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("$FreeBSD$"); 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> 68234280Smarius#include <sys/kdb.h> 6967352Sjhb#include <sys/kernel.h> 7093609Sdes#include <sys/ktr.h> 7176166Smarkm#include <sys/lock.h> 7274912Sjhb#include <sys/mutex.h> 7365557Sjasone#include <sys/proc.h> 74122514Sjhb#include <sys/queue.h> 75131259Sjhb#include <sys/sched.h> 76235459Srstone#include <sys/sdt.h> 77131259Sjhb#include <sys/sysctl.h> 78122514Sjhb#include <sys/turnstile.h> 7965557Sjasone 80169666Sjeff#include <vm/uma.h> 81169666Sjeff 82154937Sjhb#ifdef DDB 83154937Sjhb#include <ddb/ddb.h> 84161337Sjhb#include <sys/lockmgr.h> 85161337Sjhb#include <sys/sx.h> 86154937Sjhb#endif 87154937Sjhb 8865557Sjasone/* 89122514Sjhb * Constants for the hash table of turnstile chains. TC_SHIFT is a magic 90122514Sjhb * number chosen because the sleep queue's use the same value for the 91122514Sjhb * shift. Basically, we ignore the lower 8 bits of the address. 92122514Sjhb * TC_TABLESIZE must be a power of two for TC_MASK to work properly. 9371352Sjasone */ 94122514Sjhb#define TC_TABLESIZE 128 /* Must be power of 2. */ 95122514Sjhb#define TC_MASK (TC_TABLESIZE - 1) 96122514Sjhb#define TC_SHIFT 8 97122514Sjhb#define TC_HASH(lock) (((uintptr_t)(lock) >> TC_SHIFT) & TC_MASK) 98122514Sjhb#define TC_LOOKUP(lock) &turnstile_chains[TC_HASH(lock)] 9971352Sjasone 10071352Sjasone/* 101122514Sjhb * There are three different lists of turnstiles as follows. The list 102122514Sjhb * connected by ts_link entries is a per-thread list of all the turnstiles 103122514Sjhb * attached to locks that we own. This is used to fixup our priority when 104122514Sjhb * a lock is released. The other two lists use the ts_hash entries. The 105126317Sjhb * first of these two is the turnstile chain list that a turnstile is on 106126317Sjhb * when it is attached to a lock. The second list to use ts_hash is the 107126317Sjhb * free list hung off of a turnstile that is attached to a lock. 108122514Sjhb * 109154937Sjhb * Each turnstile contains three lists of threads. The two ts_blocked lists 110154937Sjhb * are linked list of threads blocked on the turnstile's lock. One list is 111154937Sjhb * for exclusive waiters, and the other is for shared waiters. The 112126884Sjhb * ts_pending list is a linked list of threads previously awakened by 113122514Sjhb * turnstile_signal() or turnstile_wait() that are waiting to be put on 114122514Sjhb * the run queue. 115122514Sjhb * 116122514Sjhb * Locking key: 117122514Sjhb * c - turnstile chain lock 118122514Sjhb * q - td_contested lock 11971352Sjasone */ 120122514Sjhbstruct turnstile { 121170295Sjeff struct mtx ts_lock; /* Spin lock for self. */ 122154937Sjhb struct threadqueue ts_blocked[2]; /* (c + q) Blocked threads. */ 123154937Sjhb struct threadqueue ts_pending; /* (c) Pending threads. */ 124122514Sjhb LIST_ENTRY(turnstile) ts_hash; /* (c) Chain and free list. */ 125122514Sjhb LIST_ENTRY(turnstile) ts_link; /* (q) Contested locks. */ 126122514Sjhb LIST_HEAD(, turnstile) ts_free; /* (c) Free turnstiles. */ 127122514Sjhb struct lock_object *ts_lockobj; /* (c) Lock we reference. */ 128122590Sjhb struct thread *ts_owner; /* (c + q) Who owns the lock. */ 12974912Sjhb}; 130122514Sjhb 131122514Sjhbstruct turnstile_chain { 132122514Sjhb LIST_HEAD(, turnstile) tc_turnstiles; /* List of turnstiles. */ 133122514Sjhb struct mtx tc_lock; /* Spin lock for this chain. */ 134131259Sjhb#ifdef TURNSTILE_PROFILING 135131259Sjhb u_int tc_depth; /* Length of tc_queues. */ 136131259Sjhb u_int tc_max_depth; /* Max length of tc_queues. */ 137131259Sjhb#endif 13874912Sjhb}; 13971352Sjasone 140131259Sjhb#ifdef TURNSTILE_PROFILING 141131259Sjhbu_int turnstile_max_depth; 142227309Sedstatic SYSCTL_NODE(_debug, OID_AUTO, turnstile, CTLFLAG_RD, 0, 143227309Sed "turnstile profiling"); 144227309Sedstatic SYSCTL_NODE(_debug_turnstile, OID_AUTO, chains, CTLFLAG_RD, 0, 145131259Sjhb "turnstile chain stats"); 146131259SjhbSYSCTL_UINT(_debug_turnstile, OID_AUTO, max_depth, CTLFLAG_RD, 147234303Sdavide &turnstile_max_depth, 0, "maximum depth achieved of a single chain"); 148131259Sjhb#endif 149122514Sjhbstatic struct mtx td_contested_lock; 150122514Sjhbstatic struct turnstile_chain turnstile_chains[TC_TABLESIZE]; 151169666Sjeffstatic uma_zone_t turnstile_zone; 15293702Sjhb 15393702Sjhb/* 15472200Sbmilekic * Prototypes for non-exported routines. 15572200Sbmilekic */ 156122514Sjhbstatic void init_turnstile0(void *dummy); 157131263Sjhb#ifdef TURNSTILE_PROFILING 158131263Sjhbstatic void init_turnstile_profiling(void *arg); 159131263Sjhb#endif 160139453Sjhbstatic void propagate_priority(struct thread *td); 161139453Sjhbstatic int turnstile_adjust_thread(struct turnstile *ts, 162139453Sjhb struct thread *td); 163154937Sjhbstatic struct thread *turnstile_first_waiter(struct turnstile *ts); 164122514Sjhbstatic void turnstile_setowner(struct turnstile *ts, struct thread *owner); 165169666Sjeff#ifdef INVARIANTS 166169666Sjeffstatic void turnstile_dtor(void *mem, int size, void *arg); 167169666Sjeff#endif 168169666Sjeffstatic int turnstile_init(void *mem, int size, int flags); 169170295Sjeffstatic void turnstile_fini(void *mem, int size); 17067352Sjhb 171235459SrstoneSDT_PROVIDER_DECLARE(sched); 172258622SavgSDT_PROBE_DEFINE(sched, , , sleep); 173258622SavgSDT_PROBE_DEFINE2(sched, , , wakeup, "struct thread *", 174235459Srstone "struct proc *"); 175235459Srstone 176122514Sjhb/* 177122514Sjhb * Walks the chain of turnstiles and their owners to propagate the priority 178122514Sjhb * of the thread being blocked to all the threads holding locks that have to 179122514Sjhb * release their locks before this thread can run again. 180122514Sjhb */ 18167352Sjhbstatic void 18283366Sjulianpropagate_priority(struct thread *td) 18367352Sjhb{ 184122514Sjhb struct turnstile *ts; 185122514Sjhb int pri; 18667352Sjhb 187170295Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 188122514Sjhb pri = td->td_priority; 189122514Sjhb ts = td->td_blocked; 190176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 191170295Sjeff /* 192170295Sjeff * Grab a recursive lock on this turnstile chain so it stays locked 193170295Sjeff * for the whole operation. The caller expects us to return with 194170295Sjeff * the original lock held. We only ever lock down the chain so 195170295Sjeff * the lock order is constant. 196170295Sjeff */ 197170295Sjeff mtx_lock_spin(&ts->ts_lock); 19867352Sjhb for (;;) { 199122514Sjhb td = ts->ts_owner; 20067352Sjhb 20183366Sjulian if (td == NULL) { 20267352Sjhb /* 203154937Sjhb * This might be a read lock with no owner. There's 204154937Sjhb * not much we can do, so just bail. 20567352Sjhb */ 206170295Sjeff mtx_unlock_spin(&ts->ts_lock); 20767352Sjhb return; 20867352Sjhb } 20972200Sbmilekic 210170295Sjeff thread_lock_flags(td, MTX_DUPOK); 211170295Sjeff mtx_unlock_spin(&ts->ts_lock); 21299072Sjulian MPASS(td->td_proc != NULL); 21383366Sjulian MPASS(td->td_proc->p_magic == P_MAGIC); 214122514Sjhb 215122514Sjhb /* 216157275Sjhb * If the thread is asleep, then we are probably about 217246923Spjd * to deadlock. To make debugging this easier, show 218246923Spjd * backtrace of misbehaving thread and panic to not 219246923Spjd * leave the kernel deadlocked. 220122514Sjhb */ 221157275Sjhb if (TD_IS_SLEEPING(td)) { 222157275Sjhb printf( 223157275Sjhb "Sleeping thread (tid %d, pid %d) owns a non-sleepable lock\n", 224157275Sjhb td->td_tid, td->td_proc->p_pid); 225234190Sjhb kdb_backtrace_thread(td); 226157275Sjhb panic("sleeping thread"); 227157275Sjhb } 228122514Sjhb 229122514Sjhb /* 230122514Sjhb * If this thread already has higher priority than the 231122514Sjhb * thread that is being blocked, we are finished. 232122514Sjhb */ 233170295Sjeff if (td->td_priority <= pri) { 234170295Sjeff thread_unlock(td); 23567352Sjhb return; 236170295Sjeff } 23769376Sjhb 23869376Sjhb /* 239139453Sjhb * Bump this thread's priority. 24067352Sjhb */ 241139453Sjhb sched_lend_prio(td, pri); 242139453Sjhb 243139453Sjhb /* 244139453Sjhb * If lock holder is actually running or on the run queue 245139453Sjhb * then we are done. 246139453Sjhb */ 247139453Sjhb if (TD_IS_RUNNING(td) || TD_ON_RUNQ(td)) { 248139453Sjhb MPASS(td->td_blocked == NULL); 249170295Sjeff thread_unlock(td); 25067352Sjhb return; 25167352Sjhb } 25272376Sjake 25373912Sjhb#ifndef SMP 25467352Sjhb /* 25583366Sjulian * For UP, we check to see if td is curthread (this shouldn't 25673912Sjhb * ever happen however as it would mean we are in a deadlock.) 25773912Sjhb */ 25883366Sjulian KASSERT(td != curthread, ("Deadlock detected")); 25973912Sjhb#endif 26073912Sjhb 26173912Sjhb /* 262122514Sjhb * If we aren't blocked on a lock, we should be. 26367352Sjhb */ 264104387Sjhb KASSERT(TD_ON_LOCK(td), ( 265139453Sjhb "thread %d(%s):%d holds %s but isn't blocked on a lock\n", 266173600Sjulian td->td_tid, td->td_name, td->td_state, 267122514Sjhb ts->ts_lockobj->lo_name)); 26867352Sjhb 26967352Sjhb /* 270122514Sjhb * Pick up the lock that td is blocked on. 27167352Sjhb */ 272122514Sjhb ts = td->td_blocked; 273122514Sjhb MPASS(ts != NULL); 274176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 275139453Sjhb /* Resort td on the list if needed. */ 276139453Sjhb if (!turnstile_adjust_thread(ts, td)) { 277170295Sjeff mtx_unlock_spin(&ts->ts_lock); 278122590Sjhb return; 279122590Sjhb } 280170295Sjeff /* The thread lock is released as ts lock above. */ 281139453Sjhb } 282139453Sjhb} 283122590Sjhb 284139453Sjhb/* 285139453Sjhb * Adjust the thread's position on a turnstile after its priority has been 286139453Sjhb * changed. 287139453Sjhb */ 288139453Sjhbstatic int 289139453Sjhbturnstile_adjust_thread(struct turnstile *ts, struct thread *td) 290139453Sjhb{ 291139453Sjhb struct thread *td1, *td2; 292154937Sjhb int queue; 29372200Sbmilekic 294170295Sjeff THREAD_LOCK_ASSERT(td, MA_OWNED); 295139453Sjhb MPASS(TD_ON_LOCK(td)); 29667352Sjhb 297139453Sjhb /* 298139453Sjhb * This thread may not be blocked on this turnstile anymore 299139453Sjhb * but instead might already be woken up on another CPU 300170295Sjeff * that is waiting on the thread lock in turnstile_unpend() to 301139453Sjhb * finish waking this thread up. We can detect this case 302139453Sjhb * by checking to see if this thread has been given a 303139453Sjhb * turnstile by either turnstile_signal() or 304139453Sjhb * turnstile_broadcast(). In this case, treat the thread as 305139453Sjhb * if it was already running. 306139453Sjhb */ 307139453Sjhb if (td->td_turnstile != NULL) 308139453Sjhb return (0); 309139453Sjhb 310139453Sjhb /* 311139453Sjhb * Check if the thread needs to be moved on the blocked chain. 312139453Sjhb * It needs to be moved if either its priority is lower than 313139453Sjhb * the previous thread or higher than the next thread. 314139453Sjhb */ 315176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 316139453Sjhb td1 = TAILQ_PREV(td, threadqueue, td_lockq); 317139453Sjhb td2 = TAILQ_NEXT(td, td_lockq); 318139453Sjhb if ((td1 != NULL && td->td_priority < td1->td_priority) || 319139453Sjhb (td2 != NULL && td->td_priority > td2->td_priority)) { 320139453Sjhb 32167352Sjhb /* 32283366Sjulian * Remove thread from blocked chain and determine where 323139453Sjhb * it should be moved to. 32467352Sjhb */ 325154937Sjhb queue = td->td_tsqueue; 326154937Sjhb MPASS(queue == TS_EXCLUSIVE_QUEUE || queue == TS_SHARED_QUEUE); 327122514Sjhb mtx_lock_spin(&td_contested_lock); 328154937Sjhb TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq); 329154937Sjhb TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) { 33083366Sjulian MPASS(td1->td_proc->p_magic == P_MAGIC); 331139453Sjhb if (td1->td_priority > td->td_priority) 33267352Sjhb break; 33367352Sjhb } 33472200Sbmilekic 335139453Sjhb if (td1 == NULL) 336154937Sjhb TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq); 337139453Sjhb else 338139453Sjhb TAILQ_INSERT_BEFORE(td1, td, td_lockq); 339122514Sjhb mtx_unlock_spin(&td_contested_lock); 340139453Sjhb if (td1 == NULL) 341139453Sjhb CTR3(KTR_LOCK, 342139453Sjhb "turnstile_adjust_thread: td %d put at tail on [%p] %s", 343139453Sjhb td->td_tid, ts->ts_lockobj, ts->ts_lockobj->lo_name); 344139453Sjhb else 345139453Sjhb CTR4(KTR_LOCK, 346139453Sjhb "turnstile_adjust_thread: td %d moved before %d on [%p] %s", 347139453Sjhb td->td_tid, td1->td_tid, ts->ts_lockobj, 348139453Sjhb ts->ts_lockobj->lo_name); 34967352Sjhb } 350139453Sjhb return (1); 35167352Sjhb} 35267352Sjhb 35371352Sjasone/* 354122514Sjhb * Early initialization of turnstiles. This is not done via a SYSINIT() 355122514Sjhb * since this needs to be initialized very early when mutexes are first 356122514Sjhb * initialized. 35793609Sdes */ 358122514Sjhbvoid 359122514Sjhbinit_turnstiles(void) 36093667Sdes{ 361122514Sjhb int i; 36293667Sdes 363122514Sjhb for (i = 0; i < TC_TABLESIZE; i++) { 364122514Sjhb LIST_INIT(&turnstile_chains[i].tc_turnstiles); 365122514Sjhb mtx_init(&turnstile_chains[i].tc_lock, "turnstile chain", 366122514Sjhb NULL, MTX_SPIN); 367131263Sjhb } 368131263Sjhb mtx_init(&td_contested_lock, "td_contested", NULL, MTX_SPIN); 369154482Sjhb LIST_INIT(&thread0.td_contested); 370131263Sjhb thread0.td_turnstile = NULL; 371131263Sjhb} 372131263Sjhb 373131259Sjhb#ifdef TURNSTILE_PROFILING 374131263Sjhbstatic void 375131263Sjhbinit_turnstile_profiling(void *arg) 376131263Sjhb{ 377131263Sjhb struct sysctl_oid *chain_oid; 378131263Sjhb char chain_name[10]; 379131263Sjhb int i; 380131263Sjhb 381131263Sjhb for (i = 0; i < TC_TABLESIZE; i++) { 382131259Sjhb snprintf(chain_name, sizeof(chain_name), "%d", i); 383131259Sjhb chain_oid = SYSCTL_ADD_NODE(NULL, 384131259Sjhb SYSCTL_STATIC_CHILDREN(_debug_turnstile_chains), OID_AUTO, 385131259Sjhb chain_name, CTLFLAG_RD, NULL, "turnstile chain stats"); 386131259Sjhb SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 387131259Sjhb "depth", CTLFLAG_RD, &turnstile_chains[i].tc_depth, 0, 388131259Sjhb NULL); 389131259Sjhb SYSCTL_ADD_UINT(NULL, SYSCTL_CHILDREN(chain_oid), OID_AUTO, 390131259Sjhb "max_depth", CTLFLAG_RD, &turnstile_chains[i].tc_max_depth, 391131259Sjhb 0, NULL); 392122514Sjhb } 39393667Sdes} 394131263SjhbSYSINIT(turnstile_profiling, SI_SUB_LOCK, SI_ORDER_ANY, 395131263Sjhb init_turnstile_profiling, NULL); 396131263Sjhb#endif 39793667Sdes 398122514Sjhbstatic void 399122514Sjhbinit_turnstile0(void *dummy) 40093609Sdes{ 40193609Sdes 402169666Sjeff turnstile_zone = uma_zcreate("TURNSTILE", sizeof(struct turnstile), 403182879Sjhb NULL, 404169666Sjeff#ifdef INVARIANTS 405182879Sjhb turnstile_dtor, 406169666Sjeff#else 407182879Sjhb NULL, 408169666Sjeff#endif 409182879Sjhb turnstile_init, turnstile_fini, UMA_ALIGN_CACHE, UMA_ZONE_NOFREE); 410122514Sjhb thread0.td_turnstile = turnstile_alloc(); 41193609Sdes} 412122514SjhbSYSINIT(turnstile0, SI_SUB_LOCK, SI_ORDER_ANY, init_turnstile0, NULL); 41393609Sdes 41493609Sdes/* 415139453Sjhb * Update a thread on the turnstile list after it's priority has been changed. 416139453Sjhb * The old priority is passed in as an argument. 417139453Sjhb */ 418139453Sjhbvoid 419139453Sjhbturnstile_adjust(struct thread *td, u_char oldpri) 420139453Sjhb{ 421139453Sjhb struct turnstile *ts; 422139453Sjhb 423139453Sjhb MPASS(TD_ON_LOCK(td)); 424139453Sjhb 425139453Sjhb /* 426139453Sjhb * Pick up the lock that td is blocked on. 427139453Sjhb */ 428139453Sjhb ts = td->td_blocked; 429139453Sjhb MPASS(ts != NULL); 430176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 431170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 432139453Sjhb 433139453Sjhb /* Resort the turnstile on the list. */ 434170295Sjeff if (!turnstile_adjust_thread(ts, td)) 435139453Sjhb return; 436139453Sjhb /* 437139453Sjhb * If our priority was lowered and we are at the head of the 438139453Sjhb * turnstile, then propagate our new priority up the chain. 439139453Sjhb * Note that we currently don't try to revoke lent priorities 440139453Sjhb * when our priority goes up. 441139453Sjhb */ 442154937Sjhb MPASS(td->td_tsqueue == TS_EXCLUSIVE_QUEUE || 443154937Sjhb td->td_tsqueue == TS_SHARED_QUEUE); 444154937Sjhb if (td == TAILQ_FIRST(&ts->ts_blocked[td->td_tsqueue]) && 445154937Sjhb td->td_priority < oldpri) { 446139453Sjhb propagate_priority(td); 447170295Sjeff } 448139453Sjhb} 449139453Sjhb 450139453Sjhb/* 451122514Sjhb * Set the owner of the lock this turnstile is attached to. 45274900Sjhb */ 453122514Sjhbstatic void 454122514Sjhbturnstile_setowner(struct turnstile *ts, struct thread *owner) 45574900Sjhb{ 45674900Sjhb 457122514Sjhb mtx_assert(&td_contested_lock, MA_OWNED); 458154937Sjhb MPASS(ts->ts_owner == NULL); 459154937Sjhb 460154937Sjhb /* A shared lock might not have an owner. */ 461154937Sjhb if (owner == NULL) 462154937Sjhb return; 463154937Sjhb 464122514Sjhb MPASS(owner->td_proc->p_magic == P_MAGIC); 465122514Sjhb ts->ts_owner = owner; 466122514Sjhb LIST_INSERT_HEAD(&owner->td_contested, ts, ts_link); 46774900Sjhb} 46874900Sjhb 469169666Sjeff#ifdef INVARIANTS 470122514Sjhb/* 471169666Sjeff * UMA zone item deallocator. 472122514Sjhb */ 473169666Sjeffstatic void 474169666Sjeffturnstile_dtor(void *mem, int size, void *arg) 47574900Sjhb{ 476122514Sjhb struct turnstile *ts; 47774900Sjhb 478169666Sjeff ts = mem; 479169666Sjeff MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE])); 480169666Sjeff MPASS(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE])); 481169666Sjeff MPASS(TAILQ_EMPTY(&ts->ts_pending)); 482169666Sjeff} 483169666Sjeff#endif 484169666Sjeff 485169666Sjeff/* 486169666Sjeff * UMA zone item initializer. 487169666Sjeff */ 488169666Sjeffstatic int 489169666Sjeffturnstile_init(void *mem, int size, int flags) 490169666Sjeff{ 491169666Sjeff struct turnstile *ts; 492169666Sjeff 493169666Sjeff bzero(mem, size); 494169666Sjeff ts = mem; 495154937Sjhb TAILQ_INIT(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]); 496154937Sjhb TAILQ_INIT(&ts->ts_blocked[TS_SHARED_QUEUE]); 497122514Sjhb TAILQ_INIT(&ts->ts_pending); 498122514Sjhb LIST_INIT(&ts->ts_free); 499170295Sjeff mtx_init(&ts->ts_lock, "turnstile lock", NULL, MTX_SPIN | MTX_RECURSE); 500169666Sjeff return (0); 50174900Sjhb} 50274900Sjhb 503170295Sjeffstatic void 504170295Sjeffturnstile_fini(void *mem, int size) 505170295Sjeff{ 506170295Sjeff struct turnstile *ts; 507170295Sjeff 508170295Sjeff ts = mem; 509170295Sjeff mtx_destroy(&ts->ts_lock); 510170295Sjeff} 511170295Sjeff 512122514Sjhb/* 513169666Sjeff * Get a turnstile for a new thread. 514169666Sjeff */ 515169666Sjeffstruct turnstile * 516169666Sjeffturnstile_alloc(void) 517169666Sjeff{ 518169666Sjeff 519169666Sjeff return (uma_zalloc(turnstile_zone, M_WAITOK)); 520169666Sjeff} 521169666Sjeff 522169666Sjeff/* 523122514Sjhb * Free a turnstile when a thread is destroyed. 524122514Sjhb */ 52574900Sjhbvoid 526122514Sjhbturnstile_free(struct turnstile *ts) 52774900Sjhb{ 52874900Sjhb 529169666Sjeff uma_zfree(turnstile_zone, ts); 53074900Sjhb} 53174900Sjhb 53274900Sjhb/* 533136445Sjhb * Lock the turnstile chain associated with the specified lock. 534136445Sjhb */ 535136445Sjhbvoid 536170295Sjeffturnstile_chain_lock(struct lock_object *lock) 537136445Sjhb{ 538136445Sjhb struct turnstile_chain *tc; 539136445Sjhb 540136445Sjhb tc = TC_LOOKUP(lock); 541136445Sjhb mtx_lock_spin(&tc->tc_lock); 542136445Sjhb} 543136445Sjhb 544170295Sjeffstruct turnstile * 545170295Sjeffturnstile_trywait(struct lock_object *lock) 546170295Sjeff{ 547170295Sjeff struct turnstile_chain *tc; 548170295Sjeff struct turnstile *ts; 549170295Sjeff 550170295Sjeff tc = TC_LOOKUP(lock); 551170295Sjeff mtx_lock_spin(&tc->tc_lock); 552170295Sjeff LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 553170295Sjeff if (ts->ts_lockobj == lock) { 554170295Sjeff mtx_lock_spin(&ts->ts_lock); 555170295Sjeff return (ts); 556170295Sjeff } 557170295Sjeff 558170295Sjeff ts = curthread->td_turnstile; 559170295Sjeff MPASS(ts != NULL); 560170295Sjeff mtx_lock_spin(&ts->ts_lock); 561170295Sjeff KASSERT(ts->ts_lockobj == NULL, ("stale ts_lockobj pointer")); 562170295Sjeff ts->ts_lockobj = lock; 563170295Sjeff 564170295Sjeff return (ts); 565170295Sjeff} 566170295Sjeff 567170295Sjeffvoid 568170295Sjeffturnstile_cancel(struct turnstile *ts) 569170295Sjeff{ 570170295Sjeff struct turnstile_chain *tc; 571170295Sjeff struct lock_object *lock; 572170295Sjeff 573170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 574170295Sjeff 575170295Sjeff mtx_unlock_spin(&ts->ts_lock); 576170295Sjeff lock = ts->ts_lockobj; 577170295Sjeff if (ts == curthread->td_turnstile) 578170295Sjeff ts->ts_lockobj = NULL; 579170295Sjeff tc = TC_LOOKUP(lock); 580170295Sjeff mtx_unlock_spin(&tc->tc_lock); 581170295Sjeff} 582170295Sjeff 583136445Sjhb/* 584122514Sjhb * Look up the turnstile for a lock in the hash table locking the associated 585136445Sjhb * turnstile chain along the way. If no turnstile is found in the hash 586136445Sjhb * table, NULL is returned. 58771352Sjasone */ 588122514Sjhbstruct turnstile * 589122514Sjhbturnstile_lookup(struct lock_object *lock) 59071352Sjasone{ 591122514Sjhb struct turnstile_chain *tc; 592122514Sjhb struct turnstile *ts; 59371352Sjasone 594122514Sjhb tc = TC_LOOKUP(lock); 595136445Sjhb mtx_assert(&tc->tc_lock, MA_OWNED); 596122514Sjhb LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 597170295Sjeff if (ts->ts_lockobj == lock) { 598170295Sjeff mtx_lock_spin(&ts->ts_lock); 599122514Sjhb return (ts); 600170295Sjeff } 601122514Sjhb return (NULL); 60271352Sjasone} 60371352Sjasone 60471352Sjasone/* 605122514Sjhb * Unlock the turnstile chain associated with a given lock. 60671352Sjasone */ 60772200Sbmilekicvoid 608170295Sjeffturnstile_chain_unlock(struct lock_object *lock) 60971352Sjasone{ 610122514Sjhb struct turnstile_chain *tc; 61171352Sjasone 612122514Sjhb tc = TC_LOOKUP(lock); 613122514Sjhb mtx_unlock_spin(&tc->tc_lock); 61472200Sbmilekic} 61572200Sbmilekic 61672200Sbmilekic/* 617154937Sjhb * Return a pointer to the thread waiting on this turnstile with the 618154937Sjhb * most important priority or NULL if the turnstile has no waiters. 619154937Sjhb */ 620154937Sjhbstatic struct thread * 621154937Sjhbturnstile_first_waiter(struct turnstile *ts) 622154937Sjhb{ 623154937Sjhb struct thread *std, *xtd; 624154937Sjhb 625154937Sjhb std = TAILQ_FIRST(&ts->ts_blocked[TS_SHARED_QUEUE]); 626154937Sjhb xtd = TAILQ_FIRST(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]); 627154937Sjhb if (xtd == NULL || (std != NULL && std->td_priority < xtd->td_priority)) 628154937Sjhb return (std); 629154937Sjhb return (xtd); 630154937Sjhb} 631154937Sjhb 632154937Sjhb/* 633122514Sjhb * Take ownership of a turnstile and adjust the priority of the new 634122514Sjhb * owner appropriately. 63572200Sbmilekic */ 63672200Sbmilekicvoid 637170295Sjeffturnstile_claim(struct turnstile *ts) 63872200Sbmilekic{ 639170295Sjeff struct thread *td, *owner; 640122514Sjhb struct turnstile_chain *tc; 64172200Sbmilekic 642170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 643170295Sjeff MPASS(ts != curthread->td_turnstile); 64472200Sbmilekic 645122514Sjhb owner = curthread; 646122514Sjhb mtx_lock_spin(&td_contested_lock); 647122514Sjhb turnstile_setowner(ts, owner); 648122514Sjhb mtx_unlock_spin(&td_contested_lock); 64972200Sbmilekic 650154937Sjhb td = turnstile_first_waiter(ts); 651122514Sjhb MPASS(td != NULL); 652122514Sjhb MPASS(td->td_proc->p_magic == P_MAGIC); 653176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 65472200Sbmilekic 655122514Sjhb /* 656122514Sjhb * Update the priority of the new owner if needed. 657122514Sjhb */ 658170295Sjeff thread_lock(owner); 659122514Sjhb if (td->td_priority < owner->td_priority) 660139453Sjhb sched_lend_prio(owner, td->td_priority); 661170295Sjeff thread_unlock(owner); 662170295Sjeff tc = TC_LOOKUP(ts->ts_lockobj); 663170295Sjeff mtx_unlock_spin(&ts->ts_lock); 664170295Sjeff mtx_unlock_spin(&tc->tc_lock); 66567352Sjhb} 66667352Sjhb 66772200Sbmilekic/* 668136445Sjhb * Block the current thread on the turnstile assicated with 'lock'. This 669136445Sjhb * function will context switch and not return until this thread has been 670136445Sjhb * woken back up. This function must be called with the appropriate 671136445Sjhb * turnstile chain locked and will return with it unlocked. 67272200Sbmilekic */ 67367352Sjhbvoid 674170295Sjeffturnstile_wait(struct turnstile *ts, struct thread *owner, int queue) 67567352Sjhb{ 676122514Sjhb struct turnstile_chain *tc; 67783366Sjulian struct thread *td, *td1; 678170295Sjeff struct lock_object *lock; 67967352Sjhb 68083366Sjulian td = curthread; 681170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 682154937Sjhb if (owner) 683154937Sjhb MPASS(owner->td_proc->p_magic == P_MAGIC); 684154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 68572200Sbmilekic 686136445Sjhb /* 687136445Sjhb * If the lock does not already have a turnstile, use this thread's 688136445Sjhb * turnstile. Otherwise insert the current thread into the 689136445Sjhb * turnstile already in use by this lock. 690136445Sjhb */ 691170295Sjeff tc = TC_LOOKUP(ts->ts_lockobj); 692218272Sjhb mtx_assert(&tc->tc_lock, MA_OWNED); 693170295Sjeff if (ts == td->td_turnstile) { 694131259Sjhb#ifdef TURNSTILE_PROFILING 695131259Sjhb tc->tc_depth++; 696131259Sjhb if (tc->tc_depth > tc->tc_max_depth) { 697131259Sjhb tc->tc_max_depth = tc->tc_depth; 698131259Sjhb if (tc->tc_max_depth > turnstile_max_depth) 699131259Sjhb turnstile_max_depth = tc->tc_max_depth; 700131259Sjhb } 701131259Sjhb#endif 702122514Sjhb LIST_INSERT_HEAD(&tc->tc_turnstiles, ts, ts_hash); 703122514Sjhb KASSERT(TAILQ_EMPTY(&ts->ts_pending), 704122514Sjhb ("thread's turnstile has pending threads")); 705154937Sjhb KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]), 706154937Sjhb ("thread's turnstile has exclusive waiters")); 707154937Sjhb KASSERT(TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]), 708154937Sjhb ("thread's turnstile has shared waiters")); 709122514Sjhb KASSERT(LIST_EMPTY(&ts->ts_free), 710122514Sjhb ("thread's turnstile has a non-empty free list")); 711170295Sjeff MPASS(ts->ts_lockobj != NULL); 712122514Sjhb mtx_lock_spin(&td_contested_lock); 713154937Sjhb TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq); 714122514Sjhb turnstile_setowner(ts, owner); 715122514Sjhb mtx_unlock_spin(&td_contested_lock); 716122514Sjhb } else { 717154937Sjhb TAILQ_FOREACH(td1, &ts->ts_blocked[queue], td_lockq) 718122514Sjhb if (td1->td_priority > td->td_priority) 719122514Sjhb break; 720122514Sjhb mtx_lock_spin(&td_contested_lock); 721122514Sjhb if (td1 != NULL) 722122514Sjhb TAILQ_INSERT_BEFORE(td1, td, td_lockq); 723122514Sjhb else 724154937Sjhb TAILQ_INSERT_TAIL(&ts->ts_blocked[queue], td, td_lockq); 725154937Sjhb MPASS(owner == ts->ts_owner); 726122514Sjhb mtx_unlock_spin(&td_contested_lock); 727122514Sjhb MPASS(td->td_turnstile != NULL); 728122514Sjhb LIST_INSERT_HEAD(&ts->ts_free, td->td_turnstile, ts_hash); 72972200Sbmilekic } 730170295Sjeff thread_lock(td); 731170295Sjeff thread_lock_set(td, &ts->ts_lock); 732122514Sjhb td->td_turnstile = NULL; 73372200Sbmilekic 734122514Sjhb /* Save who we are blocked on and switch. */ 735170295Sjeff lock = ts->ts_lockobj; 736154937Sjhb td->td_tsqueue = queue; 737122514Sjhb td->td_blocked = ts; 738122514Sjhb td->td_lockname = lock->lo_name; 739201879Sattilio td->td_blktick = ticks; 740122514Sjhb TD_SET_LOCK(td); 741170295Sjeff mtx_unlock_spin(&tc->tc_lock); 742122514Sjhb propagate_priority(td); 74372200Sbmilekic 744122514Sjhb if (LOCK_LOG_TEST(lock, 0)) 745139453Sjhb CTR4(KTR_LOCK, "%s: td %d blocked on [%p] %s", __func__, 746139453Sjhb td->td_tid, lock, lock->lo_name); 74772200Sbmilekic 748235459Srstone SDT_PROBE0(sched, , , sleep); 749235459Srstone 750176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 751178272Sjeff mi_switch(SW_VOL | SWT_TURNSTILE, NULL); 75272200Sbmilekic 753122514Sjhb if (LOCK_LOG_TEST(lock, 0)) 754139453Sjhb CTR4(KTR_LOCK, "%s: td %d free from blocked on [%p] %s", 755139453Sjhb __func__, td->td_tid, lock, lock->lo_name); 756170295Sjeff thread_unlock(td); 75767352Sjhb} 75867352Sjhb 75972200Sbmilekic/* 760122514Sjhb * Pick the highest priority thread on this turnstile and put it on the 761122514Sjhb * pending list. This must be called with the turnstile chain locked. 76272200Sbmilekic */ 763122514Sjhbint 764154937Sjhbturnstile_signal(struct turnstile *ts, int queue) 76571352Sjasone{ 766122514Sjhb struct turnstile_chain *tc; 767122514Sjhb struct thread *td; 768122514Sjhb int empty; 76980748Sjhb 770122514Sjhb MPASS(ts != NULL); 771170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 772122514Sjhb MPASS(curthread->td_proc->p_magic == P_MAGIC); 773176017Sjeff MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL); 774154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 77571352Sjasone 776122514Sjhb /* 777122514Sjhb * Pick the highest priority thread blocked on this lock and 778122514Sjhb * move it to the pending list. 779122514Sjhb */ 780154937Sjhb td = TAILQ_FIRST(&ts->ts_blocked[queue]); 781122514Sjhb MPASS(td->td_proc->p_magic == P_MAGIC); 782122514Sjhb mtx_lock_spin(&td_contested_lock); 783154937Sjhb TAILQ_REMOVE(&ts->ts_blocked[queue], td, td_lockq); 784122514Sjhb mtx_unlock_spin(&td_contested_lock); 785122514Sjhb TAILQ_INSERT_TAIL(&ts->ts_pending, td, td_lockq); 78667352Sjhb 78782304Sbmilekic /* 788122514Sjhb * If the turnstile is now empty, remove it from its chain and 789122514Sjhb * give it to the about-to-be-woken thread. Otherwise take a 790122514Sjhb * turnstile from the free list and give it to the thread. 791105782Sdes */ 792154937Sjhb empty = TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) && 793154937Sjhb TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE]); 794131259Sjhb if (empty) { 795170295Sjeff tc = TC_LOOKUP(ts->ts_lockobj); 796170295Sjeff mtx_assert(&tc->tc_lock, MA_OWNED); 797122514Sjhb MPASS(LIST_EMPTY(&ts->ts_free)); 798131259Sjhb#ifdef TURNSTILE_PROFILING 799131259Sjhb tc->tc_depth--; 800131259Sjhb#endif 801131259Sjhb } else 802122514Sjhb ts = LIST_FIRST(&ts->ts_free); 803123363Sjhb MPASS(ts != NULL); 804122514Sjhb LIST_REMOVE(ts, ts_hash); 805122514Sjhb td->td_turnstile = ts; 806122514Sjhb 807122514Sjhb return (empty); 80867352Sjhb} 809122514Sjhb 81072200Sbmilekic/* 811122514Sjhb * Put all blocked threads on the pending list. This must be called with 812122514Sjhb * the turnstile chain locked. 81393672Sarr */ 81493672Sarrvoid 815154937Sjhbturnstile_broadcast(struct turnstile *ts, int queue) 81693672Sarr{ 817122514Sjhb struct turnstile_chain *tc; 818122514Sjhb struct turnstile *ts1; 819122514Sjhb struct thread *td; 82093672Sarr 821122514Sjhb MPASS(ts != NULL); 822170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 823122514Sjhb MPASS(curthread->td_proc->p_magic == P_MAGIC); 824176017Sjeff MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL); 825170295Sjeff /* 826170295Sjeff * We must have the chain locked so that we can remove the empty 827170295Sjeff * turnstile from the hash queue. 828170295Sjeff */ 829122514Sjhb tc = TC_LOOKUP(ts->ts_lockobj); 830122514Sjhb mtx_assert(&tc->tc_lock, MA_OWNED); 831154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 832122514Sjhb 833122514Sjhb /* 834122514Sjhb * Transfer the blocked list to the pending list. 835122514Sjhb */ 836122514Sjhb mtx_lock_spin(&td_contested_lock); 837154937Sjhb TAILQ_CONCAT(&ts->ts_pending, &ts->ts_blocked[queue], td_lockq); 838122514Sjhb mtx_unlock_spin(&td_contested_lock); 839122514Sjhb 840122514Sjhb /* 841122514Sjhb * Give a turnstile to each thread. The last thread gets 842154937Sjhb * this turnstile if the turnstile is empty. 843122514Sjhb */ 844122514Sjhb TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) { 845122514Sjhb if (LIST_EMPTY(&ts->ts_free)) { 846122514Sjhb MPASS(TAILQ_NEXT(td, td_lockq) == NULL); 847122514Sjhb ts1 = ts; 848131259Sjhb#ifdef TURNSTILE_PROFILING 849131259Sjhb tc->tc_depth--; 850131259Sjhb#endif 851122514Sjhb } else 852122514Sjhb ts1 = LIST_FIRST(&ts->ts_free); 853123363Sjhb MPASS(ts1 != NULL); 854122514Sjhb LIST_REMOVE(ts1, ts_hash); 855122514Sjhb td->td_turnstile = ts1; 856122514Sjhb } 85793672Sarr} 858122590Sjhb 85993672Sarr/* 860122514Sjhb * Wakeup all threads on the pending list and adjust the priority of the 861122514Sjhb * current thread appropriately. This must be called with the turnstile 862122514Sjhb * chain locked. 863105782Sdes */ 86467352Sjhbvoid 865154937Sjhbturnstile_unpend(struct turnstile *ts, int owner_type) 86667352Sjhb{ 867122514Sjhb TAILQ_HEAD( ,thread) pending_threads; 868170295Sjeff struct turnstile *nts; 869122514Sjhb struct thread *td; 870139453Sjhb u_char cp, pri; 87172200Sbmilekic 872122514Sjhb MPASS(ts != NULL); 873170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 874176017Sjeff MPASS(ts->ts_owner == curthread || ts->ts_owner == NULL); 875122514Sjhb MPASS(!TAILQ_EMPTY(&ts->ts_pending)); 87672200Sbmilekic 877122514Sjhb /* 878122514Sjhb * Move the list of pending threads out of the turnstile and 879122514Sjhb * into a local variable. 880122514Sjhb */ 881122514Sjhb TAILQ_INIT(&pending_threads); 882122514Sjhb TAILQ_CONCAT(&pending_threads, &ts->ts_pending, td_lockq); 883122514Sjhb#ifdef INVARIANTS 884154937Sjhb if (TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) && 885154937Sjhb TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE])) 886122514Sjhb ts->ts_lockobj = NULL; 88769429Sjhb#endif 888122514Sjhb /* 889170295Sjeff * Adjust the priority of curthread based on other contested 890170295Sjeff * locks it owns. Don't lower the priority below the base 891170295Sjeff * priority however. 892170295Sjeff */ 893170295Sjeff td = curthread; 894170295Sjeff pri = PRI_MAX; 895170295Sjeff thread_lock(td); 896170295Sjeff mtx_lock_spin(&td_contested_lock); 897170295Sjeff /* 898122514Sjhb * Remove the turnstile from this thread's list of contested locks 899122514Sjhb * since this thread doesn't own it anymore. New threads will 900122514Sjhb * not be blocking on the turnstile until it is claimed by a new 901154937Sjhb * owner. There might not be a current owner if this is a shared 902154937Sjhb * lock. 903122514Sjhb */ 904154937Sjhb if (ts->ts_owner != NULL) { 905154937Sjhb ts->ts_owner = NULL; 906154937Sjhb LIST_REMOVE(ts, ts_link); 907154937Sjhb } 908170295Sjeff LIST_FOREACH(nts, &td->td_contested, ts_link) { 909170295Sjeff cp = turnstile_first_waiter(nts)->td_priority; 910122514Sjhb if (cp < pri) 911122514Sjhb pri = cp; 912122514Sjhb } 913122514Sjhb mtx_unlock_spin(&td_contested_lock); 914139453Sjhb sched_unlend_prio(td, pri); 915170295Sjeff thread_unlock(td); 916122514Sjhb /* 917122514Sjhb * Wake up all the pending threads. If a thread is not blocked 918122514Sjhb * on a lock, then it is currently executing on another CPU in 919123364Sjhb * turnstile_wait() or sitting on a run queue waiting to resume 920123364Sjhb * in turnstile_wait(). Set a flag to force it to try to acquire 921122514Sjhb * the lock again instead of blocking. 922122514Sjhb */ 923122514Sjhb while (!TAILQ_EMPTY(&pending_threads)) { 924122514Sjhb td = TAILQ_FIRST(&pending_threads); 925122514Sjhb TAILQ_REMOVE(&pending_threads, td, td_lockq); 926235459Srstone SDT_PROBE2(sched, , , wakeup, td, td->td_proc); 927170295Sjeff thread_lock(td); 928176078Sjeff THREAD_LOCKPTR_ASSERT(td, &ts->ts_lock); 929122514Sjhb MPASS(td->td_proc->p_magic == P_MAGIC); 930170295Sjeff MPASS(TD_ON_LOCK(td)); 931170295Sjeff TD_CLR_LOCK(td); 932170295Sjeff MPASS(TD_CAN_RUN(td)); 933170295Sjeff td->td_blocked = NULL; 934170295Sjeff td->td_lockname = NULL; 935201879Sattilio td->td_blktick = 0; 936154937Sjhb#ifdef INVARIANTS 937170295Sjeff td->td_tsqueue = 0xff; 938154937Sjhb#endif 939170295Sjeff sched_add(td, SRQ_BORING); 940170295Sjeff thread_unlock(td); 941122514Sjhb } 942170295Sjeff mtx_unlock_spin(&ts->ts_lock); 94367352Sjhb} 94467352Sjhb 94572200Sbmilekic/* 946157844Sjhb * Give up ownership of a turnstile. This must be called with the 947157844Sjhb * turnstile chain locked. 948157844Sjhb */ 949157844Sjhbvoid 950157844Sjhbturnstile_disown(struct turnstile *ts) 951157844Sjhb{ 952157844Sjhb struct thread *td; 953157844Sjhb u_char cp, pri; 954157844Sjhb 955157844Sjhb MPASS(ts != NULL); 956170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 957157844Sjhb MPASS(ts->ts_owner == curthread); 958157844Sjhb MPASS(TAILQ_EMPTY(&ts->ts_pending)); 959157844Sjhb MPASS(!TAILQ_EMPTY(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE]) || 960157844Sjhb !TAILQ_EMPTY(&ts->ts_blocked[TS_SHARED_QUEUE])); 961157844Sjhb 962157844Sjhb /* 963157844Sjhb * Remove the turnstile from this thread's list of contested locks 964157844Sjhb * since this thread doesn't own it anymore. New threads will 965157844Sjhb * not be blocking on the turnstile until it is claimed by a new 966157844Sjhb * owner. 967157844Sjhb */ 968157844Sjhb mtx_lock_spin(&td_contested_lock); 969157844Sjhb ts->ts_owner = NULL; 970157844Sjhb LIST_REMOVE(ts, ts_link); 971157844Sjhb mtx_unlock_spin(&td_contested_lock); 972157844Sjhb 973157844Sjhb /* 974157844Sjhb * Adjust the priority of curthread based on other contested 975157844Sjhb * locks it owns. Don't lower the priority below the base 976157844Sjhb * priority however. 977157844Sjhb */ 978157844Sjhb td = curthread; 979157844Sjhb pri = PRI_MAX; 980170295Sjeff thread_lock(td); 981170295Sjeff mtx_unlock_spin(&ts->ts_lock); 982157844Sjhb mtx_lock_spin(&td_contested_lock); 983157844Sjhb LIST_FOREACH(ts, &td->td_contested, ts_link) { 984157844Sjhb cp = turnstile_first_waiter(ts)->td_priority; 985157844Sjhb if (cp < pri) 986157844Sjhb pri = cp; 987157844Sjhb } 988157844Sjhb mtx_unlock_spin(&td_contested_lock); 989157844Sjhb sched_unlend_prio(td, pri); 990170295Sjeff thread_unlock(td); 991157844Sjhb} 992157844Sjhb 993157844Sjhb/* 994122514Sjhb * Return the first thread in a turnstile. 99572200Sbmilekic */ 996122514Sjhbstruct thread * 997154937Sjhbturnstile_head(struct turnstile *ts, int queue) 99867352Sjhb{ 999122514Sjhb#ifdef INVARIANTS 100067352Sjhb 1001122514Sjhb MPASS(ts != NULL); 1002154937Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 1003170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 1004122514Sjhb#endif 1005154937Sjhb return (TAILQ_FIRST(&ts->ts_blocked[queue])); 100671320Sjasone} 1007154937Sjhb 1008157844Sjhb/* 1009157844Sjhb * Returns true if a sub-queue of a turnstile is empty. 1010157844Sjhb */ 1011157844Sjhbint 1012157844Sjhbturnstile_empty(struct turnstile *ts, int queue) 1013157844Sjhb{ 1014157844Sjhb#ifdef INVARIANTS 1015157844Sjhb 1016157844Sjhb MPASS(ts != NULL); 1017157844Sjhb MPASS(queue == TS_SHARED_QUEUE || queue == TS_EXCLUSIVE_QUEUE); 1018170295Sjeff mtx_assert(&ts->ts_lock, MA_OWNED); 1019157844Sjhb#endif 1020157844Sjhb return (TAILQ_EMPTY(&ts->ts_blocked[queue])); 1021157844Sjhb} 1022157844Sjhb 1023154937Sjhb#ifdef DDB 1024154937Sjhbstatic void 1025154937Sjhbprint_thread(struct thread *td, const char *prefix) 1026154937Sjhb{ 1027154937Sjhb 1028154937Sjhb db_printf("%s%p (tid %d, pid %d, \"%s\")\n", prefix, td, td->td_tid, 1029295488Skib td->td_proc->p_pid, td->td_name); 1030154937Sjhb} 1031154937Sjhb 1032154937Sjhbstatic void 1033154937Sjhbprint_queue(struct threadqueue *queue, const char *header, const char *prefix) 1034154937Sjhb{ 1035154937Sjhb struct thread *td; 1036154937Sjhb 1037154937Sjhb db_printf("%s:\n", header); 1038154937Sjhb if (TAILQ_EMPTY(queue)) { 1039154937Sjhb db_printf("%sempty\n", prefix); 1040154937Sjhb return; 1041154937Sjhb } 1042154937Sjhb TAILQ_FOREACH(td, queue, td_lockq) { 1043154937Sjhb print_thread(td, prefix); 1044154937Sjhb } 1045154937Sjhb} 1046154937Sjhb 1047154937SjhbDB_SHOW_COMMAND(turnstile, db_show_turnstile) 1048154937Sjhb{ 1049154937Sjhb struct turnstile_chain *tc; 1050154937Sjhb struct turnstile *ts; 1051154937Sjhb struct lock_object *lock; 1052154937Sjhb int i; 1053154937Sjhb 1054154937Sjhb if (!have_addr) 1055154937Sjhb return; 1056154937Sjhb 1057154937Sjhb /* 1058154937Sjhb * First, see if there is an active turnstile for the lock indicated 1059154937Sjhb * by the address. 1060154937Sjhb */ 1061154937Sjhb lock = (struct lock_object *)addr; 1062154937Sjhb tc = TC_LOOKUP(lock); 1063154937Sjhb LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 1064154937Sjhb if (ts->ts_lockobj == lock) 1065154937Sjhb goto found; 1066154937Sjhb 1067154937Sjhb /* 1068154937Sjhb * Second, see if there is an active turnstile at the address 1069154937Sjhb * indicated. 1070154937Sjhb */ 1071154937Sjhb for (i = 0; i < TC_TABLESIZE; i++) 1072154937Sjhb LIST_FOREACH(ts, &turnstile_chains[i].tc_turnstiles, ts_hash) { 1073154937Sjhb if (ts == (struct turnstile *)addr) 1074154937Sjhb goto found; 1075154937Sjhb } 1076154937Sjhb 1077154937Sjhb db_printf("Unable to locate a turnstile via %p\n", (void *)addr); 1078154937Sjhb return; 1079154937Sjhbfound: 1080154937Sjhb lock = ts->ts_lockobj; 1081154937Sjhb db_printf("Lock: %p - (%s) %s\n", lock, LOCK_CLASS(lock)->lc_name, 1082154937Sjhb lock->lo_name); 1083154937Sjhb if (ts->ts_owner) 1084154937Sjhb print_thread(ts->ts_owner, "Lock Owner: "); 1085154937Sjhb else 1086154937Sjhb db_printf("Lock Owner: none\n"); 1087154937Sjhb print_queue(&ts->ts_blocked[TS_SHARED_QUEUE], "Shared Waiters", "\t"); 1088154937Sjhb print_queue(&ts->ts_blocked[TS_EXCLUSIVE_QUEUE], "Exclusive Waiters", 1089154937Sjhb "\t"); 1090154937Sjhb print_queue(&ts->ts_pending, "Pending Threads", "\t"); 1091154937Sjhb 1092154937Sjhb} 1093158031Sjhb 1094161324Sjhb/* 1095161324Sjhb * Show all the threads a particular thread is waiting on based on 1096161324Sjhb * non-sleepable and non-spin locks. 1097161324Sjhb */ 1098158031Sjhbstatic void 1099161324Sjhbprint_lockchain(struct thread *td, const char *prefix) 1100158031Sjhb{ 1101158031Sjhb struct lock_object *lock; 1102158031Sjhb struct lock_class *class; 1103158031Sjhb struct turnstile *ts; 1104158031Sjhb 1105158031Sjhb /* 1106158031Sjhb * Follow the chain. We keep walking as long as the thread is 1107158031Sjhb * blocked on a turnstile that has an owner. 1108158031Sjhb */ 1109160313Sjhb while (!db_pager_quit) { 1110158031Sjhb db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid, 1111295488Skib td->td_proc->p_pid, td->td_name); 1112158031Sjhb switch (td->td_state) { 1113158031Sjhb case TDS_INACTIVE: 1114158031Sjhb db_printf("is inactive\n"); 1115158031Sjhb return; 1116158031Sjhb case TDS_CAN_RUN: 1117158031Sjhb db_printf("can run\n"); 1118158031Sjhb return; 1119158031Sjhb case TDS_RUNQ: 1120158031Sjhb db_printf("is on a run queue\n"); 1121158031Sjhb return; 1122158031Sjhb case TDS_RUNNING: 1123158031Sjhb db_printf("running on CPU %d\n", td->td_oncpu); 1124158031Sjhb return; 1125158031Sjhb case TDS_INHIBITED: 1126158031Sjhb if (TD_ON_LOCK(td)) { 1127158031Sjhb ts = td->td_blocked; 1128158031Sjhb lock = ts->ts_lockobj; 1129158031Sjhb class = LOCK_CLASS(lock); 1130158031Sjhb db_printf("blocked on lock %p (%s) \"%s\"\n", 1131158031Sjhb lock, class->lc_name, lock->lo_name); 1132158031Sjhb if (ts->ts_owner == NULL) 1133158031Sjhb return; 1134158031Sjhb td = ts->ts_owner; 1135158031Sjhb break; 1136158031Sjhb } 1137158031Sjhb db_printf("inhibited\n"); 1138158031Sjhb return; 1139158031Sjhb default: 1140158031Sjhb db_printf("??? (%#x)\n", td->td_state); 1141158031Sjhb return; 1142158031Sjhb } 1143158031Sjhb } 1144158031Sjhb} 1145158031Sjhb 1146161324SjhbDB_SHOW_COMMAND(lockchain, db_show_lockchain) 1147158031Sjhb{ 1148158031Sjhb struct thread *td; 1149158031Sjhb 1150158031Sjhb /* Figure out which thread to start with. */ 1151158031Sjhb if (have_addr) 1152283248Spfg td = db_lookup_thread(addr, true); 1153158031Sjhb else 1154158031Sjhb td = kdb_thread; 1155158031Sjhb 1156161324Sjhb print_lockchain(td, ""); 1157158031Sjhb} 1158158031Sjhb 1159183054SsamDB_SHOW_ALL_COMMAND(chains, db_show_allchains) 1160158031Sjhb{ 1161158031Sjhb struct thread *td; 1162158031Sjhb struct proc *p; 1163158031Sjhb int i; 1164158031Sjhb 1165158031Sjhb i = 1; 1166166073Sdelphij FOREACH_PROC_IN_SYSTEM(p) { 1167158031Sjhb FOREACH_THREAD_IN_PROC(p, td) { 1168158031Sjhb if (TD_ON_LOCK(td) && LIST_EMPTY(&td->td_contested)) { 1169158031Sjhb db_printf("chain %d:\n", i++); 1170161324Sjhb print_lockchain(td, " "); 1171158031Sjhb } 1172160313Sjhb if (db_pager_quit) 1173160313Sjhb return; 1174158031Sjhb } 1175158031Sjhb } 1176158031Sjhb} 1177183054SsamDB_SHOW_ALIAS(allchains, db_show_allchains) 1178158031Sjhb 1179161337Sjhb/* 1180161337Sjhb * Show all the threads a particular thread is waiting on based on 1181161337Sjhb * sleepable locks. 1182161337Sjhb */ 1183161337Sjhbstatic void 1184161337Sjhbprint_sleepchain(struct thread *td, const char *prefix) 1185161337Sjhb{ 1186161337Sjhb struct thread *owner; 1187161337Sjhb 1188161337Sjhb /* 1189161337Sjhb * Follow the chain. We keep walking as long as the thread is 1190161337Sjhb * blocked on a sleep lock that has an owner. 1191161337Sjhb */ 1192161337Sjhb while (!db_pager_quit) { 1193161337Sjhb db_printf("%sthread %d (pid %d, %s) ", prefix, td->td_tid, 1194295488Skib td->td_proc->p_pid, td->td_name); 1195161337Sjhb switch (td->td_state) { 1196161337Sjhb case TDS_INACTIVE: 1197161337Sjhb db_printf("is inactive\n"); 1198161337Sjhb return; 1199161337Sjhb case TDS_CAN_RUN: 1200161337Sjhb db_printf("can run\n"); 1201161337Sjhb return; 1202161337Sjhb case TDS_RUNQ: 1203161337Sjhb db_printf("is on a run queue\n"); 1204161337Sjhb return; 1205161337Sjhb case TDS_RUNNING: 1206161337Sjhb db_printf("running on CPU %d\n", td->td_oncpu); 1207161337Sjhb return; 1208161337Sjhb case TDS_INHIBITED: 1209161337Sjhb if (TD_ON_SLEEPQ(td)) { 1210161337Sjhb if (lockmgr_chain(td, &owner) || 1211161337Sjhb sx_chain(td, &owner)) { 1212161337Sjhb if (owner == NULL) 1213161337Sjhb return; 1214161337Sjhb td = owner; 1215161337Sjhb break; 1216161337Sjhb } 1217161337Sjhb db_printf("sleeping on %p \"%s\"\n", 1218161337Sjhb td->td_wchan, td->td_wmesg); 1219161337Sjhb return; 1220161337Sjhb } 1221161337Sjhb db_printf("inhibited\n"); 1222161337Sjhb return; 1223161337Sjhb default: 1224161337Sjhb db_printf("??? (%#x)\n", td->td_state); 1225161337Sjhb return; 1226161337Sjhb } 1227161337Sjhb } 1228161337Sjhb} 1229161337Sjhb 1230161337SjhbDB_SHOW_COMMAND(sleepchain, db_show_sleepchain) 1231161337Sjhb{ 1232161337Sjhb struct thread *td; 1233161337Sjhb 1234161337Sjhb /* Figure out which thread to start with. */ 1235161337Sjhb if (have_addr) 1236283248Spfg td = db_lookup_thread(addr, true); 1237161337Sjhb else 1238161337Sjhb td = kdb_thread; 1239161337Sjhb 1240161337Sjhb print_sleepchain(td, ""); 1241161337Sjhb} 1242161337Sjhb 1243158031Sjhbstatic void print_waiters(struct turnstile *ts, int indent); 1244158031Sjhb 1245158031Sjhbstatic void 1246158031Sjhbprint_waiter(struct thread *td, int indent) 1247158031Sjhb{ 1248158031Sjhb struct turnstile *ts; 1249158031Sjhb int i; 1250158031Sjhb 1251160313Sjhb if (db_pager_quit) 1252160313Sjhb return; 1253158031Sjhb for (i = 0; i < indent; i++) 1254158031Sjhb db_printf(" "); 1255158031Sjhb print_thread(td, "thread "); 1256158031Sjhb LIST_FOREACH(ts, &td->td_contested, ts_link) 1257158031Sjhb print_waiters(ts, indent + 1); 1258158031Sjhb} 1259158031Sjhb 1260158031Sjhbstatic void 1261158031Sjhbprint_waiters(struct turnstile *ts, int indent) 1262158031Sjhb{ 1263158031Sjhb struct lock_object *lock; 1264158031Sjhb struct lock_class *class; 1265158031Sjhb struct thread *td; 1266158031Sjhb int i; 1267158031Sjhb 1268160313Sjhb if (db_pager_quit) 1269160313Sjhb return; 1270158031Sjhb lock = ts->ts_lockobj; 1271158031Sjhb class = LOCK_CLASS(lock); 1272158031Sjhb for (i = 0; i < indent; i++) 1273158031Sjhb db_printf(" "); 1274158031Sjhb db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, lock->lo_name); 1275158031Sjhb TAILQ_FOREACH(td, &ts->ts_blocked[TS_EXCLUSIVE_QUEUE], td_lockq) 1276158031Sjhb print_waiter(td, indent + 1); 1277158031Sjhb TAILQ_FOREACH(td, &ts->ts_blocked[TS_SHARED_QUEUE], td_lockq) 1278158031Sjhb print_waiter(td, indent + 1); 1279158031Sjhb TAILQ_FOREACH(td, &ts->ts_pending, td_lockq) 1280158031Sjhb print_waiter(td, indent + 1); 1281158031Sjhb} 1282158031Sjhb 1283161324SjhbDB_SHOW_COMMAND(locktree, db_show_locktree) 1284158031Sjhb{ 1285158031Sjhb struct lock_object *lock; 1286158031Sjhb struct lock_class *class; 1287158031Sjhb struct turnstile_chain *tc; 1288158031Sjhb struct turnstile *ts; 1289158031Sjhb 1290158031Sjhb if (!have_addr) 1291158031Sjhb return; 1292158031Sjhb lock = (struct lock_object *)addr; 1293158031Sjhb tc = TC_LOOKUP(lock); 1294158031Sjhb LIST_FOREACH(ts, &tc->tc_turnstiles, ts_hash) 1295158031Sjhb if (ts->ts_lockobj == lock) 1296158031Sjhb break; 1297158031Sjhb if (ts == NULL) { 1298158031Sjhb class = LOCK_CLASS(lock); 1299158031Sjhb db_printf("lock %p (%s) \"%s\"\n", lock, class->lc_name, 1300158031Sjhb lock->lo_name); 1301158031Sjhb } else 1302158031Sjhb print_waiters(ts, 0); 1303158031Sjhb} 1304154937Sjhb#endif 1305