1 /* $NetBSD: subr_pserialize.c,v 1.6 2013/01/07 23:21:32 rmind Exp $ */ 2 3 /*- 4 * Copyright (c) 2010, 2011 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 17 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 18 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 19 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 20 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 21 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 22 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 23 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 24 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 25 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 26 * POSSIBILITY OF SUCH DAMAGE. 27 */ 28 29 /* 30 * Passive serialization. 31 * 32 * Implementation accurately matches the lapsed US patent 4809168, therefore 33 * code is patent-free in the United States. Your use of this code is at 34 * your own risk. 35 * 36 * Note for NetBSD developers: all changes to this source file must be 37 * approved by the <core>. 38 */ 39 40 #include <sys/cdefs.h> 41 __KERNEL_RCSID(0, "$NetBSD: subr_pserialize.c,v 1.6 2013/01/07 23:21:32 rmind Exp $"); 42 43 #include <sys/param.h> 44 45 #include <sys/condvar.h> 46 #include <sys/cpu.h> 47 #include <sys/evcnt.h> 48 #include <sys/kmem.h> 49 #include <sys/mutex.h> 50 #include <sys/pserialize.h> 51 #include <sys/queue.h> 52 #include <sys/xcall.h> 53 54 struct pserialize { 55 TAILQ_ENTRY(pserialize) psz_chain; 56 lwp_t * psz_owner; 57 kcondvar_t psz_notifier; 58 kcpuset_t * psz_target; 59 kcpuset_t * psz_pass; 60 }; 61 62 static u_int psz_work_todo __cacheline_aligned; 63 static kmutex_t psz_lock __cacheline_aligned; 64 static struct evcnt psz_ev_excl __cacheline_aligned; 65 66 /* 67 * As defined in "Method 1": 68 * q0: "0 MP checkpoints have occured". 69 * q1: "1 MP checkpoint has occured". 70 * q2: "2 MP checkpoints have occured". 71 */ 72 static TAILQ_HEAD(, pserialize) psz_queue0 __cacheline_aligned; 73 static TAILQ_HEAD(, pserialize) psz_queue1 __cacheline_aligned; 74 static TAILQ_HEAD(, pserialize) psz_queue2 __cacheline_aligned; 75 76 /* 77 * pserialize_init: 78 * 79 * Initialize passive serialization structures. 80 */ 81 void 82 pserialize_init(void) 83 { 84 85 psz_work_todo = 0; 86 TAILQ_INIT(&psz_queue0); 87 TAILQ_INIT(&psz_queue1); 88 TAILQ_INIT(&psz_queue2); 89 mutex_init(&psz_lock, MUTEX_DEFAULT, IPL_SCHED); 90 evcnt_attach_dynamic(&psz_ev_excl, EVCNT_TYPE_MISC, NULL, 91 "pserialize", "exclusive access"); 92 } 93 94 /* 95 * pserialize_create: 96 * 97 * Create and initialize a passive serialization object. 98 */ 99 pserialize_t 100 pserialize_create(void) 101 { 102 pserialize_t psz; 103 104 psz = kmem_zalloc(sizeof(struct pserialize), KM_SLEEP); 105 cv_init(&psz->psz_notifier, "psrlz"); 106 kcpuset_create(&psz->psz_target, true); 107 kcpuset_create(&psz->psz_pass, true); 108 psz->psz_owner = NULL; 109 110 return psz; 111 } 112 113 /* 114 * pserialize_destroy: 115 * 116 * Destroy a passive serialization object. 117 */ 118 void 119 pserialize_destroy(pserialize_t psz) 120 { 121 122 KASSERT(psz->psz_owner == NULL); 123 124 cv_destroy(&psz->psz_notifier); 125 kcpuset_destroy(psz->psz_target); 126 kcpuset_destroy(psz->psz_pass); 127 kmem_free(psz, sizeof(struct pserialize)); 128 } 129 130 /* 131 * pserialize_perform: 132 * 133 * Perform the write side of passive serialization. The calling 134 * thread holds an exclusive lock on the data object(s) being updated. 135 * We wait until every processor in the system has made at least two 136 * passes through cpu_swichto(). The wait is made with the caller's 137 * update lock held, but is short term. 138 */ 139 void 140 pserialize_perform(pserialize_t psz) 141 { 142 uint64_t xc; 143 144 KASSERT(!cpu_intr_p()); 145 KASSERT(!cpu_softintr_p()); 146 147 if (__predict_false(panicstr != NULL)) { 148 return; 149 } 150 KASSERT(psz->psz_owner == NULL); 151 KASSERT(ncpu > 0); 152 153 /* 154 * Set up the object and put it onto the queue. The lock 155 * activity here provides the necessary memory barrier to 156 * make the caller's data update completely visible to 157 * other processors. 158 */ 159 psz->psz_owner = curlwp; 160 kcpuset_copy(psz->psz_target, kcpuset_running); 161 kcpuset_zero(psz->psz_pass); 162 163 mutex_spin_enter(&psz_lock); 164 TAILQ_INSERT_TAIL(&psz_queue0, psz, psz_chain); 165 psz_work_todo++; 166 mutex_spin_exit(&psz_lock); 167 168 /* 169 * Force some context switch activity on every CPU, as the system 170 * may not be busy. Note: should pass the point twice. 171 */ 172 xc = xc_broadcast(XC_HIGHPRI, (xcfunc_t)nullop, NULL, NULL); 173 xc_wait(xc); 174 175 /* No need to xc_wait() as we implement our own condvar. */ 176 xc_broadcast(XC_HIGHPRI, (xcfunc_t)nullop, NULL, NULL); 177 178 /* 179 * Wait for all CPUs to cycle through mi_switch() twice. 180 * The last one through will remove our update from the 181 * queue and awaken us. 182 */ 183 mutex_spin_enter(&psz_lock); 184 while (!kcpuset_iszero(psz->psz_target)) { 185 cv_wait(&psz->psz_notifier, &psz_lock); 186 } 187 psz_ev_excl.ev_count++; 188 mutex_spin_exit(&psz_lock); 189 190 psz->psz_owner = NULL; 191 } 192 193 int 194 pserialize_read_enter(void) 195 { 196 197 KASSERT(!cpu_intr_p()); 198 return splsoftserial(); 199 } 200 201 void 202 pserialize_read_exit(int s) 203 { 204 205 splx(s); 206 } 207 208 /* 209 * pserialize_switchpoint: 210 * 211 * Monitor system context switch activity. Called from machine 212 * independent code after mi_switch() returns. 213 */ 214 void 215 pserialize_switchpoint(void) 216 { 217 pserialize_t psz, next; 218 cpuid_t cid; 219 220 /* 221 * If no updates pending, bail out. No need to lock in order to 222 * test psz_work_todo; the only ill effect of missing an update 223 * would be to delay LWPs waiting in pserialize_perform(). That 224 * will not happen because updates are on the queue before an 225 * xcall is generated (serialization) to tickle every CPU. 226 */ 227 if (__predict_true(psz_work_todo == 0)) { 228 return; 229 } 230 mutex_spin_enter(&psz_lock); 231 cid = cpu_index(curcpu()); 232 233 /* 234 * At first, scan through the second queue and update each request, 235 * if passed all processors, then transfer to the third queue. 236 */ 237 for (psz = TAILQ_FIRST(&psz_queue1); psz != NULL; psz = next) { 238 next = TAILQ_NEXT(psz, psz_chain); 239 if (!kcpuset_match(psz->psz_pass, psz->psz_target)) { 240 kcpuset_set(psz->psz_pass, cid); 241 continue; 242 } 243 kcpuset_zero(psz->psz_pass); 244 TAILQ_REMOVE(&psz_queue1, psz, psz_chain); 245 TAILQ_INSERT_TAIL(&psz_queue2, psz, psz_chain); 246 } 247 /* 248 * Scan through the first queue and update each request, 249 * if passed all processors, then move to the second queue. 250 */ 251 for (psz = TAILQ_FIRST(&psz_queue0); psz != NULL; psz = next) { 252 next = TAILQ_NEXT(psz, psz_chain); 253 if (!kcpuset_match(psz->psz_pass, psz->psz_target)) { 254 kcpuset_set(psz->psz_pass, cid); 255 continue; 256 } 257 kcpuset_zero(psz->psz_pass); 258 TAILQ_REMOVE(&psz_queue0, psz, psz_chain); 259 TAILQ_INSERT_TAIL(&psz_queue1, psz, psz_chain); 260 } 261 /* 262 * Process the third queue: entries have been seen twice on every 263 * processor, remove from the queue and notify the updating thread. 264 */ 265 while ((psz = TAILQ_FIRST(&psz_queue2)) != NULL) { 266 TAILQ_REMOVE(&psz_queue2, psz, psz_chain); 267 kcpuset_zero(psz->psz_target); 268 cv_signal(&psz->psz_notifier); 269 psz_work_todo--; 270 } 271 mutex_spin_exit(&psz_lock); 272 } 273