1 /* $NetBSD: subr_pserialize.c,v 1.10 2017/12/28 03:39:48 msaitoh 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 1.10 2017/12/28 03:39:48 msaitoh 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/proc.h> 52 #include <sys/queue.h> 53 #include <sys/xcall.h> 54 55 struct pserialize { 56 TAILQ_ENTRY(pserialize) psz_chain; 57 lwp_t * psz_owner; 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 #ifdef LOCKDEBUG 77 #include <sys/percpu.h> 78 79 static percpu_t *psz_debug_nreads __cacheline_aligned; 80 #endif 81 82 /* 83 * pserialize_init: 84 * 85 * Initialize passive serialization structures. 86 */ 87 void 88 pserialize_init(void) 89 { 90 91 psz_work_todo = 0; 92 TAILQ_INIT(&psz_queue0); 93 TAILQ_INIT(&psz_queue1); 94 TAILQ_INIT(&psz_queue2); 95 mutex_init(&psz_lock, MUTEX_DEFAULT, IPL_SCHED); 96 evcnt_attach_dynamic(&psz_ev_excl, EVCNT_TYPE_MISC, NULL, 97 "pserialize", "exclusive access"); 98 #ifdef LOCKDEBUG 99 psz_debug_nreads = percpu_alloc(sizeof(uint32_t)); 100 #endif 101 } 102 103 /* 104 * pserialize_create: 105 * 106 * Create and initialize a passive serialization object. 107 */ 108 pserialize_t 109 pserialize_create(void) 110 { 111 pserialize_t psz; 112 113 psz = kmem_zalloc(sizeof(struct pserialize), KM_SLEEP); 114 kcpuset_create(&psz->psz_target, true); 115 kcpuset_create(&psz->psz_pass, true); 116 psz->psz_owner = NULL; 117 118 return psz; 119 } 120 121 /* 122 * pserialize_destroy: 123 * 124 * Destroy a passive serialization object. 125 */ 126 void 127 pserialize_destroy(pserialize_t psz) 128 { 129 130 KASSERT(psz->psz_owner == NULL); 131 132 kcpuset_destroy(psz->psz_target); 133 kcpuset_destroy(psz->psz_pass); 134 kmem_free(psz, sizeof(struct pserialize)); 135 } 136 137 /* 138 * pserialize_perform: 139 * 140 * Perform the write side of passive serialization. The calling 141 * thread holds an exclusive lock on the data object(s) being updated. 142 * We wait until every processor in the system has made at least two 143 * passes through cpu_switchto(). The wait is made with the caller's 144 * update lock held, but is short term. 145 */ 146 void 147 pserialize_perform(pserialize_t psz) 148 { 149 uint64_t xc; 150 151 KASSERT(!cpu_intr_p()); 152 KASSERT(!cpu_softintr_p()); 153 154 if (__predict_false(panicstr != NULL)) { 155 return; 156 } 157 KASSERT(psz->psz_owner == NULL); 158 KASSERT(ncpu > 0); 159 160 if (__predict_false(mp_online == false)) { 161 psz_ev_excl.ev_count++; 162 return; 163 } 164 165 /* 166 * Set up the object and put it onto the queue. The lock 167 * activity here provides the necessary memory barrier to 168 * make the caller's data update completely visible to 169 * other processors. 170 */ 171 psz->psz_owner = curlwp; 172 kcpuset_copy(psz->psz_target, kcpuset_running); 173 kcpuset_zero(psz->psz_pass); 174 175 mutex_spin_enter(&psz_lock); 176 TAILQ_INSERT_TAIL(&psz_queue0, psz, psz_chain); 177 psz_work_todo++; 178 179 do { 180 mutex_spin_exit(&psz_lock); 181 182 /* 183 * Force some context switch activity on every CPU, as 184 * the system may not be busy. Pause to not flood. 185 */ 186 xc = xc_broadcast(XC_HIGHPRI, (xcfunc_t)nullop, NULL, NULL); 187 xc_wait(xc); 188 kpause("psrlz", false, 1, NULL); 189 190 mutex_spin_enter(&psz_lock); 191 } while (!kcpuset_iszero(psz->psz_target)); 192 193 psz_ev_excl.ev_count++; 194 mutex_spin_exit(&psz_lock); 195 196 psz->psz_owner = NULL; 197 } 198 199 int 200 pserialize_read_enter(void) 201 { 202 int s; 203 204 KASSERT(!cpu_intr_p()); 205 s = splsoftserial(); 206 #ifdef LOCKDEBUG 207 { 208 uint32_t *nreads; 209 nreads = percpu_getref(psz_debug_nreads); 210 (*nreads)++; 211 if (*nreads == 0) 212 panic("nreads overflow"); 213 percpu_putref(psz_debug_nreads); 214 } 215 #endif 216 return s; 217 } 218 219 void 220 pserialize_read_exit(int s) 221 { 222 223 #ifdef LOCKDEBUG 224 { 225 uint32_t *nreads; 226 nreads = percpu_getref(psz_debug_nreads); 227 (*nreads)--; 228 if (*nreads == UINT_MAX) 229 panic("nreads underflow"); 230 percpu_putref(psz_debug_nreads); 231 } 232 #endif 233 splx(s); 234 } 235 236 /* 237 * pserialize_switchpoint: 238 * 239 * Monitor system context switch activity. Called from machine 240 * independent code after mi_switch() returns. 241 */ 242 void 243 pserialize_switchpoint(void) 244 { 245 pserialize_t psz, next; 246 cpuid_t cid; 247 248 /* We must to ensure not to come here from inside a read section. */ 249 KASSERT(pserialize_not_in_read_section()); 250 251 /* 252 * If no updates pending, bail out. No need to lock in order to 253 * test psz_work_todo; the only ill effect of missing an update 254 * would be to delay LWPs waiting in pserialize_perform(). That 255 * will not happen because updates are on the queue before an 256 * xcall is generated (serialization) to tickle every CPU. 257 */ 258 if (__predict_true(psz_work_todo == 0)) { 259 return; 260 } 261 mutex_spin_enter(&psz_lock); 262 cid = cpu_index(curcpu()); 263 264 /* 265 * At first, scan through the second queue and update each request, 266 * if passed all processors, then transfer to the third queue. 267 */ 268 for (psz = TAILQ_FIRST(&psz_queue1); psz != NULL; psz = next) { 269 next = TAILQ_NEXT(psz, psz_chain); 270 kcpuset_set(psz->psz_pass, cid); 271 if (!kcpuset_match(psz->psz_pass, psz->psz_target)) { 272 continue; 273 } 274 kcpuset_zero(psz->psz_pass); 275 TAILQ_REMOVE(&psz_queue1, psz, psz_chain); 276 TAILQ_INSERT_TAIL(&psz_queue2, psz, psz_chain); 277 } 278 /* 279 * Scan through the first queue and update each request, 280 * if passed all processors, then move to the second queue. 281 */ 282 for (psz = TAILQ_FIRST(&psz_queue0); psz != NULL; psz = next) { 283 next = TAILQ_NEXT(psz, psz_chain); 284 kcpuset_set(psz->psz_pass, cid); 285 if (!kcpuset_match(psz->psz_pass, psz->psz_target)) { 286 continue; 287 } 288 kcpuset_zero(psz->psz_pass); 289 TAILQ_REMOVE(&psz_queue0, psz, psz_chain); 290 TAILQ_INSERT_TAIL(&psz_queue1, psz, psz_chain); 291 } 292 /* 293 * Process the third queue: entries have been seen twice on every 294 * processor, remove from the queue and notify the updating thread. 295 */ 296 while ((psz = TAILQ_FIRST(&psz_queue2)) != NULL) { 297 TAILQ_REMOVE(&psz_queue2, psz, psz_chain); 298 kcpuset_zero(psz->psz_target); 299 psz_work_todo--; 300 } 301 mutex_spin_exit(&psz_lock); 302 } 303 304 /* 305 * pserialize_in_read_section: 306 * 307 * True if the caller is in a pserialize read section. To be used only 308 * for diagnostic assertions where we want to guarantee the condition like: 309 * 310 * KASSERT(pserialize_in_read_section()); 311 */ 312 bool 313 pserialize_in_read_section(void) 314 { 315 #ifdef LOCKDEBUG 316 uint32_t *nreads; 317 bool in; 318 319 /* Not initialized yet */ 320 if (__predict_false(psz_debug_nreads == NULL)) 321 return true; 322 323 nreads = percpu_getref(psz_debug_nreads); 324 in = *nreads != 0; 325 percpu_putref(psz_debug_nreads); 326 327 return in; 328 #else 329 return true; 330 #endif 331 } 332 333 /* 334 * pserialize_not_in_read_section: 335 * 336 * True if the caller is not in a pserialize read section. To be used only 337 * for diagnostic assertions where we want to guarantee the condition like: 338 * 339 * KASSERT(pserialize_not_in_read_section()); 340 */ 341 bool 342 pserialize_not_in_read_section(void) 343 { 344 #ifdef LOCKDEBUG 345 uint32_t *nreads; 346 bool notin; 347 348 /* Not initialized yet */ 349 if (__predict_false(psz_debug_nreads == NULL)) 350 return true; 351 352 nreads = percpu_getref(psz_debug_nreads); 353 notin = *nreads == 0; 354 percpu_putref(psz_debug_nreads); 355 356 return notin; 357 #else 358 return true; 359 #endif 360 } 361