410Smini } 195103410Smini return (ret); 196103410Smini} 197103410Smini 198103410Smini/* 19999026Sjulian * Initialize global thread allocation resources. 20099026Sjulian */ 20199026Sjulianvoid 20299026Sjulianthreadinit(void) 20399026Sjulian{ 20499026Sjulian 20599026Sjulian thread_zone = uma_zcreate("THREAD", sizeof (struct thread), 20699026Sjulian thread_ctor, thread_dtor, thread_init, thread_fini, 20799026Sjulian UMA_ALIGN_CACHE, 0); 208103367Sjulian ksegrp_zone = uma_zcreate("KSEGRP", sizeof (struct ksegrp), 209103367Sjulian NULL, NULL, NULL, NULL, 210103367Sjulian UMA_ALIGN_CACHE, 0); 211103367Sjulian kse_zone = uma_zcreate("KSE", sizeof (struct kse), 212103367Sjulian NULL, NULL, NULL, NULL, 213103367Sjulian UMA_ALIGN_CACHE, 0); 21499026Sjulian} 21599026Sjulian 21699026Sjulian/* 217103002Sjulian * Stash an embarasingly extra thread into the zombie thread queue. 21899026Sjulian */ 21999026Sjulianvoid 22099026Sjulianthread_stash(struct thread *td) 22199026Sjulian{ 22299026Sjulian mtx_lock_spin(&zombie_thread_lock); 22399026Sjulian TAILQ_INSERT_HEAD(&zombie_threads, td, td_runq); 22499026Sjulian mtx_unlock_spin(&zombie_thread_lock); 22599026Sjulian} 22699026Sjulian 227103410Smini/* 228103410Smini * Reap zombie threads. 22999026Sjulian */ 23099026Sjulianvoid 23199026Sjulianthread_reap(void) 23299026Sjulian{ 23399026Sjulian struct thread *td_reaped; 23499026Sjulian 23599026Sjulian /* 23699026Sjulian * don't even bother to lock if none at this instant 23799026Sjulian * We really don't care about the next instant.. 23899026Sjulian */ 23999026Sjulian if (!TAILQ_EMPTY(&zombie_threads)) { 24099026Sjulian mtx_lock_spin(&zombie_thread_lock); 24199026Sjulian while (!TAILQ_EMPTY(&zombie_threads)) { 24299026Sjulian td_reaped = TAILQ_FIRST(&zombie_threads); 24399026Sjulian TAILQ_REMOVE(&zombie_threads, td_reaped, td_runq); 24499026Sjulian mtx_unlock_spin(&zombie_thread_lock); 24599026Sjulian thread_free(td_reaped); 24699026Sjulian mtx_lock_spin(&zombie_thread_lock); 24799026Sjulian } 24899026Sjulian mtx_unlock_spin(&zombie_thread_lock); 24999026Sjulian } 25099026Sjulian} 25199026Sjulian 25299026Sjulian/* 253103367Sjulian * Allocate a ksegrp. 254103367Sjulian */ 255103367Sjulianstruct ksegrp * 256103367Sjulianksegrp_alloc(void) 257103367Sjulian{ 258103367Sjulian return (uma_zalloc(ksegrp_zone, M_WAITOK)); 259103367Sjulian} 260103367Sjulian 261103367Sjulian/* 262103367Sjulian * Allocate a kse. 263103367Sjulian */ 264103367Sjulianstruct kse * 265103367Sjuliankse_alloc(void) 266103367Sjulian{ 267103367Sjulian return (uma_zalloc(kse_zone, M_WAITOK)); 268103367Sjulian} 269103367Sjulian 270103367Sjulian/* 27199026Sjulian * Allocate a thread. 27299026Sjulian */ 27399026Sjulianstruct thread * 27499026Sjulianthread_alloc(void) 27599026Sjulian{ 27699026Sjulian thread_reap(); /* check if any zombies to get */ 27799026Sjulian return (uma_zalloc(thread_zone, M_WAITOK)); 27899026Sjulian} 27999026Sjulian 28099026Sjulian/* 281103367Sjulian * Deallocate a ksegrp. 282103367Sjulian */ 283103367Sjulianvoid 284103367Sjulianksegrp_free(struct ksegrp *td) 285103367Sjulian{ 286103367Sjulian uma_zfree(ksegrp_zone, td); 287103367Sjulian} 288103367Sjulian 289103367Sjulian/* 290103367Sjulian * Deallocate a kse. 291103367Sjulian */ 292103367Sjulianvoid 293103367Sjuliankse_free(struct kse *td) 294103367Sjulian{ 295103367Sjulian uma_zfree(kse_zone, td); 296103367Sjulian} 297103367Sjulian 298103367Sjulian/* 29999026Sjulian * Deallocate a thread. 30099026Sjulian */ 30199026Sjulianvoid 30299026Sjulianthread_free(struct thread *td) 30399026Sjulian{ 30499026Sjulian uma_zfree(thread_zone, td); 30599026Sjulian} 30699026Sjulian 30799026Sjulian/* 30899026Sjulian * Store the thread context in the UTS's mailbox. 30999026Sjulian */ 31099026Sjulianint 31199026Sjulianthread_export_context(struct thread *td) 31299026Sjulian{ 31399026Sjulian struct kse *ke; 31499026Sjulian uintptr_t td2_mbx; 31599026Sjulian void *addr1; 31699026Sjulian void *addr2; 31799026Sjulian int error; 318103410Smini ucontext_t uc; 31999026Sjulian 320100271Speter#ifdef __ia64__ 321100271Speter td2_mbx = 0; /* pacify gcc (!) */ 322100271Speter#endif 323103410Smini /* Export the user/machine context. */ 324103410Smini error = copyin((caddr_t)td->td_mailbox + 325103410Smini offsetof(struct thread_mailbox, tm_context), 326103410Smini &uc, 327103410Smini sizeof(ucontext_t)); 328103410Smini if (error == 0) { 329103410Smini thread_getcontext(td, &uc); 330103410Smini error = copyout(&uc, (caddr_t)td->td_mailbox + 331103410Smini offsetof(struct thread_mailbox, tm_context), 332103410Smini sizeof(ucontext_t)); 333103410Smini } 33499026Sjulian 33599026Sjulian ke = td->td_kse; 33699026Sjulian addr1 = (caddr_t)ke->ke_mailbox 337103410Smini + offsetof(struct kse_mailbox, km_completed); 33899026Sjulian addr2 = (caddr_t)td->td_mailbox 339103410Smini + offsetof(struct thread_mailbox , tm_next); 34099026Sjulian /* Then link it into it's KSE's list of completed threads. */ 34199026Sjulian if (!error) { 34299026Sjulian error = td2_mbx = fuword(addr1); 34399026Sjulian if (error == -1) 34499026Sjulian error = EFAULT; 34599026Sjulian else 34699026Sjulian error = 0; 34799026Sjulian } 34899026Sjulian if (!error) 34999026Sjulian error = suword(addr2, td2_mbx); 35099026Sjulian if (!error) 35199026Sjulian error = suword(addr1, (u_long)td->td_mailbox); 35299026Sjulian if (error == -1) 35399026Sjulian error = EFAULT; 35499026Sjulian return (error); 35599026Sjulian} 35699026Sjulian 35799026Sjulian 35899026Sjulian/* 35999026Sjulian * Discard the current thread and exit from its context. 36099026Sjulian * 36199026Sjulian * Because we can't free a thread while we're operating under its context, 36299026Sjulian * push the current thread into our KSE's ke_tdspare slot, freeing the 36399026Sjulian * thread that might be there currently. Because we know that only this 36499026Sjulian * processor will run our KSE, we needn't worry about someone else grabbing 36599026Sjulian * our context before we do a cpu_throw. 36699026Sjulian */ 36799026Sjulianvoid 36899026Sjulianthread_exit(void) 36999026Sjulian{ 37099026Sjulian struct thread *td; 37199026Sjulian struct kse *ke; 37299026Sjulian struct proc *p; 37399026Sjulian struct ksegrp *kg; 37499026Sjulian 37599026Sjulian td = curthread; 37699026Sjulian kg = td->td_ksegrp; 37799026Sjulian p = td->td_proc; 37899026Sjulian ke = td->td_kse; 37999026Sjulian 38099026Sjulian mtx_assert(&sched_lock, MA_OWNED); 381102581Sjulian KASSERT(p != NULL, ("thread exiting without a process")); 382102581Sjulian KASSERT(ke != NULL, ("thread exiting without a kse")); 383102581Sjulian KASSERT(kg != NULL, ("thread exiting without a kse group")); 38499026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 38599026Sjulian CTR1(KTR_PROC, "thread_exit: thread %p", td); 38699026Sjulian KASSERT(!mtx_owned(&Giant), ("dying thread owns giant")); 38799026Sjulian 38899026Sjulian if (ke->ke_tdspare != NULL) { 389103216Sjulian thread_stash(ke->ke_tdspare); 39099026Sjulian ke->ke_tdspare = NULL; 39199026Sjulian } 39299026Sjulian cpu_thread_exit(td); /* XXXSMP */ 39399026Sjulian 394102581Sjulian /* 395103002Sjulian * The last thread is left attached to the process 396103002Sjulian * So that the whole bundle gets recycled. Skip 397103002Sjulian * all this stuff. 398102581Sjulian */ 399103002Sjulian if (p->p_numthreads > 1) { 400103002Sjulian /* Reassign this thread's KSE. */ 401103002Sjulian ke->ke_thread = NULL; 402103002Sjulian td->td_kse = NULL; 403103002Sjulian ke->ke_state = KES_UNQUEUED; 404103002Sjulian kse_reassign(ke); 405103002Sjulian 406103002Sjulian /* Unlink this thread from its proc. and the kseg */ 407103002Sjulian TAILQ_REMOVE(&p->p_threads, td, td_plist); 408103002Sjulian p->p_numthreads--; 409103002Sjulian TAILQ_REMOVE(&kg->kg_threads, td, td_kglist); 410103002Sjulian kg->kg_numthreads--; 411103002Sjulian /* 412103002Sjulian * The test below is NOT true if we are the 413103002Sjulian * sole exiting thread. P_STOPPED_SNGL is unset 414103002Sjulian * in exit1() after it is the only survivor. 415103002Sjulian */ 416103002Sjulian if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 417103002Sjulian if (p->p_numthreads == p->p_suspcount) { 418103216Sjulian thread_unsuspend_one(p->p_singlethread); 419103002Sjulian } 42099026Sjulian } 421103002Sjulian PROC_UNLOCK(p); 422103216Sjulian td->td_state = TDS_INACTIVE; 423103002Sjulian td->td_proc = NULL; 424103002Sjulian td->td_ksegrp = NULL; 425103002Sjulian td->td_last_kse = NULL; 426103002Sjulian ke->ke_tdspare = td; 427103002Sjulian } else { 428103002Sjulian PROC_UNLOCK(p); 42999026Sjulian } 430103002Sjulian 43199026Sjulian cpu_throw(); 43299026Sjulian /* NOTREACHED */ 43399026Sjulian} 43499026Sjulian 43599026Sjulian/* 43699026Sjulian * Link a thread to a process. 437103002Sjulian * set up anything that needs to be initialized for it to 438103002Sjulian * be used by the process. 43999026Sjulian * 44099026Sjulian * Note that we do not link to the proc's ucred here. 44199026Sjulian * The thread is linked as if running but no KSE assigned. 44299026Sjulian */ 44399026Sjulianvoid 44499026Sjulianthread_link(struct thread *td, struct ksegrp *kg) 44599026Sjulian{ 44699026Sjulian struct proc *p; 44799026Sjulian 44899026Sjulian p = kg->kg_proc; 449103216Sjulian td->td_state = TDS_INACTIVE; 45099026Sjulian td->td_proc = p; 45199026Sjulian td->td_ksegrp = kg; 45299026Sjulian td->td_last_kse = NULL; 45399026Sjulian 454103002Sjulian LIST_INIT(&td->td_contested); 455103002Sjulian callout_init(&td->td_slpcallout, 1); 45699026Sjulian TAILQ_INSERT_HEAD(&p->p_threads, td, td_plist); 45799026Sjulian TAILQ_INSERT_HEAD(&kg->kg_threads, td, td_kglist); 45899026Sjulian p->p_numthreads++; 45999026Sjulian kg->kg_numthreads++; 460103367Sjulian if (oiks_debug && p->p_numthreads > max_threads_per_proc) { 46199026Sjulian printf("OIKS %d\n", p->p_numthreads); 46299026Sjulian if (oiks_debug > 1) 46399026Sjulian Debugger("OIKS"); 46499026Sjulian } 46599026Sjulian td->td_kse = NULL; 46699026Sjulian} 46799026Sjulian 46899026Sjulian/* 469103410Smini * Create a thread and schedule it for upcall on the KSE given. 47099026Sjulian */ 47199026Sjulianstruct thread * 47299026Sjulianthread_schedule_upcall(struct thread *td, struct kse *ke) 47399026Sjulian{ 47499026Sjulian struct thread *td2; 47599026Sjulian 47699026Sjulian mtx_assert(&sched_lock, MA_OWNED); 47799026Sjulian if (ke->ke_tdspare != NULL) { 47899026Sjulian td2 = ke->ke_tdspare; 47999026Sjulian ke->ke_tdspare = NULL; 48099026Sjulian } else { 48199026Sjulian mtx_unlock_spin(&sched_lock); 48299026Sjulian td2 = thread_alloc(); 48399026Sjulian mtx_lock_spin(&sched_lock); 48499026Sjulian } 48599026Sjulian CTR3(KTR_PROC, "thread_schedule_upcall: thread %p (pid %d, %s)", 48699026Sjulian td, td->td_proc->p_pid, td->td_proc->p_comm); 487103072Sjulian bzero(&td2->td_startzero, 488103002Sjulian (unsigned)RANGEOF(struct thread, td_startzero, td_endzero)); 489103002Sjulian bcopy(&td->td_startcopy, &td2->td_startcopy, 490103002Sjulian (unsigned) RANGEOF(struct thread, td_startcopy, td_endcopy)); 49199026Sjulian thread_link(td2, ke->ke_ksegrp); 492103410Smini cpu_set_upcall(td2, td->td_pcb); 493103410Smini bcopy(td->td_frame, td2->td_frame, sizeof(struct trapframe)); 494103410Smini /* 495103410Smini * The user context for this thread is selected when we choose 496103410Smini * a KSE and return to userland on it. All we need do here is 497103410Smini * note that the thread exists in order to perform an upcall. 498103410Smini * 499103410Smini * Since selecting a KSE to perform the upcall involves locking 500103410Smini * that KSE's context to our upcall, its best to wait until the 501103410Smini * last possible moment before grabbing a KSE. We do this in 502103410Smini * userret(). 503103410Smini */ 50499026Sjulian td2->td_ucred = crhold(td->td_ucred); 50599026Sjulian td2->td_flags = TDF_UNBOUND|TDF_UPCALLING; 506103216Sjulian TD_SET_CAN_RUN(td2); 50799026Sjulian setrunqueue(td2); 50899026Sjulian return (td2); 50999026Sjulian} 51099026Sjulian 51199026Sjulian/* 512103410Smini * Schedule an upcall to notify a KSE process recieved signals. 51399026Sjulian * 514103410Smini * XXX - Modifying a sigset_t like this is totally bogus. 515103410Smini */ 516103410Sministruct thread * 517103410Sminisignal_upcall(struct proc *p, int sig) 518103410Smini{ 519103410Smini struct thread *td, *td2; 520103410Smini struct kse *ke; 521103410Smini sigset_t ss; 522103410Smini int error; 523103410Smini 524103410Smini PROC_LOCK_ASSERT(p, MA_OWNED); 525103410Smini 526103410Smini td = FIRST_THREAD_IN_PROC(p); 527103410Smini ke = td->td_kse; 528103410Smini PROC_UNLOCK(p); 529103410Smini error = copyin(&ke->ke_mailbox->km_sigscaught, &ss, sizeof(sigset_t)); 530103410Smini PROC_LOCK(p); 531103410Smini if (error) 532103410Smini return (NULL); 533103410Smini SIGADDSET(ss, sig); 534103410Smini PROC_UNLOCK(p); 535103410Smini error = copyout(&ss, &ke->ke_mailbox->km_sigscaught, sizeof(sigset_t)); 536103410Smini PROC_LOCK(p); 537103410Smini if (error) 538103410Smini return (NULL); 539103410Smini mtx_lock_spin(&sched_lock); 540103410Smini td2 = thread_schedule_upcall(td, ke); 541103410Smini mtx_unlock_spin(&sched_lock); 542103410Smini return (td2); 543103410Smini} 544103410Smini 545103410Smini/* 546103410Smini * Consider whether or not an upcall should be made, and update the 547103410Smini * TDF_UPCALLING flag appropriately. 548103410Smini * 549103410Smini * This function is called when the current thread had been bound to a user 550103410Smini * thread that performed a syscall that blocked, and is now returning. 551103410Smini * Got that? syscall -> msleep -> wakeup -> syscall_return -> us. 552103410Smini * 553103410Smini * This thread will be returned to the UTS in its mailbox as a completed 554103410Smini * thread. We need to decide whether or not to perform an upcall now, 555103410Smini * or simply queue the thread for later. 556103410Smini * 557103410Smini * XXXKSE Future enhancement: We could also return back to 558103410Smini * the thread if we haven't had to do an upcall since then. 559103410Smini * If the KSE's copy is == the thread's copy, and there are 560103410Smini * no other completed threads. 561103410Smini */ 562103410Sministatic int 563103410Sminithread_consider_upcalling(struct proc *p, struct ksegrp *kg, struct kse *ke, 564103410Smini struct thread *td, struct trapframe *frame) 565103410Smini{ 566103410Smini int error; 567103410Smini 568103410Smini /* 569103410Smini * Save the thread's context, and link it 570103410Smini * into the KSE's list of completed threads. 571103410Smini */ 572103410Smini error = thread_export_context(td); 573103410Smini td->td_mailbox = NULL; 574103410Smini if (error) 575103410Smini /* 576103410Smini * Failing to do the KSE operation just defaults 577103410Smini * back to synchonous operation, so just return from 578103410Smini * the syscall. 579103410Smini */ 580103410Smini return (error); 581103410Smini 582103410Smini /* 583103410Smini * Decide whether to perfom an upcall now. 584103410Smini */ 585103410Smini /* Make sure there are no other threads waiting to run. */ 586103410Smini if (TAILQ_FIRST(&kg->kg_runq)) { 587103410Smini /* 588103410Smini * Another thread in this KSEG needs to run. 589103410Smini * Switch to it instead of performing an upcall, 590103410Smini * abondoning this thread. Perform the upcall 591103410Smini * later; discard this thread for now. 592103410Smini * 593103410Smini * XXXKSE - As for the other threads to run; 594103410Smini * we COULD rush through all the threads 595103410Smini * in this KSEG at this priority, or we 596103410Smini * could throw the ball back into the court 597103410Smini * and just run the highest prio kse available. 598103410Smini * What is OUR priority? The priority of the highest 599103410Smini * sycall waiting to be returned? 600103410Smini * For now, just let another KSE run (easiest). 601103410Smini * 602103410Smini * XXXKSE Future enhancement: Shove threads in this 603103410Smini * state onto a list of completed threads hanging 604103410Smini * off the KSEG. Then, collect them before performing 605103410Smini * an upcall. This way, we don't commit to an upcall 606103410Smini * on a particular KSE, but report completed threads on 607103410Smini * the next upcall to any KSE in this KSEG. 608103410Smini * 609103410Smini */ 610103410Smini PROC_LOCK(p); 611103410Smini mtx_lock_spin(&sched_lock); 612103410Smini thread_exit(); /* Abandon current thread. */ 613103410Smini /* NOTREACHED */ 614103410Smini } else 615103410Smini /* 616103410Smini * Perform an upcall now. 617103410Smini * 618103410Smini * XXXKSE - Assumes we are going to userland, and not 619103410Smini * nested in the kernel. 620103410Smini */ 621103410Smini td->td_flags |= TDF_UPCALLING; 622103410Smini return (0); 623103410Smini} 624103410Smini 625103410Smini/* 626103410Smini * The extra work we go through if we are a threaded process when we 627103410Smini * return to userland. 628103410Smini * 62999026Sjulian * If we are a KSE process and returning to user mode, check for 63099026Sjulian * extra work to do before we return (e.g. for more syscalls 63199026Sjulian * to complete first). If we were in a critical section, we should 63299026Sjulian * just return to let it finish. Same if we were in the UTS (in 633103410Smini * which case the mailbox's context's busy indicator will be set). 634103410Smini * The only traps we suport will have set the mailbox. 635103410Smini * We will clear it here. 63699026Sjulian */ 63799026Sjulianint 63899026Sjulianthread_userret(struct proc *p, struct ksegrp *kg, struct kse *ke, 63999026Sjulian struct thread *td, struct trapframe *frame) 64099026Sjulian{ 641103410Smini int error; 64299026Sjulian 643103410Smini /* 644103410Smini * Ensure that we have a spare thread available. 645103410Smini */ 64699026Sjulian if (ke->ke_tdspare == NULL) { 647103410Smini mtx_lock(&Giant); 64899026Sjulian ke->ke_tdspare = thread_alloc(); 649103410Smini mtx_unlock(&Giant); 65099026Sjulian } 651103410Smini 652103410Smini /* 653103410Smini * Bound threads need no additional work. 654103410Smini */ 655103410Smini if ((td->td_flags & TDF_UNBOUND) == 0) 656103410Smini return (0); 657103410Smini error = 0; 658103410Smini 659103410Smini /* 660103410Smini * Decide whether or not we should perform an upcall now. 661103410Smini */ 662103410Smini if (((td->td_flags & TDF_UPCALLING) == 0) && td->td_mailbox) { 663103410Smini error = thread_consider_upcalling(p, kg, ke, td, frame); 664103410Smini if (error != 0) 665103410Smini /* 666103410Smini * Failing to do the KSE operation just defaults 667103410Smini * back to synchonous operation, so just return from 668103410Smini * the syscall. 669103410Smini */ 670103410Smini goto cont; 671103410Smini } 672103410Smini if (td->td_flags & TDF_UPCALLING) { 67399026Sjulian /* 674103410Smini * There is no more work to do and we are going to ride 675103410Smini * this thead/KSE up to userland. 67699026Sjulian */ 677103410Smini CTR3(KTR_PROC, "userret: upcall thread %p (pid %d, %s)", 678103410Smini td, p->p_pid, p->p_comm); 67999026Sjulian 68099026Sjulian /* 681103410Smini * Set user context to the UTS. 68299026Sjulian */ 683103410Smini cpu_set_upcall_kse(td, ke); 684103410Smini if (error) 68599026Sjulian /* 686103410Smini * Failing to do the KSE operation just defaults 687103410Smini * back to synchonous operation, so just return from 688103410Smini * the syscall. 68999026Sjulian */ 690103410Smini goto cont; 69199026Sjulian 69299026Sjulian /* 693103410Smini * Set state and mailbox. 69499026Sjulian */ 695103410Smini td->td_flags &= ~TDF_UPCALLING; 696103410Smini error = suword((caddr_t)td->td_kse->ke_mailbox + 697103410Smini offsetof(struct kse_mailbox, km_curthread), 698103410Smini 0); 699103410Smini } 70099026Sjuliancont: 701103410Smini /* 702103410Smini * Stop any chance that we may be separated from 703103410Smini * the KSE we are currently on. This is "biting the bullet", 704103410Smini * we are committing to go to user space as as this KSE here. 705103410Smini */ 706103410Smini td->td_flags &= ~TDF_UNBOUND; /* Bind to this user thread. */ 70799026Sjulian return (error); 70899026Sjulian} 70999026Sjulian 71099026Sjulian/* 71199026Sjulian * Enforce single-threading. 71299026Sjulian * 71399026Sjulian * Returns 1 if the caller must abort (another thread is waiting to 71499026Sjulian * exit the process or similar). Process is locked! 71599026Sjulian * Returns 0 when you are successfully the only thread running. 71699026Sjulian * A process has successfully single threaded in the suspend mode when 71799026Sjulian * There are no threads in user mode. Threads in the kernel must be 71899026Sjulian * allowed to continue until they get to the user boundary. They may even 71999026Sjulian * copy out their return values and data before suspending. They may however be 72099026Sjulian * accellerated in reaching the user boundary as we will wake up 72199026Sjulian * any sleeping threads that are interruptable. (PCATCH). 72299026Sjulian */ 72399026Sjulianint 72499026Sjulianthread_single(int force_exit) 72599026Sjulian{ 72699026Sjulian struct thread *td; 72799026Sjulian struct thread *td2; 72899026Sjulian struct proc *p; 72999026Sjulian 73099026Sjulian td = curthread; 73199026Sjulian p = td->td_proc; 73299026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 73399026Sjulian KASSERT((td != NULL), ("curthread is NULL")); 73499026Sjulian 73599026Sjulian if ((p->p_flag & P_KSES) == 0) 73699026Sjulian return (0); 73799026Sjulian 738100648Sjulian /* Is someone already single threading? */ 739100648Sjulian if (p->p_singlethread) 74099026Sjulian return (1); 74199026Sjulian 742102950Sdavidxu if (force_exit == SINGLE_EXIT) 74399026Sjulian p->p_flag |= P_SINGLE_EXIT; 74499026Sjulian else 74599026Sjulian p->p_flag &= ~P_SINGLE_EXIT; 746102950Sdavidxu p->p_flag |= P_STOPPED_SINGLE; 74799026Sjulian p->p_singlethread = td; 74899026Sjulian while ((p->p_numthreads - p->p_suspcount) != 1) { 749103216Sjulian mtx_lock_spin(&sched_lock); 75099026Sjulian FOREACH_THREAD_IN_PROC(p, td2) { 75199026Sjulian if (td2 == td) 75299026Sjulian continue; 753103216Sjulian if (TD_IS_INHIBITED(td2)) { 754103216Sjulian if (TD_IS_SUSPENDED(td2)) { 755103216Sjulian if (force_exit == SINGLE_EXIT) { 756103216Sjulian thread_unsuspend_one(td2); 757103216Sjulian } 75899026Sjulian } 759103216Sjulian if ( TD_IS_SLEEPING(td2)) { 760103216Sjulian if (td2->td_flags & TDF_CVWAITQ) 761103216Sjulian cv_waitq_remove(td2); 762103216Sjulian else 763103216Sjulian unsleep(td2); 764103216Sjulian break; 765103216Sjulian } 766103216Sjulian if (TD_CAN_RUN(td2)) 767103216Sjulian setrunqueue(td2); 76899026Sjulian } 76999026Sjulian } 77099026Sjulian /* 77199026Sjulian * Wake us up when everyone else has suspended. 772100648Sjulian * In the mean time we suspend as well. 77399026Sjulian */ 774103216Sjulian thread_suspend_one(td); 77599026Sjulian mtx_unlock(&Giant); 77699026Sjulian PROC_UNLOCK(p); 77799026Sjulian mi_switch(); 77899026Sjulian mtx_unlock_spin(&sched_lock); 77999026Sjulian mtx_lock(&Giant); 78099026Sjulian PROC_LOCK(p); 78199026Sjulian } 78299026Sjulian return (0); 78399026Sjulian} 78499026Sjulian 78599026Sjulian/* 78699026Sjulian * Called in from locations that can safely check to see 78799026Sjulian * whether we have to suspend or at least throttle for a 78899026Sjulian * single-thread event (e.g. fork). 78999026Sjulian * 79099026Sjulian * Such locations include userret(). 79199026Sjulian * If the "return_instead" argument is non zero, the thread must be able to 79299026Sjulian * accept 0 (caller may continue), or 1 (caller must abort) as a result. 79399026Sjulian * 79499026Sjulian * The 'return_instead' argument tells the function if it may do a 79599026Sjulian * thread_exit() or suspend, or whether the caller must abort and back 79699026Sjulian * out instead. 79799026Sjulian * 79899026Sjulian * If the thread that set the single_threading request has set the 79999026Sjulian * P_SINGLE_EXIT bit in the process flags then this call will never return 80099026Sjulian * if 'return_instead' is false, but will exit. 80199026Sjulian * 80299026Sjulian * P_SINGLE_EXIT | return_instead == 0| return_instead != 0 80399026Sjulian *---------------+--------------------+--------------------- 80499026Sjulian * 0 | returns 0 | returns 0 or 1 80599026Sjulian * | when ST ends | immediatly 80699026Sjulian *---------------+--------------------+--------------------- 80799026Sjulian * 1 | thread exits | returns 1 80899026Sjulian * | | immediatly 80999026Sjulian * 0 = thread_exit() or suspension ok, 81099026Sjulian * other = return error instead of stopping the thread. 81199026Sjulian * 81299026Sjulian * While a full suspension is under effect, even a single threading 81399026Sjulian * thread would be suspended if it made this call (but it shouldn't). 81499026Sjulian * This call should only be made from places where 81599026Sjulian * thread_exit() would be safe as that may be the outcome unless 81699026Sjulian * return_instead is set. 81799026Sjulian */ 81899026Sjulianint 81999026Sjulianthread_suspend_check(int return_instead) 82099026Sjulian{ 82199026Sjulian struct thread *td = curthread; 82299026Sjulian struct proc *p = td->td_proc; 82399026Sjulian 82499026Sjulian td = curthread; 82599026Sjulian p = td->td_proc; 82699026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 82799026Sjulian while (P_SHOULDSTOP(p)) { 828102950Sdavidxu if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 82999026Sjulian KASSERT(p->p_singlethread != NULL, 83099026Sjulian ("singlethread not set")); 83199026Sjulian /* 832100648Sjulian * The only suspension in action is a 833100648Sjulian * single-threading. Single threader need not stop. 834100646Sjulian * XXX Should be safe to access unlocked 835100646Sjulian * as it can only be set to be true by us. 83699026Sjulian */ 837100648Sjulian if (p->p_singlethread == td) 83899026Sjulian return (0); /* Exempt from stopping. */ 83999026Sjulian } 840100648Sjulian if (return_instead) 84199026Sjulian return (1); 84299026Sjulian 84399026Sjulian /* 84499026Sjulian * If the process is waiting for us to exit, 84599026Sjulian * this thread should just suicide. 846102950Sdavidxu * Assumes that P_SINGLE_EXIT implies P_STOPPED_SINGLE. 84799026Sjulian */ 84899026Sjulian if ((p->p_flag & P_SINGLE_EXIT) && (p->p_singlethread != td)) { 84999026Sjulian mtx_lock_spin(&sched_lock); 85099026Sjulian while (mtx_owned(&Giant)) 85199026Sjulian mtx_unlock(&Giant); 85299026Sjulian thread_exit(); 85399026Sjulian } 85499026Sjulian 85599026Sjulian /* 85699026Sjulian * When a thread suspends, it just 85799026Sjulian * moves to the processes's suspend queue 85899026Sjulian * and stays there. 85999026Sjulian * 86099026Sjulian * XXXKSE if TDF_BOUND is true 86199026Sjulian * it will not release it's KSE which might 86299026Sjulian * lead to deadlock if there are not enough KSEs 86399026Sjulian * to complete all waiting threads. 86499026Sjulian * Maybe be able to 'lend' it out again. 86599026Sjulian * (lent kse's can not go back to userland?) 86699026Sjulian * and can only be lent in STOPPED state. 86799026Sjulian */ 868102238Sjulian mtx_lock_spin(&sched_lock); 869102950Sdavidxu if ((p->p_flag & P_STOPPED_SIG) && 870102238Sjulian (p->p_suspcount+1 == p->p_numthreads)) { 871102238Sjulian mtx_unlock_spin(&sched_lock); 872102238Sjulian PROC_LOCK(p->p_pptr); 873102238Sjulian if ((p->p_pptr->p_procsig->ps_flag & 874102238Sjulian PS_NOCLDSTOP) == 0) { 875102238Sjulian psignal(p->p_pptr, SIGCHLD); 876102238Sjulian } 877102238Sjulian PROC_UNLOCK(p->p_pptr); 878103055Sjulian mtx_lock_spin(&sched_lock); 879102238Sjulian } 88099026Sjulian mtx_assert(&Giant, MA_NOTOWNED); 881103216Sjulian thread_suspend_one(td); 88299026Sjulian PROC_UNLOCK(p); 883102950Sdavidxu if (P_SHOULDSTOP(p) == P_STOPPED_SINGLE) { 884100632Sjulian if (p->p_numthreads == p->p_suspcount) { 885103216Sjulian thread_unsuspend_one(p->p_singlethread); 886100632Sjulian } 887100632Sjulian } 888100594Sjulian p->p_stats->p_ru.ru_nivcsw++; 88999026Sjulian mi_switch(); 89099026Sjulian mtx_unlock_spin(&sched_lock); 89199026Sjulian PROC_LOCK(p); 89299026Sjulian } 89399026Sjulian return (0); 89499026Sjulian} 89599026Sjulian 896102898Sdavidxuvoid 897102898Sdavidxuthread_suspend_one(struct thread *td) 898102898Sdavidxu{ 899102898Sdavidxu struct proc *p = td->td_proc; 900102898Sdavidxu 901102898Sdavidxu mtx_assert(&sched_lock, MA_OWNED); 902102898Sdavidxu p->p_suspcount++; 903103216Sjulian TD_SET_SUSPENDED(td); 904102898Sdavidxu TAILQ_INSERT_TAIL(&p->p_suspended, td, td_runq); 905103216Sjulian /* 906103216Sjulian * Hack: If we are suspending but are on the sleep queue 907103216Sjulian * then we are in msleep or the cv equivalent. We 908103216Sjulian * want to look like we have two Inhibitors. 909103216Sjulian */ 910103216Sjulian if (TD_ON_SLEEPQ(td)) 911103216Sjulian TD_SET_SLEEPING(td); 912102898Sdavidxu} 913102898Sdavidxu 914102898Sdavidxuvoid 915102898Sdavidxuthread_unsuspend_one(struct thread *td) 916102898Sdavidxu{ 917102898Sdavidxu struct proc *p = td->td_proc; 918102898Sdavidxu 919102898Sdavidxu mtx_assert(&sched_lock, MA_OWNED); 920102898Sdavidxu TAILQ_REMOVE(&p->p_suspended, td, td_runq); 921103216Sjulian TD_CLR_SUSPENDED(td); 922102898Sdavidxu p->p_suspcount--; 923103216Sjulian setrunnable(td); 924102898Sdavidxu} 925102898Sdavidxu 92699026Sjulian/* 92799026Sjulian * Allow all threads blocked by single threading to continue running. 92899026Sjulian */ 92999026Sjulianvoid 93099026Sjulianthread_unsuspend(struct proc *p) 93199026Sjulian{ 93299026Sjulian struct thread *td; 93399026Sjulian 934100646Sjulian mtx_assert(&sched_lock, MA_OWNED); 93599026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 93699026Sjulian if (!P_SHOULDSTOP(p)) { 93799026Sjulian while (( td = TAILQ_FIRST(&p->p_suspended))) { 938102898Sdavidxu thread_unsuspend_one(td); 93999026Sjulian } 940102950Sdavidxu } else if ((P_SHOULDSTOP(p) == P_STOPPED_SINGLE) && 94199026Sjulian (p->p_numthreads == p->p_suspcount)) { 94299026Sjulian /* 94399026Sjulian * Stopping everything also did the job for the single 94499026Sjulian * threading request. Now we've downgraded to single-threaded, 94599026Sjulian * let it continue. 94699026Sjulian */ 947102898Sdavidxu thread_unsuspend_one(p->p_singlethread); 94899026Sjulian } 94999026Sjulian} 95099026Sjulian 95199026Sjulianvoid 95299026Sjulianthread_single_end(void) 95399026Sjulian{ 95499026Sjulian struct thread *td; 95599026Sjulian struct proc *p; 95699026Sjulian 95799026Sjulian td = curthread; 95899026Sjulian p = td->td_proc; 95999026Sjulian PROC_LOCK_ASSERT(p, MA_OWNED); 960102950Sdavidxu p->p_flag &= ~P_STOPPED_SINGLE; 96199026Sjulian p->p_singlethread = NULL; 962102292Sjulian /* 963102292Sjulian * If there are other threads they mey now run, 964102292Sjulian * unless of course there is a blanket 'stop order' 965102292Sjulian * on the process. The single threader must be allowed 966102292Sjulian * to continue however as this is a bad place to stop. 967102292Sjulian */ 968102292Sjulian if ((p->p_numthreads != 1) && (!P_SHOULDSTOP(p))) { 969102292Sjulian mtx_lock_spin(&sched_lock); 970102292Sjulian while (( td = TAILQ_FIRST(&p->p_suspended))) { 971103216Sjulian thread_unsuspend_one(td); 972102292Sjulian } 973102292Sjulian mtx_unlock_spin(&sched_lock); 974102292Sjulian } 97599026Sjulian} 97699026Sjulian 977102292Sjulian 978