subr_turnstile.c revision 176078
165557Sjasone/*- 265557Sjasone * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 365557Sjasone * 465557Sjasone * Redistribution and use in source and binary forms, with or without 565557Sjasone * modification, are permitted provided that the following conditions 665557Sjasone * are met: 765557Sjasone * 1. Redistributions of source code must retain the above copyright 865557Sjasone * notice, this list of conditions and the following disclaimer. 965557Sjasone * 2. Redistributions in binary form must reproduce the above copyright 1065557Sjasone * notice, this list of conditions and the following disclaimer in the 1165557Sjasone * documentation and/or other materials provided with the distribution. 1265557Sjasone * 3. Berkeley Software Design Inc's name may not be used to endorse or 1365557Sjasone * promote products derived from this software without specific prior 1465557Sjasone * written permission. 1565557Sjasone * 1665557Sjasone * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 1765557Sjasone * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 1865557Sjasone * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 1965557Sjasone * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 2065557Sjasone * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2165557Sjasone * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2265557Sjasone * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2365557Sjasone * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 2465557Sjasone * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 2565557Sjasone * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 2665557Sjasone * SUCH DAMAGE. 2765557Sjasone * 2865557Sjasone * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 2967352Sjhb * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 3065557Sjasone */ 3165557Sjasone 3265557Sjasone/* 33122514Sjhb * Implementation of turnstiles used to hold queue of threads blocked on 34122514Sjhb * non-sleepable locks. Sleepable locks use condition variables to 35122514Sjhb * implement their queues. Turnstiles differ from a sleep queue in that 36122514Sjhb * turnstile queue's are assigned to a lock held by an owning thread. Thus, 37122514Sjhb * when one thread is enqueued onto a turnstile, it can lend its priority 38122514Sjhb * to the owning thread. 39122514Sjhb * 40122514Sjhb * We wish to avoid bloating locks with an embedded turnstile and we do not 41122514Sjhb * want to use back-pointers in the locks for the same reason. Thus, we 42122514Sjhb * use a similar approach to that of Solaris 7 as described in Solaris 43122514Sjhb * Internals by Jim Mauro and Richard McDougall. Turnstiles are looked up 44122514Sjhb * in a hash table based on the address of the lock. Each entry in the 45122514Sjhb * hash table is a linked-lists of turnstiles and is called a turnstile 46122514Sjhb * chain. Each chain contains a spin mutex that protects all of the 47122514Sjhb * turnstiles in the chain. 48122514Sjhb * 49169666Sjeff * Each time a thread is created, a turnstile is allocated from a UMA zone 50169666Sjeff * and attached to that thread. When a thread blocks on a lock, if it is the 51169666Sjeff * first thread to block, it lends its turnstile to the lock. If the lock 52169666Sjeff * already has a turnstile, then it gives its turnstile to the lock's 53169666Sjeff * turnstile's free list. When a thread is woken up, it takes a turnstile from 54169666Sjeff * the free list if there are any other waiters. If it is the only thread 55169666Sjeff * blocked on the lock, then it reclaims the turnstile associated with the lock 56169666Sjeff * and removes it from the hash table. 5772200Sbmilekic */ 5872200Sbmilekic 59116182Sobrien#include <sys/cdefs.h> 60116182Sobrien__FBSDID("$FreeBSD: head/sys/kern/subr_turnstile.c 176078 2008-02-07 06:55:38Z jeff $"); 61116182Sobrien 62154937Sjhb#include "opt_ddb.h" 63154937Sjhb#include "opt_turnstile_profiling.h" 64170640Sjeff#include "opt_sched.h" 65154937Sjhb 6665557Sjasone#include <sys/param.h> 6793609Sdes#include <sys/systm.h> 6867352Sjhb#include <sys/kernel.h> 6993609Sdes#include <sys/ktr.h> 7076166Smarkm#include <sys/lock.h> 7174912Sjhb#include <sys/mutex.h> 7265557Sjasone#include <sys/proc.h> 73122514Sjhb#include <sys/queue.h> 74131259Sjhb#include <sys/sched.h> 75131259Sjhb#include <sys/sysctl.h> 76122514Sjhb#include <sys/turnstile.h> 7765557Sjasone 78169666Sjeff#include <vm/uma.h> 79169666Sjeff 80154937Sjhb#ifdef DDB 81158031Sjhb#include <sys/kdb.h> 82154937Sjhb#include <ddb/ddb.h> 83161337Sjhb#include <sys/lockmgr.h> 84161337Sjhb#include <sys/sx.h> 85154937Sjhb#endif 86154937Sjhb 8765557Sjasone/* 88122514Sjhb * Constants for the hash table of turnstile chains. TC_SHIFT is a magic 89122514Sjhb * number chosen because the sleep queue's use the same value for the 90122514Sjhb * shift. Basically, we ignore the lower 8 bits of the address. 91122514Sjhb * TC_TABLESIZE must be a power of two for TC_MASK to work properly. 9271352Sjasone */ 93122514Sjhb#define TC_TABLESIZE 128 /* Must be power of 2. */ 94122514Sjhb#define TC_MASK (TC_TABLESIZE - 1) 95122514Sjhb#define TC_SHIFT 8 96122514Sjhb#define TC_HASH(lock) (((uintptr_t)(lock) >> TC_SHIFT) & TC_MASK) 97122514Sjhb#define TC_LOOKUP(lock) &turnstile_chains[TC_HASH(lock)] 9871352Sjasone 9971352Sjasone/* 100122514Sjhb * There are three different lists of turnstiles as follows. The list 101122514Sjhb * connected by ts_link entries is a per-thread list of all the turnstiles 102122514Sjhb * attached to locks that we own. This is used to fixup our priority when 103122514Sjhb * a lock is released. The other two lists use the ts_hash entries. The 104126317Sjhb * first of these two is the turnstile chain list that a turnstile is on 105126317Sjhb * when it is attached to a lock. The second list to use ts_hash is the 106126317Sjhb * free list hung off of a turnstile that is attached to a lock. 107122514Sjhb * 108154937Sjhb * Each turnstile contains three lists of threads. The two ts_blocked lists 109154937Sjhb * are linked list of threads blocked on the turnstile's lock. One list is 110154937Sjhb * for exclusive waiters, and the other is for shared waiters. The 111126884Sjhb * ts_pending list is a linked list of threads previously awakened by 112122514Sjhb * turnstile_signal() or turnstile_wait() that are waiting to be put on 113122514Sjhb * the run queue. 114122514Sjhb * 115122514Sjhb * Locking key: 116122514Sjhb * c - turnstile chain lock 117122514Sjhb * q - td_contested lock 11871352Sjasone */ 119122514Sjhbstruct turnstile { 120170295Sjeff struct mtx ts_lock; /* Spin lock for self. */ 121154937Sjhb struct threadqueue ts_blocked[2]; /* (c + q) Blocked threads. */ 122154937Sjhb struct threadqueue ts_pending; /* (c) Pending threads. */ 123122514Sjhb LIST_ENTRY(turnstile) ts_hash; /* (c) Chain and free list. */ 124122514Sjhb LIST_ENTRY(turnstile) ts_link; /* (q) Contested locks. */ 125122514Sjhb LIST_HEAD(, turnstile) ts_free; /* (c) Free turnstiles. */ 126122514Sjhb struct lock_object *ts_lockobj; /* (c) Lock we reference. */ 127122590Sjhb struct thread *ts_owner; /* (c + q) Who owns the lock. */ 12874912Sjhb}; 129122514Sjhb 130122514Sjhbstruct turnstile_chain { 131122514Sjhb LIST_HEAD(, turnstile) tc_turnstiles; /* List of turnstiles. */ 132122514Sjhb struct mtx tc_lock; /* Spin lock for this chain. */ 133131259Sjhb#ifdef TURNSTILE_PROFILING 134131259Sjhb u_int tc_depth; /* Length of tc_queues. */ 135131259Sjhb u_int tc_max_depth; /* Max length of tc_queues. */ 136131259Sjhb#endif 13774912Sjhb}; 13871352Sjasone 139131259Sjhb#ifdef TURNSTILE_PROFILING 140131259Sjhbu_int turnstile_max_depth; 141131259SjhbSYSCTL_NODE(_debug, OID_AUTO, turnstile, CTLFLAG_RD, 0, "turnstile profiling"); 142131259SjhbSYSCTL_NODE(_debug_turnstile, OID_AUTO, chains, CTLFLAG_RD, 0, 143131259Sjhb "turnstile chain stats"); 144131259SjhbSYSCTL_UINT(_debug_turnstile, OID_AUTO, max_depth, CTLFLAG_RD, 145131259Sjhb &turnstile_max_depth, 0, "maxmimum depth achieved of a single chain"); 146131259Sjhb#endif 147122514Sjhbstatic struct mtx td_contested_lock; 148122514Sjhbstatic struct turnstile_chain turnstile_chains[TC_TABLESIZE]; 149169666Sjeffstatic uma_zone_t turnstile_zone; 15093702Sjhb 15193702Sjhb/* 15272200Sbmilekic * Prototypes for non-exported routines. 15372200Sbmilekic */ 154122514Sjhbstatic void init_turnstile0(void *dummy); 155131263Sjhb#ifdef TURNSTILE_PROFILING 156131263Sjhbstatic void init_turnstile_profiling(void *arg); 157131263Sjhb#endif 158139453Sjhbstatic void propagate_priority(struct thread *td); 159139453Sjhbstatic int turnstile_adjust_thread(struct turnstile *ts, 160139453Sjhb struct thread *td); 161154937Sjhbstatic struct thread *turnstile_first_waiter(struct turnstile *ts); 162122514Sjhbstatic void turnstile_setowner(struct turnstile *ts, struct thread *owner); 163169666Sjeff#ifdef INVARIANTS 164169666Sjeffstatic void turnstile_dtor(void *mem, int size, void *arg); 165169666Sjeff#endif 166169666Sjeffstatic int turnstile_init(void *mem, int size, int flags); 167170295Sjeffstatic void turnstile_fini(void *mem, int size); 16867352Sjhb 169122514Sjhb/* 170122514Sjhb * Walks the chain of turnstiles and their owners to propagate the priority 171122514Sjhb * of the thread being blocked to all the threads holding locks that have to 172122514Sjhb * release their locks before this thread can run again. 173122514Sjhb */ 17467352Sjhbstatic void 17583366Sjulianpropagate_priority(struct thread *td) 17667352Sjhb{ 177122514Sjhb struct turnstile *ts; 178122514Sjhb int pri; 17967352Sjhb 180170295Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 181122514Sjhb pri = td->td_priority; 182122514Sjhb ts = td->td_blocked; 183176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 184170295Sjeff /* 185170295Sjeff * Grab a recursive lock on this turnstile chain so it stays locked 186170295Sjeff * for the whole operation. The caller expects us to return with 187170295Sjeff * the original lock held. We only ever lock down the chain so 188170295Sjeff * the lock order is constant. 189170295Sjeff */ 190170295Sjeff mtx_lock_spin(&ts->ts_lock); 19167352Sjhb for (;;) { 192122514Sjhb td = ts->ts_owner; 19367352Sjhb 19483366Sjulian if (td == NULL) { 19567352Sjhb /* 196154937Sjhb * This might be a read lock with no owner. There's 197154937Sjhb * not much we can do, so just bail. 19867352Sjhb */ 199170295Sjeff mtx_unlock_spin(&ts->ts_lock); 20067352Sjhb return; 20167352Sjhb } 20272200Sbmilekic 203170295Sjeff thread_lock_flags(td, MTX_DUPOK); 204170295Sjeff mtx_unlock_spin(&ts->ts_lock); 20599072Sjulian MPASS(td->td_proc != NULL); 20683366Sjulian MPASS(td->td_proc->p_magic == P_MAGIC); 207122514Sjhb 208122514Sjhb /* 209157275Sjhb * If the thread is asleep, then we are probably about 210157275Sjhb * to deadlock. To make debugging this easier, just 211157275Sjhb * panic and tell the user which thread misbehaved so 212157275Sjhb * they can hopefully get a stack trace from the truly 213157275Sjhb * misbehaving thread. 214122514Sjhb */ 215157275Sjhb if (TD_IS_SLEEPING(td)) { 216157275Sjhb printf( 217157275Sjhb "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n", 218157275Sjhb td->td_tid, td->td_proc->p_pid); 219157275Sjhb#ifdef DDB 220157275Sjhb db_trace_thread(td, -1); 221157275Sjhb#endif 222157275Sjhb panic("sleeping thread"); 223157275Sjhb } 224122514Sjhb 225122514Sjhb /* 226122514Sjhb * If this thread already has higher priority than the 227122514Sjhb * thread that is being blocked, we are finished. 228122514Sjhb */ 229170295Sjeff if (td->td_priority <= pri) { 230170295Sjeff thread_unlock(td); 23167352Sjhb return; 232170295Sjeff } 23369376Sjhb 23469376Sjhb /* 235139453Sjhb * Bump this thread's priority. 23667352Sjhb */ 237139453Sjhb sched_lend_prio(td, pri); 238139453Sjhb 239139453Sjhb /* 240139453Sjhb * If lock holder is actually running or on the run queue 241139453Sjhb * then we are done. 242139453Sjhb */ 243139453Sjhb if (TD_IS_RUNNING(td) || TD_ON_RUNQ(td)) { 244139453Sjhb MPASS(td->td_blocked == NULL); 245170295Sjeff thread_unlock(td); 24667352Sjhb return; 24767352Sjhb } 24872376Sjake 24973912Sjhb#ifndef SMP 25067352Sjhb /* 25183366Sjulian * For UP, we check to see if td is curthread (this shouldn't 25273912Sjhb * ever happen however as it would mean we are in a deadlock.) 25373912Sjhb */ 25483366Sjulian KASSERT(td != curthread, ("Deadlock detected")); 25573912Sjhb#endif 25673912Sjhb 25773912Sjhb /* 258122514Sjhb * If we aren't blocked on a lock, we should be. 25967352Sjhb */ 260104387Sjhb KASSERT(TD_ON_LOCK(td), ( 261139453Sjhb "thread %d(%s):%d holds %s but isn't blocked on a lock\n", 262173600Sjulian td->td_tid, td->td_name, td->td_state, 263122514Sjhb ts->ts_lockobj->lo_name)); 26467352Sjhb 26567352Sjhb /* 266122514Sjhb * Pick up the lock that td is blocked on. 26767352Sjhb */ 268122514Sjhb ts = td->td_blocked; 269122514Sjhb MPASS(ts != NULL); 270176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 271139453Sjhb /* Resort td on the list if needed. */ 272139453Sjhb if (!turnstile_adjust_thread(ts, td)) { 273170295Sjeff mtx_unlock_spin(&ts->ts_lock); 274122590Sjhb return; 275122590Sjhb } 276170295Sjeff /* The thread lock is released as ts lock above. */ 277139453Sjhb } 278139453Sjhb} 279122590Sjhb 280139453Sjhb/* 281139453Sjhb * Adjust the thread's position on a turnstile after its priority has been 282139453Sjhb * changed. 283139453Sjhb */ 284139453Sjhbstatic int 285139453Sjhbturnstile_adjust_thread(struct turnstile *ts, struct thread *td) 286139453Sjhb{ 287139453Sjhb struct thread *td1, *td2; 288154937Sjhb int queue; 28972200Sbmilekic 290170295Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 291139453Sjhb MPASS(TD_ON_LOCK(td)); 29267352Sjhb 293139453Sjhb /* 294139453Sjhb * This thread may not be blocked on this turnstile anymore 295139453Sjhb * but instead might already be woken up on another CPU 296170295Sjeff * that is waiting on the thread lock in turnstile_unpend() to 297139453Sjhb * finish waking this thread up. We can detect this case 298139453Sjhb * by checking to see if this thread has been given a 299139453Sjhb * turnstile by either turnstile_signal() or 300139453Sjhb * turnstile_broadcast(). In this case, treat the thread as 301139453Sjhb * if it was already running. 302139453Sjhb */ 303139453Sjhb if (td->td_turnstile != NULL) 304139453Sjhb return (0); 305139453Sjhb 306139453Sjhb /* 307139453Sjhb * Check if the thread needs to be moved on the blocked chain. 308139453Sjhb * It needs to be moved if either its priority is lower than 309139453Sjhb * the previous thread or higher than the next thread. 310139453Sjhb */ 311176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 312139453Sjhb td1 = TAILQ_PREV(td, threadqueue, td_lockq); 313139453Sjhb td2 = TAILQ_NEXT(td, td_lockq); 314139453Sjhb if ((td1 != NULL && td->td_priority < td1->td_priority) || 315139453Sjhb (td2 != NULL && td->td_priority > td2->td_priority)) { 316139453Sjhb 31767352Sjhb /* 31883366Sjulian * Remove thread from blocked chain and determine where 319139453Sjhb * it should be moved to. 32067352Sjhb */ 321154937Sjhb queue = td->td_tsqueue; 322154937Sjhb MPASS(queue == TS_EXCLUSIVE_QUEUE || queue == TS_SHARED_QUEUE); 323122514Sjhb mtx_lock_spin(&td_contested_lock); 324154937Sjhb TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq); 325154937Sjhb TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) { 32683366Sjulian MPASS(td1->td_proc->p_magic == P_MAGIC); 327139453Sjhb if (td1->td_priority > td->td_priority) 32867352Sjhb break; 32967352Sjhb } 33072200Sbmilekic 331139453Sjhb if (td1 == NULL) 332154937Sjhb TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq); 333139453Sjhb else 334139453Sjhb TAILQ_INSERT_BEFORE(td1, td, td_lockq); 335122514Sjhb mtx_unlock_spin(&td_contested_lock); 336139453Sjhb if (td1 == NULL) 337139453Sjhb CTR3(KTR_LOCK, 338139453Sjhb "turnstile_adjust_thread: td %d put at tail on [%p] %s", 339139453Sjhb td->td_tid, ts->ts_lockobj, ts->ts_lockobj->lo_name); 340139453Sjhb else 341139453Sjhb CTR4(KTR_LOCK, 342139453Sjhb "turnstile_adjust_thread: td %d moved before %d on [%p] %s", 343139453Sjhb td->td_tid, td1->td_tid, ts->ts_lockobj, 344139453Sjhb ts->ts_lockobj->lo_name); 34567352Sjhb } 346139453Sjhb return (1); 34767352Sjhb} 34867352Sjhb 34971352Sjasone/* 350122514Sjhb * Early initialization of turnstiles. This is not done via a SYSINIT() 351122514Sjhb * since this needs to be initialized very early when mutexes are first 352122514Sjhb * initialized. 35393609Sdes */ 354122514Sjhbvoid 355122514Sjhbinit_turnstiles(void) 35693667Sdes{ 357122514Sjhb int i; 35893667Sdes 359122514Sjhb for (i = 0; i < TC_TABLESIZE; i++) { 360122514Sjhb LIST_INIT(&turnstile_chains[i].tc_turnstiles); 361122514Sjhb mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain", 362122514Sjhb NULL, MTX_SPIN); 363131263Sjhb } 364131263Sjhb mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN); 365154482Sjhb LIST_INIT(&thread0.td_contested); 366131263Sjhb thread0.td_turnstile = NULL; 367131263Sjhb} 368131263Sjhb 369131259Sjhb#ifdef TURNSTILE_PROFILING 370131263Sjhbstatic void 371131263Sjhbinit_turnstile_profiling(void *arg) 372131263Sjhb{ 373131263Sjhb struct sysctl_oid *chain_oid; 374131263Sjhb char chain_name[10]; 375131263Sjhb int i; 376131263Sjhb 377131263Sjhb for (i = 0; i < TC_TABLESIZE; i++) { 378131259Sjhb snprintf(chain_name, sizeof(chain_name), "%d", i); 379131259Sjhb chain_oid = SYSCTL_ADD_NODE(NULL, 380131259Sjhb SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO, 381131259Sjhb chain_name, CTLFLAG_RD, NULL, "turnstile chain stats"); 382131259Sjhb SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 383131259Sjhb "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0, 384131259Sjhb NULL); 385131259Sjhb SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 386131259Sjhb "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth, 387131259Sjhb 0, NULL); 388122514Sjhb } 38993667Sdes} 390131263SjhbSYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY, 391131263Sjhb init_turnstile_profiling, NULL); 392131263Sjhb#endif 39393667Sdes 394122514Sjhbstatic void 395122514Sjhbinit_turnstile0(void *dummy) 39693609Sdes{ 39793609Sdes 398169666Sjeff turnstile_zone = uma_zcreate("TURNSTILE", sizeof(struct turnstile), 399169666Sjeff#ifdef INVARIANTS 400170295Sjeff NULL, turnstile_dtor, turnstile_init, turnstile_fini, 401170295Sjeff UMA_ALIGN_CACHE, 0); 402169666Sjeff#else 403170295Sjeff NULL, NULL, turnstile_init, turnstile_fini, UMA_ALIGN_CACHE, 0); 404169666Sjeff#endif 405122514Sjhb thread0.td_turnstile = turnstile_alloc(); 40693609Sdes} 407122514SjhbSYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL); 40893609Sdes 40993609Sdes/* 410139453Sjhb * Update a thread on the turnstile list after it's priority has been changed. 411139453Sjhb * The old priority is passed in as an argument. 412139453Sjhb */ 413139453Sjhbvoid 414139453Sjhbturnstile_adjust(struct thread *td, u_char oldpri) 415139453Sjhb{ 416139453Sjhb struct turnstile *ts; 417139453Sjhb 418139453Sjhb MPASS(TD_ON_LOCK(td)); 419139453Sjhb 420139453Sjhb /* 421139453Sjhb * Pick up the lock that td is blocked on. 422139453Sjhb */ 423139453Sjhb ts = td->td_blocked; 424139453Sjhb MPASS(ts != NULL); 425176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 426170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 427139453Sjhb 428139453Sjhb /* Resort the turnstile on the list. */ 429170295Sjeff if (!turnstile_adjust_thread(ts, td)) 430139453Sjhb return; 431139453Sjhb /* 432139453Sjhb * If our priority was lowered and we are at the head of the 433139453Sjhb * turnstile, then propagate our new priority up the chain. 434139453Sjhb * Note that we currently don't try to revoke lent priorities 435139453Sjhb * when our priority goes up. 436139453Sjhb */ 437154937Sjhb MPASS(td->td_tsqueue == TS_EXCLUSIVE_QUEUE || 438154937Sjhb td->td_tsqueue == TS_SHARED_QUEUE); 439154937Sjhb if (td == TAILQ_FIRST(&ts->ts_blocked[td->td_tsqueue]) && 440154937Sjhb td->td_priority < oldpri) { 441139453Sjhb propagate_priority(td); 442170295Sjeff } 443139453Sjhb} 444139453Sjhb 445139453Sjhb/* 446122514Sjhb * Set the owner of the lock this turnstile is attached to. 44774900Sjhb */ 448122514Sjhbstatic void 449122514Sjhbturnstile_setowner(struct turnstile *ts, struct thread *owner) 45074900Sjhb{ 45174900Sjhb 452122514Sjhb mtx_assert(&td_contested_lock, MA_OWNED); 453154937Sjhb MPASS(ts->ts_owner == NULL); 454154937Sjhb 455154937Sjhb /* A shared lock might not have an owner. */ 456154937Sjhb if (owner == NULL) 457154937Sjhb return; 458154937Sjhb 459122514Sjhb MPASS(owner->td_proc->p_magic == P_MAGIC); 460122514Sjhb ts->ts_owner = owner; 461122514Sjhb LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link); 46274900Sjhb} 46374900Sjhb 464169666Sjeff#ifdef INVARIANTS 465122514Sjhb/* 466169666Sjeff * UMA zone item deallocator. 467122514Sjhb */ 468169666Sjeffstatic void 469169666Sjeffturnstile_dtor(void *mem, int size, void *arg) 47074900Sjhb{ 471122514Sjhb struct turnstile *ts; 47274900Sjhb 473169666Sjeff ts = mem; 474169666Sjeff MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE])); 475169666Sjeff MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE])); 476169666Sjeff MPASS(TAILQ_EMPTY(&ts->ts_pending)); 477169666Sjeff} 478169666Sjeff#endif 479169666Sjeff 480169666Sjeff/* 481169666Sjeff * UMA zone item initializer. 482169666Sjeff */ 483169666Sjeffstatic int 484169666Sjeffturnstile_init(void *mem, int size, int flags) 485169666Sjeff{ 486169666Sjeff struct turnstile *ts; 487169666Sjeff 488169666Sjeff bzero(mem, size); 489169666Sjeff ts = mem; 490154937Sjhb TAILQ_INIT(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]); 491154937Sjhb TAILQ_INIT(&ts->ts_blocked[TS_SHARED_QUEUE]); 492122514Sjhb TAILQ_INIT(&ts->ts_pending); 493122514Sjhb LIST_INIT(&ts->ts_free); 494170295Sjeff mtx_init(&ts->ts_lock, "turnstile lock", NULL, MTX_SPIN | MTX_RECURSE); 495169666Sjeff return (0); 49674900Sjhb} 49774900Sjhb 498170295Sjeffstatic void 499170295Sjeffturnstile_fini(void *mem, int size) 500170295Sjeff{ 501170295Sjeff struct turnstile *ts; 502170295Sjeff 503170295Sjeff ts = mem; 504170295Sjeff mtx_destroy(&ts->ts_lock); 505170295Sjeff} 506170295Sjeff 507122514Sjhb/* 508169666Sjeff * Get a turnstile for a new thread. 509169666Sjeff */ 510169666Sjeffstruct turnstile * 511169666Sjeffturnstile_alloc(void) 512169666Sjeff{ 513169666Sjeff 514169666Sjeff return (uma_zalloc(turnstile_zone, M_WAITOK)); 515169666Sjeff} 516169666Sjeff 517169666Sjeff/* 518122514Sjhb * Free a turnstile when a thread is destroyed. 519122514Sjhb */ 52074900Sjhbvoid 521122514Sjhbturnstile_free(struct turnstile *ts) 52274900Sjhb{ 52374900Sjhb 524169666Sjeff uma_zfree(turnstile_zone, ts); 52574900Sjhb} 52674900Sjhb 52774900Sjhb/* 528136445Sjhb * Lock the turnstile chain associated with the specified lock. 529136445Sjhb */ 530136445Sjhbvoid 531170295Sjeffturnstile_chain_lock(struct lock_object *lock) 532136445Sjhb{ 533136445Sjhb struct turnstile_chain *tc; 534136445Sjhb 535136445Sjhb tc = TC_LOOKUP(lock); 536136445Sjhb mtx_lock_spin(&tc->tc_lock); 537136445Sjhb} 538136445Sjhb 539170295Sjeffstruct turnstile * 540170295Sjeffturnstile_trywait(struct lock_object *lock) 541170295Sjeff{ 542170295Sjeff struct turnstile_chain *tc; 543170295Sjeff struct turnstile *ts; 544170295Sjeff 545170295Sjeff tc = TC_LOOKUP(lock); 546170295Sjeff mtx_lock_spin(&tc->tc_lock); 547170295Sjeff LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 548170295Sjeff if (ts->ts_lockobj == lock) { 549170295Sjeff mtx_lock_spin(&ts->ts_lock); 550170295Sjeff return (ts); 551170295Sjeff } 552170295Sjeff 553170295Sjeff ts = curthread->td_turnstile; 554170295Sjeff MPASS(ts != NULL); 555170295Sjeff mtx_lock_spin(&ts->ts_lock); 556170295Sjeff KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer")); 557170295Sjeff ts->ts_lockobj = lock; 558170295Sjeff 559170295Sjeff return (ts); 560170295Sjeff} 561170295Sjeff 562170295Sjeffvoid 563170295Sjeffturnstile_cancel(struct turnstile *ts) 564170295Sjeff{ 565170295Sjeff struct turnstile_chain *tc; 566170295Sjeff struct lock_object *lock; 567170295Sjeff 568170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 569170295Sjeff 570170295Sjeff mtx_unlock_spin(&ts->ts_lock); 571170295Sjeff lock = ts->ts_lockobj; 572170295Sjeff if (ts == curthread->td_turnstile) 573170295Sjeff ts->ts_lockobj = NULL; 574170295Sjeff tc = TC_LOOKUP(lock); 575170295Sjeff mtx_unlock_spin(&tc->tc_lock); 576170295Sjeff} 577170295Sjeff 578136445Sjhb/* 579122514Sjhb * Look up the turnstile for a lock in the hash table locking the associated 580136445Sjhb * turnstile chain along the way. If no turnstile is found in the hash 581136445Sjhb * table, NULL is returned. 58271352Sjasone */ 583122514Sjhbstruct turnstile * 584122514Sjhbturnstile_lookup(struct lock_object *lock) 58571352Sjasone{ 586122514Sjhb struct turnstile_chain *tc; 587122514Sjhb struct turnstile *ts; 58871352Sjasone 589122514Sjhb tc = TC_LOOKUP(lock); 590136445Sjhb mtx_assert(&tc->tc_lock, MA_OWNED); 591122514Sjhb LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 592170295Sjeff if (ts->ts_lockobj == lock) { 593170295Sjeff mtx_lock_spin(&ts->ts_lock); 594122514Sjhb return (ts); 595170295Sjeff } 596122514Sjhb return (NULL); 59771352Sjasone} 59871352Sjasone 59971352Sjasone/* 600122514Sjhb * Unlock the turnstile chain associated with a given lock. 60171352Sjasone */ 60272200Sbmilekicvoid 603170295Sjeffturnstile_chain_unlock(struct lock_object *lock) 60471352Sjasone{ 605122514Sjhb struct turnstile_chain *tc; 60671352Sjasone 607122514Sjhb tc = TC_LOOKUP(lock); 608122514Sjhb mtx_unlock_spin(&tc->tc_lock); 60972200Sbmilekic} 61072200Sbmilekic 61172200Sbmilekic/* 612154937Sjhb * Return a pointer to the thread waiting on this turnstile with the 613154937Sjhb * most important priority or NULL if the turnstile has no waiters. 614154937Sjhb */ 615154937Sjhbstatic struct thread * 616154937Sjhbturnstile_first_waiter(struct turnstile *ts) 617154937Sjhb{ 618154937Sjhb struct thread *std, *xtd; 619154937Sjhb 620154937Sjhb std = TAILQ_FIRST(&ts->ts_blocked[TS_SHARED_QUEUE]); 621154937Sjhb xtd = TAILQ_FIRST(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]); 622154937Sjhb if (xtd == NULL || (std != NULL && std->td_priority < xtd->td_priority)) 623154937Sjhb return (std); 624154937Sjhb return (xtd); 625154937Sjhb} 626154937Sjhb 627154937Sjhb/* 628122514Sjhb * Take ownership of a turnstile and adjust the priority of the new 629122514Sjhb * owner appropriately. 63072200Sbmilekic */ 63172200Sbmilekicvoid 632170295Sjeffturnstile_claim(struct turnstile *ts) 63372200Sbmilekic{ 634170295Sjeff struct thread *td, *owner; 635122514Sjhb struct turnstile_chain *tc; 63672200Sbmilekic 637170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 638170295Sjeff MPASS(ts != curthread->td_turnstile); 63972200Sbmilekic 640122514Sjhb owner = curthread; 641122514Sjhb mtx_lock_spin(&td_contested_lock); 642122514Sjhb turnstile_setowner(ts, owner); 643122514Sjhb mtx_unlock_spin(&td_contested_lock); 64472200Sbmilekic 645154937Sjhb td = turnstile_first_waiter(ts); 646122514Sjhb MPASS(td != NULL); 647122514Sjhb MPASS(td->td_proc->p_magic == P_MAGIC); 648176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 64972200Sbmilekic 650122514Sjhb /* 651122514Sjhb * Update the priority of the new owner if needed. 652122514Sjhb */ 653170295Sjeff thread_lock(owner); 654122514Sjhb if (td->td_priority < owner->td_priority) 655139453Sjhb sched_lend_prio(owner, td->td_priority); 656170295Sjeff thread_unlock(owner); 657170295Sjeff tc = TC_LOOKUP(ts->ts_lockobj); 658170295Sjeff mtx_unlock_spin(&ts->ts_lock); 659170295Sjeff mtx_unlock_spin(&tc->tc_lock); 66067352Sjhb} 66167352Sjhb 66272200Sbmilekic/* 663136445Sjhb * Block the current thread on the turnstile assicated with 'lock'. This 664136445Sjhb * function will context switch and not return until this thread has been 665136445Sjhb * woken back up. This function must be called with the appropriate 666136445Sjhb * turnstile chain locked and will return with it unlocked. 66772200Sbmilekic */ 66867352Sjhbvoid 669170295Sjeffturnstile_wait(struct turnstile *ts, struct thread *owner, int queue) 67067352Sjhb{ 671122514Sjhb struct turnstile_chain *tc; 67283366Sjulian struct thread *td, *td1; 673170295Sjeff struct lock_object *lock; 67467352Sjhb 67583366Sjulian td = curthread; 676170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 677154937Sjhb if (owner) 678154937Sjhb MPASS(owner->td_proc->p_magic == P_MAGIC); 679154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 68072200Sbmilekic 681136445Sjhb /* 682136445Sjhb * If the lock does not already have a turnstile, use this thread's 683136445Sjhb * turnstile. Otherwise insert the current thread into the 684136445Sjhb * turnstile already in use by this lock. 685136445Sjhb */ 686170295Sjeff tc = TC_LOOKUP(ts->ts_lockobj); 687170295Sjeff if (ts == td->td_turnstile) { 688170295Sjeff mtx_assert(&tc->tc_lock, MA_OWNED); 689131259Sjhb#ifdef TURNSTILE_PROFILING 690131259Sjhb tc->tc_depth++; 691131259Sjhb if (tc->tc_depth > tc->tc_max_depth) { 692131259Sjhb tc->tc_max_depth = tc->tc_depth; 693131259Sjhb if (tc->tc_max_depth > turnstile_max_depth) 694131259Sjhb turnstile_max_depth = tc->tc_max_depth; 695131259Sjhb } 696131259Sjhb#endif 697170295Sjeff tc = TC_LOOKUP(ts->ts_lockobj); 698122514Sjhb LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash); 699122514Sjhb KASSERT(TAILQ_EMPTY(&ts->ts_pending), 700122514Sjhb ("thread's turnstile has pending threads")); 701154937Sjhb KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]), 702154937Sjhb ("thread's turnstile has exclusive waiters")); 703154937Sjhb KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]), 704154937Sjhb ("thread's turnstile has shared waiters")); 705122514Sjhb KASSERT(LIST_EMPTY(&ts->ts_free), 706122514Sjhb ("thread's turnstile has a non-empty free list")); 707170295Sjeff MPASS(ts->ts_lockobj != NULL); 708122514Sjhb mtx_lock_spin(&td_contested_lock); 709154937Sjhb TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq); 710122514Sjhb turnstile_setowner(ts, owner); 711122514Sjhb mtx_unlock_spin(&td_contested_lock); 712122514Sjhb } else { 713154937Sjhb TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) 714122514Sjhb if (td1->td_priority > td->td_priority) 715122514Sjhb break; 716122514Sjhb mtx_lock_spin(&td_contested_lock); 717122514Sjhb if (td1 != NULL) 718122514Sjhb TAILQ_INSERT_BEFORE(td1, td, td_lockq); 719122514Sjhb else 720154937Sjhb TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq); 721154937Sjhb MPASS(owner == ts->ts_owner); 722122514Sjhb mtx_unlock_spin(&td_contested_lock); 723122514Sjhb MPASS(td->td_turnstile != NULL); 724122514Sjhb LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash); 72572200Sbmilekic } 726170295Sjeff thread_lock(td); 727170295Sjeff thread_lock_set(td, &ts->ts_lock); 728122514Sjhb td->td_turnstile = NULL; 72972200Sbmilekic 730122514Sjhb /* Save who we are blocked on and switch. */ 731170295Sjeff lock = ts->ts_lockobj; 732154937Sjhb td->td_tsqueue = queue; 733122514Sjhb td->td_blocked = ts; 734122514Sjhb td->td_lockname = lock->lo_name; 735122514Sjhb TD_SET_LOCK(td); 736170295Sjeff mtx_unlock_spin(&tc->tc_lock); 737122514Sjhb propagate_priority(td); 73872200Sbmilekic 739122514Sjhb if (LOCK_LOG_TEST(lock, 0)) 740139453Sjhb CTR4(KTR_LOCK, "%s: td %d blocked on [%p] %s", __func__, 741139453Sjhb td->td_tid, lock, lock->lo_name); 74272200Sbmilekic 743176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 744170295Sjeff SCHED_STAT_INC(switch_turnstile); 745131473Sjhb mi_switch(SW_VOL, NULL); 74672200Sbmilekic 747122514Sjhb if (LOCK_LOG_TEST(lock, 0)) 748139453Sjhb CTR4(KTR_LOCK, "%s: td %d free from blocked on [%p] %s", 749139453Sjhb __func__, td->td_tid, lock, lock->lo_name); 750170295Sjeff thread_unlock(td); 75167352Sjhb} 75267352Sjhb 75372200Sbmilekic/* 754122514Sjhb * Pick the highest priority thread on this turnstile and put it on the 755122514Sjhb * pending list. This must be called with the turnstile chain locked. 75672200Sbmilekic */ 757122514Sjhbint 758154937Sjhbturnstile_signal(struct turnstile *ts, int queue) 75971352Sjasone{ 760122514Sjhb struct turnstile_chain *tc; 761122514Sjhb struct thread *td; 762122514Sjhb int empty; 76380748Sjhb 764122514Sjhb MPASS(ts != NULL); 765170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 766122514Sjhb MPASS(curthread->td_proc->p_magic == P_MAGIC); 767176017Sjeff MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL); 768154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 76971352Sjasone 770122514Sjhb /* 771122514Sjhb * Pick the highest priority thread blocked on this lock and 772122514Sjhb * move it to the pending list. 773122514Sjhb */ 774154937Sjhb td = TAILQ_FIRST(&ts->ts_blocked[queue]); 775122514Sjhb MPASS(td->td_proc->p_magic == P_MAGIC); 776122514Sjhb mtx_lock_spin(&td_contested_lock); 777154937Sjhb TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq); 778122514Sjhb mtx_unlock_spin(&td_contested_lock); 779122514Sjhb TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq); 78067352Sjhb 78182304Sbmilekic /* 782122514Sjhb * If the turnstile is now empty, remove it from its chain and 783122514Sjhb * give it to the about-to-be-woken thread. Otherwise take a 784122514Sjhb * turnstile from the free list and give it to the thread. 785105782Sdes */ 786154937Sjhb empty = TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) && 787154937Sjhb TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]); 788131259Sjhb if (empty) { 789170295Sjeff tc = TC_LOOKUP(ts->ts_lockobj); 790170295Sjeff mtx_assert(&tc->tc_lock, MA_OWNED); 791122514Sjhb MPASS(LIST_EMPTY(&ts->ts_free)); 792131259Sjhb#ifdef TURNSTILE_PROFILING 793131259Sjhb tc->tc_depth--; 794131259Sjhb#endif 795131259Sjhb } else 796122514Sjhb ts = LIST_FIRST(&ts->ts_free); 797123363Sjhb MPASS(ts != NULL); 798122514Sjhb LIST_REMOVE(ts, ts_hash); 799122514Sjhb td->td_turnstile = ts; 800122514Sjhb 801122514Sjhb return (empty); 80267352Sjhb} 803122514Sjhb 80472200Sbmilekic/* 805122514Sjhb * Put all blocked threads on the pending list. This must be called with 806122514Sjhb * the turnstile chain locked. 80793672Sarr */ 80893672Sarrvoid 809154937Sjhbturnstile_broadcast(struct turnstile *ts, int queue) 81093672Sarr{ 811122514Sjhb struct turnstile_chain *tc; 812122514Sjhb struct turnstile *ts1; 813122514Sjhb struct thread *td; 81493672Sarr 815122514Sjhb MPASS(ts != NULL); 816170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 817122514Sjhb MPASS(curthread->td_proc->p_magic == P_MAGIC); 818176017Sjeff MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL); 819170295Sjeff /* 820170295Sjeff * We must have the chain locked so that we can remove the empty 821170295Sjeff * turnstile from the hash queue. 822170295Sjeff */ 823122514Sjhb tc = TC_LOOKUP(ts->ts_lockobj); 824122514Sjhb mtx_assert(&tc->tc_lock, MA_OWNED); 825154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 826122514Sjhb 827122514Sjhb /* 828122514Sjhb * Transfer the blocked list to the pending list. 829122514Sjhb */ 830122514Sjhb mtx_lock_spin(&td_contested_lock); 831154937Sjhb TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked[queue], td_lockq); 832122514Sjhb mtx_unlock_spin(&td_contested_lock); 833122514Sjhb 834122514Sjhb /* 835122514Sjhb * Give a turnstile to each thread. The last thread gets 836154937Sjhb * this turnstile if the turnstile is empty. 837122514Sjhb */ 838122514Sjhb TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) { 839122514Sjhb if (LIST_EMPTY(&ts->ts_free)) { 840122514Sjhb MPASS(TAILQ_NEXT(td, td_lockq) == NULL); 841122514Sjhb ts1 = ts; 842131259Sjhb#ifdef TURNSTILE_PROFILING 843131259Sjhb tc->tc_depth--; 844131259Sjhb#endif 845122514Sjhb } else 846122514Sjhb ts1 = LIST_FIRST(&ts->ts_free); 847123363Sjhb MPASS(ts1 != NULL); 848122514Sjhb LIST_REMOVE(ts1, ts_hash); 849122514Sjhb td->td_turnstile = ts1; 850122514Sjhb } 85193672Sarr} 852122590Sjhb 85393672Sarr/* 854122514Sjhb * Wakeup all threads on the pending list and adjust the priority of the 855122514Sjhb * current thread appropriately. This must be called with the turnstile 856122514Sjhb * chain locked. 857105782Sdes */ 85867352Sjhbvoid 859154937Sjhbturnstile_unpend(struct turnstile *ts, int owner_type) 86067352Sjhb{ 861122514Sjhb TAILQ_HEAD( ,thread) pending_threads; 862170295Sjeff struct turnstile *nts; 863122514Sjhb struct thread *td; 864139453Sjhb u_char cp, pri; 86572200Sbmilekic 866122514Sjhb MPASS(ts != NULL); 867170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 868176017Sjeff MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL); 869122514Sjhb MPASS(!TAILQ_EMPTY(&ts->ts_pending)); 87072200Sbmilekic 871122514Sjhb /* 872122514Sjhb * Move the list of pending threads out of the turnstile and 873122514Sjhb * into a local variable. 874122514Sjhb */ 875122514Sjhb TAILQ_INIT(&pending_threads); 876122514Sjhb TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq); 877122514Sjhb#ifdef INVARIANTS 878154937Sjhb if (TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) && 879154937Sjhb TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE])) 880122514Sjhb ts->ts_lockobj = NULL; 88169429Sjhb#endif 882122514Sjhb /* 883170295Sjeff * Adjust the priority of curthread based on other contested 884170295Sjeff * locks it owns. Don't lower the priority below the base 885170295Sjeff * priority however. 886170295Sjeff */ 887170295Sjeff td = curthread; 888170295Sjeff pri = PRI_MAX; 889170295Sjeff thread_lock(td); 890170295Sjeff mtx_lock_spin(&td_contested_lock); 891170295Sjeff /* 892122514Sjhb * Remove the turnstile from this thread's list of contested locks 893122514Sjhb * since this thread doesn't own it anymore. New threads will 894122514Sjhb * not be blocking on the turnstile until it is claimed by a new 895154937Sjhb * owner. There might not be a current owner if this is a shared 896154937Sjhb * lock. 897122514Sjhb */ 898154937Sjhb if (ts->ts_owner != NULL) { 899154937Sjhb ts->ts_owner = NULL; 900154937Sjhb LIST_REMOVE(ts, ts_link); 901154937Sjhb } 902170295Sjeff LIST_FOREACH(nts, &td->td_contested, ts_link) { 903170295Sjeff cp = turnstile_first_waiter(nts)->td_priority; 904122514Sjhb if (cp < pri) 905122514Sjhb pri = cp; 906122514Sjhb } 907122514Sjhb mtx_unlock_spin(&td_contested_lock); 908139453Sjhb sched_unlend_prio(td, pri); 909170295Sjeff thread_unlock(td); 910122514Sjhb /* 911122514Sjhb * Wake up all the pending threads. If a thread is not blocked 912122514Sjhb * on a lock, then it is currently executing on another CPU in 913123364Sjhb * turnstile_wait() or sitting on a run queue waiting to resume 914123364Sjhb * in turnstile_wait(). Set a flag to force it to try to acquire 915122514Sjhb * the lock again instead of blocking. 916122514Sjhb */ 917122514Sjhb while (!TAILQ_EMPTY(&pending_threads)) { 918122514Sjhb td = TAILQ_FIRST(&pending_threads); 919122514Sjhb TAILQ_REMOVE(&pending_threads, td, td_lockq); 920170295Sjeff thread_lock(td); 921176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 922122514Sjhb MPASS(td->td_proc->p_magic == P_MAGIC); 923170295Sjeff MPASS(TD_ON_LOCK(td)); 924170295Sjeff TD_CLR_LOCK(td); 925170295Sjeff MPASS(TD_CAN_RUN(td)); 926170295Sjeff td->td_blocked = NULL; 927170295Sjeff td->td_lockname = NULL; 928154937Sjhb#ifdef INVARIANTS 929170295Sjeff td->td_tsqueue = 0xff; 930154937Sjhb#endif 931170295Sjeff sched_add(td, SRQ_BORING); 932170295Sjeff thread_unlock(td); 933122514Sjhb } 934170295Sjeff mtx_unlock_spin(&ts->ts_lock); 93567352Sjhb} 93667352Sjhb 93772200Sbmilekic/* 938157844Sjhb * Give up ownership of a turnstile. This must be called with the 939157844Sjhb * turnstile chain locked. 940157844Sjhb */ 941157844Sjhbvoid 942157844Sjhbturnstile_disown(struct turnstile *ts) 943157844Sjhb{ 944157844Sjhb struct thread *td; 945157844Sjhb u_char cp, pri; 946157844Sjhb 947157844Sjhb MPASS(ts != NULL); 948170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 949157844Sjhb MPASS(ts->ts_owner == curthread); 950157844Sjhb MPASS(TAILQ_EMPTY(&ts->ts_pending)); 951157844Sjhb MPASS(!TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) || 952157844Sjhb !TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE])); 953157844Sjhb 954157844Sjhb /* 955157844Sjhb * Remove the turnstile from this thread's list of contested locks 956157844Sjhb * since this thread doesn't own it anymore. New threads will 957157844Sjhb * not be blocking on the turnstile until it is claimed by a new 958157844Sjhb * owner. 959157844Sjhb */ 960157844Sjhb mtx_lock_spin(&td_contested_lock); 961157844Sjhb ts->ts_owner = NULL; 962157844Sjhb LIST_REMOVE(ts, ts_link); 963157844Sjhb mtx_unlock_spin(&td_contested_lock); 964157844Sjhb 965157844Sjhb /* 966157844Sjhb * Adjust the priority of curthread based on other contested 967157844Sjhb * locks it owns. Don't lower the priority below the base 968157844Sjhb * priority however. 969157844Sjhb */ 970157844Sjhb td = curthread; 971157844Sjhb pri = PRI_MAX; 972170295Sjeff thread_lock(td); 973170295Sjeff mtx_unlock_spin(&ts->ts_lock); 974157844Sjhb mtx_lock_spin(&td_contested_lock); 975157844Sjhb LIST_FOREACH(ts, &td->td_contested, ts_link) { 976157844Sjhb cp = turnstile_first_waiter(ts)->td_priority; 977157844Sjhb if (cp < pri) 978157844Sjhb pri = cp; 979157844Sjhb } 980157844Sjhb mtx_unlock_spin(&td_contested_lock); 981157844Sjhb sched_unlend_prio(td, pri); 982170295Sjeff thread_unlock(td); 983157844Sjhb} 984157844Sjhb 985157844Sjhb/* 986122514Sjhb * Return the first thread in a turnstile. 98772200Sbmilekic */ 988122514Sjhbstruct thread * 989154937Sjhbturnstile_head(struct turnstile *ts, int queue) 99067352Sjhb{ 991122514Sjhb#ifdef INVARIANTS 99267352Sjhb 993122514Sjhb MPASS(ts != NULL); 994154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 995170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 996122514Sjhb#endif 997154937Sjhb return (TAILQ_FIRST(&ts->ts_blocked[queue])); 99871320Sjasone} 999154937Sjhb 1000157844Sjhb/* 1001157844Sjhb * Returns true if a sub-queue of a turnstile is empty. 1002157844Sjhb */ 1003157844Sjhbint 1004157844Sjhbturnstile_empty(struct turnstile *ts, int queue) 1005157844Sjhb{ 1006157844Sjhb#ifdef INVARIANTS 1007157844Sjhb 1008157844Sjhb MPASS(ts != NULL); 1009157844Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 1010170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 1011157844Sjhb#endif 1012157844Sjhb return (TAILQ_EMPTY(&ts->ts_blocked[queue])); 1013157844Sjhb} 1014157844Sjhb 1015154937Sjhb#ifdef DDB 1016154937Sjhbstatic void 1017154937Sjhbprint_thread(struct thread *td, const char *prefix) 1018154937Sjhb{ 1019154937Sjhb 1020154937Sjhb db_printf("%s%p (tid %d, pid %d, \"%s\")\n", prefix, td, td->td_tid, 1021157952Sjhb td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name : 1022173600Sjulian td->td_name); 1023154937Sjhb} 1024154937Sjhb 1025154937Sjhbstatic void 1026154937Sjhbprint_queue(struct threadqueue *queue, const char *header, const char *prefix) 1027154937Sjhb{ 1028154937Sjhb struct thread *td; 1029154937Sjhb 1030154937Sjhb db_printf("%s:\n", header); 1031154937Sjhb if (TAILQ_EMPTY(queue)) { 1032154937Sjhb db_printf("%sempty\n", prefix); 1033154937Sjhb return; 1034154937Sjhb } 1035154937Sjhb TAILQ_FOREACH(td, queue, td_lockq) { 1036154937Sjhb print_thread(td, prefix); 1037154937Sjhb } 1038154937Sjhb} 1039154937Sjhb 1040154937SjhbDB_SHOW_COMMAND(turnstile, db_show_turnstile) 1041154937Sjhb{ 1042154937Sjhb struct turnstile_chain *tc; 1043154937Sjhb struct turnstile *ts; 1044154937Sjhb struct lock_object *lock; 1045154937Sjhb int i; 1046154937Sjhb 1047154937Sjhb if (!have_addr) 1048154937Sjhb return; 1049154937Sjhb 1050154937Sjhb /* 1051154937Sjhb * First, see if there is an active turnstile for the lock indicated 1052154937Sjhb * by the address. 1053154937Sjhb */ 1054154937Sjhb lock = (struct lock_object *)addr; 1055154937Sjhb tc = TC_LOOKUP(lock); 1056154937Sjhb LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 1057154937Sjhb if (ts->ts_lockobj == lock) 1058154937Sjhb goto found; 1059154937Sjhb 1060154937Sjhb /* 1061154937Sjhb * Second, see if there is an active turnstile at the address 1062154937Sjhb * indicated. 1063154937Sjhb */ 1064154937Sjhb for (i = 0; i < TC_TABLESIZE; i++) 1065154937Sjhb LIST_FOREACH(ts, &turnstile_chains[i].tc_turnstiles, ts_hash) { 1066154937Sjhb if (ts == (struct turnstile *)addr) 1067154937Sjhb goto found; 1068154937Sjhb } 1069154937Sjhb 1070154937Sjhb db_printf("Unable to locate a turnstile via %p\n", (void *)addr); 1071154937Sjhb return; 1072154937Sjhbfound: 1073154937Sjhb lock = ts->ts_lockobj; 1074154937Sjhb db_printf("Lock: %p - (%s) %s\n", lock, LOCK_CLASS(lock)->lc_name, 1075154937Sjhb lock->lo_name); 1076154937Sjhb if (ts->ts_owner) 1077154937Sjhb print_thread(ts->ts_owner, "Lock Owner: "); 1078154937Sjhb else 1079154937Sjhb db_printf("Lock Owner: none\n"); 1080154937Sjhb print_queue(&ts->ts_blocked[TS_SHARED_QUEUE], "Shared Waiters", "\t"); 1081154937Sjhb print_queue(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE], "Exclusive Waiters", 1082154937Sjhb "\t"); 1083154937Sjhb print_queue(&ts->ts_pending, "Pending Threads", "\t"); 1084154937Sjhb 1085154937Sjhb} 1086158031Sjhb 1087161324Sjhb/* 1088161324Sjhb * Show all the threads a particular thread is waiting on based on 1089161324Sjhb * non-sleepable and non-spin locks. 1090161324Sjhb */ 1091158031Sjhbstatic void 1092161324Sjhbprint_lockchain(struct thread *td, const char *prefix) 1093158031Sjhb{ 1094158031Sjhb struct lock_object *lock; 1095158031Sjhb struct lock_class *class; 1096158031Sjhb struct turnstile *ts; 1097158031Sjhb 1098158031Sjhb /* 1099158031Sjhb * Follow the chain. We keep walking as long as the thread is 1100158031Sjhb * blocked on a turnstile that has an owner. 1101158031Sjhb */ 1102160313Sjhb while (!db_pager_quit) { 1103158031Sjhb db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid, 1104158031Sjhb td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name : 1105173600Sjulian td->td_name); 1106158031Sjhb switch (td->td_state) { 1107158031Sjhb case TDS_INACTIVE: 1108158031Sjhb db_printf("is inactive\n"); 1109158031Sjhb return; 1110158031Sjhb case TDS_CAN_RUN: 1111158031Sjhb db_printf("can run\n"); 1112158031Sjhb return; 1113158031Sjhb case TDS_RUNQ: 1114158031Sjhb db_printf("is on a run queue\n"); 1115158031Sjhb return; 1116158031Sjhb case TDS_RUNNING: 1117158031Sjhb db_printf("running on CPU %d\n", td->td_oncpu); 1118158031Sjhb return; 1119158031Sjhb case TDS_INHIBITED: 1120158031Sjhb if (TD_ON_LOCK(td)) { 1121158031Sjhb ts = td->td_blocked; 1122158031Sjhb lock = ts->ts_lockobj; 1123158031Sjhb class = LOCK_CLASS(lock); 1124158031Sjhb db_printf("blocked on lock %p (%s) \"%s\"\n", 1125158031Sjhb lock, class->lc_name, lock->lo_name); 1126158031Sjhb if (ts->ts_owner == NULL) 1127158031Sjhb return; 1128158031Sjhb td = ts->ts_owner; 1129158031Sjhb break; 1130158031Sjhb } 1131158031Sjhb db_printf("inhibited\n"); 1132158031Sjhb return; 1133158031Sjhb default: 1134158031Sjhb db_printf("??? (%#x)\n", td->td_state); 1135158031Sjhb return; 1136158031Sjhb } 1137158031Sjhb } 1138158031Sjhb} 1139158031Sjhb 1140161324SjhbDB_SHOW_COMMAND(lockchain, db_show_lockchain) 1141158031Sjhb{ 1142158031Sjhb struct thread *td; 1143158031Sjhb 1144158031Sjhb /* Figure out which thread to start with. */ 1145158031Sjhb if (have_addr) 1146158031Sjhb td = db_lookup_thread(addr, TRUE); 1147158031Sjhb else 1148158031Sjhb td = kdb_thread; 1149158031Sjhb 1150161324Sjhb print_lockchain(td, ""); 1151158031Sjhb} 1152158031Sjhb 1153158031SjhbDB_SHOW_COMMAND(allchains, db_show_allchains) 1154158031Sjhb{ 1155158031Sjhb struct thread *td; 1156158031Sjhb struct proc *p; 1157158031Sjhb int i; 1158158031Sjhb 1159158031Sjhb i = 1; 1160166073Sdelphij FOREACH_PROC_IN_SYSTEM(p) { 1161158031Sjhb FOREACH_THREAD_IN_PROC(p, td) { 1162158031Sjhb if (TD_ON_LOCK(td) && LIST_EMPTY(&td->td_contested)) { 1163158031Sjhb db_printf("chain %d:\n", i++); 1164161324Sjhb print_lockchain(td, " "); 1165158031Sjhb } 1166160313Sjhb if (db_pager_quit) 1167160313Sjhb return; 1168158031Sjhb } 1169158031Sjhb } 1170158031Sjhb} 1171158031Sjhb 1172161337Sjhb/* 1173161337Sjhb * Show all the threads a particular thread is waiting on based on 1174161337Sjhb * sleepable locks. 1175161337Sjhb */ 1176161337Sjhbstatic void 1177161337Sjhbprint_sleepchain(struct thread *td, const char *prefix) 1178161337Sjhb{ 1179161337Sjhb struct thread *owner; 1180161337Sjhb 1181161337Sjhb /* 1182161337Sjhb * Follow the chain. We keep walking as long as the thread is 1183161337Sjhb * blocked on a sleep lock that has an owner. 1184161337Sjhb */ 1185161337Sjhb while (!db_pager_quit) { 1186161337Sjhb db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid, 1187161337Sjhb td->td_proc->p_pid, td->td_name[0] != '\0' ? td->td_name : 1188173600Sjulian td->td_name); 1189161337Sjhb switch (td->td_state) { 1190161337Sjhb case TDS_INACTIVE: 1191161337Sjhb db_printf("is inactive\n"); 1192161337Sjhb return; 1193161337Sjhb case TDS_CAN_RUN: 1194161337Sjhb db_printf("can run\n"); 1195161337Sjhb return; 1196161337Sjhb case TDS_RUNQ: 1197161337Sjhb db_printf("is on a run queue\n"); 1198161337Sjhb return; 1199161337Sjhb case TDS_RUNNING: 1200161337Sjhb db_printf("running on CPU %d\n", td->td_oncpu); 1201161337Sjhb return; 1202161337Sjhb case TDS_INHIBITED: 1203161337Sjhb if (TD_ON_SLEEPQ(td)) { 1204161337Sjhb if (lockmgr_chain(td, &owner) || 1205161337Sjhb sx_chain(td, &owner)) { 1206161337Sjhb if (owner == NULL) 1207161337Sjhb return; 1208161337Sjhb td = owner; 1209161337Sjhb break; 1210161337Sjhb } 1211161337Sjhb db_printf("sleeping on %p \"%s\"\n", 1212161337Sjhb td->td_wchan, td->td_wmesg); 1213161337Sjhb return; 1214161337Sjhb } 1215161337Sjhb db_printf("inhibited\n"); 1216161337Sjhb return; 1217161337Sjhb default: 1218161337Sjhb db_printf("??? (%#x)\n", td->td_state); 1219161337Sjhb return; 1220161337Sjhb } 1221161337Sjhb } 1222161337Sjhb} 1223161337Sjhb 1224161337SjhbDB_SHOW_COMMAND(sleepchain, db_show_sleepchain) 1225161337Sjhb{ 1226161337Sjhb struct thread *td; 1227161337Sjhb 1228161337Sjhb /* Figure out which thread to start with. */ 1229161337Sjhb if (have_addr) 1230161337Sjhb td = db_lookup_thread(addr, TRUE); 1231161337Sjhb else 1232161337Sjhb td = kdb_thread; 1233161337Sjhb 1234161337Sjhb print_sleepchain(td, ""); 1235161337Sjhb} 1236161337Sjhb 1237158031Sjhbstatic void print_waiters(struct turnstile *ts, int indent); 1238158031Sjhb 1239158031Sjhbstatic void 1240158031Sjhbprint_waiter(struct thread *td, int indent) 1241158031Sjhb{ 1242158031Sjhb struct turnstile *ts; 1243158031Sjhb int i; 1244158031Sjhb 1245160313Sjhb if (db_pager_quit) 1246160313Sjhb return; 1247158031Sjhb for (i = 0; i < indent; i++) 1248158031Sjhb db_printf(" "); 1249158031Sjhb print_thread(td, "thread "); 1250158031Sjhb LIST_FOREACH(ts, &td->td_contested, ts_link) 1251158031Sjhb print_waiters(ts, indent + 1); 1252158031Sjhb} 1253158031Sjhb 1254158031Sjhbstatic void 1255158031Sjhbprint_waiters(struct turnstile *ts, int indent) 1256158031Sjhb{ 1257158031Sjhb struct lock_object *lock; 1258158031Sjhb struct lock_class *class; 1259158031Sjhb struct thread *td; 1260158031Sjhb int i; 1261158031Sjhb 1262160313Sjhb if (db_pager_quit) 1263160313Sjhb return; 1264158031Sjhb lock = ts->ts_lockobj; 1265158031Sjhb class = LOCK_CLASS(lock); 1266158031Sjhb for (i = 0; i < indent; i++) 1267158031Sjhb db_printf(" "); 1268158031Sjhb db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, lock->lo_name); 1269158031Sjhb TAILQ_FOREACH(td, &ts->ts_blocked[TS_EXCLUSIVE_QUEUE], td_lockq) 1270158031Sjhb print_waiter(td, indent + 1); 1271158031Sjhb TAILQ_FOREACH(td, &ts->ts_blocked[TS_SHARED_QUEUE], td_lockq) 1272158031Sjhb print_waiter(td, indent + 1); 1273158031Sjhb TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) 1274158031Sjhb print_waiter(td, indent + 1); 1275158031Sjhb} 1276158031Sjhb 1277161324SjhbDB_SHOW_COMMAND(locktree, db_show_locktree) 1278158031Sjhb{ 1279158031Sjhb struct lock_object *lock; 1280158031Sjhb struct lock_class *class; 1281158031Sjhb struct turnstile_chain *tc; 1282158031Sjhb struct turnstile *ts; 1283158031Sjhb 1284158031Sjhb if (!have_addr) 1285158031Sjhb return; 1286158031Sjhb lock = (struct lock_object *)addr; 1287158031Sjhb tc = TC_LOOKUP(lock); 1288158031Sjhb LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 1289158031Sjhb if (ts->ts_lockobj == lock) 1290158031Sjhb break; 1291158031Sjhb if (ts == NULL) { 1292158031Sjhb class = LOCK_CLASS(lock); 1293158031Sjhb db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, 1294158031Sjhb lock->lo_name); 1295158031Sjhb } else 1296158031Sjhb print_waiters(ts, 0); 1297158031Sjhb} 1298154937Sjhb#endif 1299