| 1 | /* |
| 2 | * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com> |
| 3 | * All rights reserved. |
| 4 | * |
| 5 | * Redistribution and use in source and binary forms, with or without |
| 6 | * modification, are permitted provided that the following conditions |
| 7 | * are met: |
| 8 | * 1. Redistributions of source code must retain the above copyright |
| 9 | * notice, this list of conditions and the following disclaimer. |
| 10 | * 2. Redistributions in binary form must reproduce the above copyright |
| 11 | * notice, this list of conditions and the following disclaimer in the |
| 12 | * documentation and/or other materials provided with the distribution. |
| 13 | * |
| 14 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND |
| 15 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 16 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 17 | * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE |
| 18 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 19 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 20 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 21 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 22 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 23 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 24 | * SUCH DAMAGE. |
| 25 | * |
| 26 | * $DragonFly: src/sys/kern/lwkt_ipiq.c,v 1.2 2004/02/15 05:15:25 dillon Exp $ |
| 27 | */ |
| 28 | |
| 29 | /* |
| 30 | * This module implements IPI message queueing and the MI portion of IPI |
| 31 | * message processing. |
| 32 | */ |
| 33 | |
| 34 | #ifdef _KERNEL |
| 35 | |
| 36 | #include <sys/param.h> |
| 37 | #include <sys/systm.h> |
| 38 | #include <sys/kernel.h> |
| 39 | #include <sys/proc.h> |
| 40 | #include <sys/rtprio.h> |
| 41 | #include <sys/queue.h> |
| 42 | #include <sys/thread2.h> |
| 43 | #include <sys/sysctl.h> |
| 44 | #include <sys/kthread.h> |
| 45 | #include <machine/cpu.h> |
| 46 | #include <sys/lock.h> |
| 47 | #include <sys/caps.h> |
| 48 | |
| 49 | #include <vm/vm.h> |
| 50 | #include <vm/vm_param.h> |
| 51 | #include <vm/vm_kern.h> |
| 52 | #include <vm/vm_object.h> |
| 53 | #include <vm/vm_page.h> |
| 54 | #include <vm/vm_map.h> |
| 55 | #include <vm/vm_pager.h> |
| 56 | #include <vm/vm_extern.h> |
| 57 | #include <vm/vm_zone.h> |
| 58 | |
| 59 | #include <machine/stdarg.h> |
| 60 | #include <machine/ipl.h> |
| 61 | #include <machine/smp.h> |
| 62 | #include <machine/atomic.h> |
| 63 | |
| 64 | #define THREAD_STACK (UPAGES * PAGE_SIZE) |
| 65 | |
| 66 | #else |
| 67 | |
| 68 | #include <sys/stdint.h> |
| 69 | #include <libcaps/thread.h> |
| 70 | #include <sys/thread.h> |
| 71 | #include <sys/msgport.h> |
| 72 | #include <sys/errno.h> |
| 73 | #include <libcaps/globaldata.h> |
| 74 | #include <sys/thread2.h> |
| 75 | #include <sys/msgport2.h> |
| 76 | #include <stdio.h> |
| 77 | #include <stdlib.h> |
| 78 | #include <string.h> |
| 79 | #include <machine/cpufunc.h> |
| 80 | #include <machine/lock.h> |
| 81 | |
| 82 | #endif |
| 83 | |
| 84 | #ifdef SMP |
| 85 | static __int64_t ipiq_count = 0; |
| 86 | static __int64_t ipiq_fifofull = 0; |
| 87 | #endif |
| 88 | |
| 89 | #ifdef _KERNEL |
| 90 | |
| 91 | #ifdef SMP |
| 92 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, ""); |
| 93 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, ""); |
| 94 | #endif |
| 95 | |
| 96 | #endif |
| 97 | |
| 98 | #ifdef SMP |
| 99 | |
| 100 | static int lwkt_process_ipiq1(lwkt_ipiq_t ip, struct intrframe *frame); |
| 101 | static void lwkt_cpusync_remote1(lwkt_cpusync_t poll); |
| 102 | static void lwkt_cpusync_remote2(lwkt_cpusync_t poll); |
| 103 | |
| 104 | /* |
| 105 | * Send a function execution request to another cpu. The request is queued |
| 106 | * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every |
| 107 | * possible target cpu. The FIFO can be written. |
| 108 | * |
| 109 | * YYY If the FIFO fills up we have to enable interrupts and process the |
| 110 | * IPIQ while waiting for it to empty or we may deadlock with another cpu. |
| 111 | * Create a CPU_*() function to do this! |
| 112 | * |
| 113 | * We can safely bump gd_intr_nesting_level because our crit_exit() at the |
| 114 | * end will take care of any pending interrupts. |
| 115 | * |
| 116 | * Must be called from a critical section. |
| 117 | */ |
| 118 | int |
| 119 | lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg) |
| 120 | { |
| 121 | lwkt_ipiq_t ip; |
| 122 | int windex; |
| 123 | struct globaldata *gd = mycpu; |
| 124 | |
| 125 | if (target == gd) { |
| 126 | func(arg); |
| 127 | return(0); |
| 128 | } |
| 129 | crit_enter(); |
| 130 | ++gd->gd_intr_nesting_level; |
| 131 | #ifdef INVARIANTS |
| 132 | if (gd->gd_intr_nesting_level > 20) |
| 133 | panic("lwkt_send_ipiq: TOO HEAVILY NESTED!"); |
| 134 | #endif |
| 135 | KKASSERT(curthread->td_pri >= TDPRI_CRIT); |
| 136 | ++ipiq_count; |
| 137 | ip = &gd->gd_ipiq[target->gd_cpuid]; |
| 138 | |
| 139 | /* |
| 140 | * We always drain before the FIFO becomes full so it should never |
| 141 | * become full. We need to leave enough entries to deal with |
| 142 | * reentrancy. |
| 143 | */ |
| 144 | KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO); |
| 145 | windex = ip->ip_windex & MAXCPUFIFO_MASK; |
| 146 | ip->ip_func[windex] = (ipifunc2_t)func; |
| 147 | ip->ip_arg[windex] = arg; |
| 148 | /* YYY memory barrier */ |
| 149 | ++ip->ip_windex; |
| 150 | if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) { |
| 151 | unsigned int eflags = read_eflags(); |
| 152 | cpu_enable_intr(); |
| 153 | ++ipiq_fifofull; |
| 154 | while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) { |
| 155 | KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1); |
| 156 | lwkt_process_ipiq(); |
| 157 | } |
| 158 | write_eflags(eflags); |
| 159 | } |
| 160 | --gd->gd_intr_nesting_level; |
| 161 | cpu_send_ipiq(target->gd_cpuid); /* issues mem barrier if appropriate */ |
| 162 | crit_exit(); |
| 163 | return(ip->ip_windex); |
| 164 | } |
| 165 | |
| 166 | /* |
| 167 | * deprecated, used only by fast int forwarding. |
| 168 | */ |
| 169 | int |
| 170 | lwkt_send_ipiq_bycpu(int dcpu, ipifunc_t func, void *arg) |
| 171 | { |
| 172 | return(lwkt_send_ipiq(globaldata_find(dcpu), func, arg)); |
| 173 | } |
| 174 | |
| 175 | /* |
| 176 | * Send a message to several target cpus. Typically used for scheduling. |
| 177 | * The message will not be sent to stopped cpus. |
| 178 | */ |
| 179 | int |
| 180 | lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg) |
| 181 | { |
| 182 | int cpuid; |
| 183 | int count = 0; |
| 184 | |
| 185 | mask &= ~stopped_cpus; |
| 186 | while (mask) { |
| 187 | cpuid = bsfl(mask); |
| 188 | lwkt_send_ipiq(globaldata_find(cpuid), func, arg); |
| 189 | mask &= ~(1 << cpuid); |
| 190 | ++count; |
| 191 | } |
| 192 | return(count); |
| 193 | } |
| 194 | |
| 195 | /* |
| 196 | * Wait for the remote cpu to finish processing a function. |
| 197 | * |
| 198 | * YYY we have to enable interrupts and process the IPIQ while waiting |
| 199 | * for it to empty or we may deadlock with another cpu. Create a CPU_*() |
| 200 | * function to do this! YYY we really should 'block' here. |
| 201 | * |
| 202 | * MUST be called from a critical section. This routine may be called |
| 203 | * from an interrupt (for example, if an interrupt wakes a foreign thread |
| 204 | * up). |
| 205 | */ |
| 206 | void |
| 207 | lwkt_wait_ipiq(globaldata_t target, int seq) |
| 208 | { |
| 209 | lwkt_ipiq_t ip; |
| 210 | int maxc = 100000000; |
| 211 | |
| 212 | if (target != mycpu) { |
| 213 | ip = &mycpu->gd_ipiq[target->gd_cpuid]; |
| 214 | if ((int)(ip->ip_xindex - seq) < 0) { |
| 215 | unsigned int eflags = read_eflags(); |
| 216 | cpu_enable_intr(); |
| 217 | while ((int)(ip->ip_xindex - seq) < 0) { |
| 218 | lwkt_process_ipiq(); |
| 219 | if (--maxc == 0) |
| 220 | printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq); |
| 221 | if (maxc < -1000000) |
| 222 | panic("LWKT_WAIT_IPIQ"); |
| 223 | } |
| 224 | write_eflags(eflags); |
| 225 | } |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | /* |
| 230 | * Called from IPI interrupt (like a fast interrupt), which has placed |
| 231 | * us in a critical section. The MP lock may or may not be held. |
| 232 | * May also be called from doreti or splz, or be reentrantly called |
| 233 | * indirectly through the ip_func[] we run. |
| 234 | * |
| 235 | * There are two versions, one where no interrupt frame is available (when |
| 236 | * called from the send code and from splz, and one where an interrupt |
| 237 | * frame is available. |
| 238 | */ |
| 239 | void |
| 240 | lwkt_process_ipiq(void) |
| 241 | { |
| 242 | globaldata_t gd = mycpu; |
| 243 | lwkt_ipiq_t ip; |
| 244 | int n; |
| 245 | |
| 246 | again: |
| 247 | for (n = 0; n < ncpus; ++n) { |
| 248 | if (n != gd->gd_cpuid) { |
| 249 | ip = globaldata_find(n)->gd_ipiq; |
| 250 | if (ip != NULL) { |
| 251 | while (lwkt_process_ipiq1(&ip[gd->gd_cpuid], NULL)) |
| 252 | ; |
| 253 | } |
| 254 | } |
| 255 | } |
| 256 | if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) { |
| 257 | if (lwkt_process_ipiq1(&gd->gd_cpusyncq, NULL)) |
| 258 | goto again; |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | #ifdef _KERNEL |
| 263 | void |
| 264 | lwkt_process_ipiq_frame(struct intrframe frame) |
| 265 | { |
| 266 | globaldata_t gd = mycpu; |
| 267 | lwkt_ipiq_t ip; |
| 268 | int n; |
| 269 | |
| 270 | again: |
| 271 | for (n = 0; n < ncpus; ++n) { |
| 272 | if (n != gd->gd_cpuid) { |
| 273 | ip = globaldata_find(n)->gd_ipiq; |
| 274 | if (ip != NULL) { |
| 275 | while (lwkt_process_ipiq1(&ip[gd->gd_cpuid], &frame)) |
| 276 | ; |
| 277 | } |
| 278 | } |
| 279 | } |
| 280 | if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) { |
| 281 | if (lwkt_process_ipiq1(&gd->gd_cpusyncq, &frame)) |
| 282 | goto again; |
| 283 | } |
| 284 | } |
| 285 | #endif |
| 286 | |
| 287 | static int |
| 288 | lwkt_process_ipiq1(lwkt_ipiq_t ip, struct intrframe *frame) |
| 289 | { |
| 290 | int ri; |
| 291 | int wi = ip->ip_windex; |
| 292 | /* |
| 293 | * Note: xindex is only updated after we are sure the function has |
| 294 | * finished execution. Beware lwkt_process_ipiq() reentrancy! The |
| 295 | * function may send an IPI which may block/drain. |
| 296 | */ |
| 297 | while ((ri = ip->ip_rindex) != wi) { |
| 298 | ip->ip_rindex = ri + 1; |
| 299 | ri &= MAXCPUFIFO_MASK; |
| 300 | ip->ip_func[ri](ip->ip_arg[ri], frame); |
| 301 | /* YYY memory barrier */ |
| 302 | ip->ip_xindex = ip->ip_rindex; |
| 303 | } |
| 304 | return(wi != ip->ip_windex); |
| 305 | } |
| 306 | |
| 307 | /* |
| 308 | * CPU Synchronization Support |
| 309 | * |
| 310 | * lwkt_cpusync_simple() |
| 311 | * |
| 312 | * The function is executed synchronously before return on remote cpus. |
| 313 | * A lwkt_cpusync_t pointer is passed as an argument. The data can |
| 314 | * be accessed via arg->cs_data. |
| 315 | * |
| 316 | * XXX should I just pass the data as an argument to be consistent? |
| 317 | */ |
| 318 | |
| 319 | void |
| 320 | lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data) |
| 321 | { |
| 322 | struct lwkt_cpusync cmd; |
| 323 | |
| 324 | cmd.cs_run_func = NULL; |
| 325 | cmd.cs_fin1_func = func; |
| 326 | cmd.cs_fin2_func = NULL; |
| 327 | cmd.cs_data = data; |
| 328 | lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd); |
| 329 | if (mask & (1 << mycpu->gd_cpuid)) |
| 330 | func(&cmd); |
| 331 | lwkt_cpusync_finish(&cmd); |
| 332 | } |
| 333 | |
| 334 | /* |
| 335 | * lwkt_cpusync_fastdata() |
| 336 | * |
| 337 | * The function is executed in tandem with return on remote cpus. |
| 338 | * The data is directly passed as an argument. Do not pass pointers to |
| 339 | * temporary storage as the storage might have |
| 340 | * gone poof by the time the target cpu executes |
| 341 | * the function. |
| 342 | * |
| 343 | * At the moment lwkt_cpusync is declared on the stack and we must wait |
| 344 | * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future |
| 345 | * optimization we should be able to put a counter in the globaldata |
| 346 | * structure (if it is not otherwise being used) and just poke it and |
| 347 | * return without waiting. XXX |
| 348 | */ |
| 349 | void |
| 350 | lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data) |
| 351 | { |
| 352 | struct lwkt_cpusync cmd; |
| 353 | |
| 354 | cmd.cs_run_func = NULL; |
| 355 | cmd.cs_fin1_func = NULL; |
| 356 | cmd.cs_fin2_func = func; |
| 357 | cmd.cs_data = NULL; |
| 358 | lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd); |
| 359 | if (mask & (1 << mycpu->gd_cpuid)) |
| 360 | func(data); |
| 361 | lwkt_cpusync_finish(&cmd); |
| 362 | } |
| 363 | |
| 364 | /* |
| 365 | * lwkt_cpusync_start() |
| 366 | * |
| 367 | * Start synchronization with a set of target cpus, return once they are |
| 368 | * known to be in a synchronization loop. The target cpus will execute |
| 369 | * poll->cs_run_func() IN TANDEM WITH THE RETURN. |
| 370 | * |
| 371 | * XXX future: add lwkt_cpusync_start_quick() and require a call to |
| 372 | * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to |
| 373 | * potentially absorb the IPI latency doing something useful. |
| 374 | */ |
| 375 | void |
| 376 | lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll) |
| 377 | { |
| 378 | poll->cs_count = 0; |
| 379 | poll->cs_mask = mask; |
| 380 | poll->cs_maxcount = lwkt_send_ipiq_mask(mask & mycpu->gd_other_cpus, |
| 381 | (ipifunc_t)lwkt_cpusync_remote1, poll); |
| 382 | if (mask & (1 << mycpu->gd_cpuid)) { |
| 383 | if (poll->cs_run_func) |
| 384 | poll->cs_run_func(poll); |
| 385 | } |
| 386 | while (poll->cs_count != poll->cs_maxcount) { |
| 387 | crit_enter(); |
| 388 | lwkt_process_ipiq(); |
| 389 | crit_exit(); |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | void |
| 394 | lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll) |
| 395 | { |
| 396 | mask &= ~poll->cs_mask; |
| 397 | poll->cs_mask |= mask; |
| 398 | poll->cs_maxcount += lwkt_send_ipiq_mask(mask & mycpu->gd_other_cpus, |
| 399 | (ipifunc_t)lwkt_cpusync_remote1, poll); |
| 400 | if (mask & (1 << mycpu->gd_cpuid)) { |
| 401 | if (poll->cs_run_func) |
| 402 | poll->cs_run_func(poll); |
| 403 | } |
| 404 | while (poll->cs_count != poll->cs_maxcount) { |
| 405 | crit_enter(); |
| 406 | lwkt_process_ipiq(); |
| 407 | crit_exit(); |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | /* |
| 412 | * Finish synchronization with a set of target cpus. The target cpus will |
| 413 | * execute cs_fin1_func(poll) prior to this function returning, and will |
| 414 | * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN. |
| 415 | */ |
| 416 | void |
| 417 | lwkt_cpusync_finish(lwkt_cpusync_t poll) |
| 418 | { |
| 419 | int count; |
| 420 | |
| 421 | count = -(poll->cs_maxcount + 1); |
| 422 | poll->cs_count = -1; |
| 423 | if (poll->cs_mask & (1 << mycpu->gd_cpuid)) { |
| 424 | if (poll->cs_fin1_func) |
| 425 | poll->cs_fin1_func(poll); |
| 426 | if (poll->cs_fin2_func) |
| 427 | poll->cs_fin2_func(poll->cs_data); |
| 428 | } |
| 429 | while (poll->cs_count != count) { |
| 430 | crit_enter(); |
| 431 | lwkt_process_ipiq(); |
| 432 | crit_exit(); |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | /* |
| 437 | * helper IPI remote messaging function. |
| 438 | * |
| 439 | * Called on remote cpu when a new cpu synchronization request has been |
| 440 | * sent to us. Execute the run function and adjust cs_count, then requeue |
| 441 | * the request so we spin on it. |
| 442 | */ |
| 443 | static void |
| 444 | lwkt_cpusync_remote1(lwkt_cpusync_t poll) |
| 445 | { |
| 446 | atomic_add_int(&poll->cs_count, 1); |
| 447 | if (poll->cs_run_func) |
| 448 | poll->cs_run_func(poll); |
| 449 | lwkt_cpusync_remote2(poll); |
| 450 | } |
| 451 | |
| 452 | /* |
| 453 | * helper IPI remote messaging function. |
| 454 | * |
| 455 | * Poll for the originator telling us to finish. If it hasn't, requeue |
| 456 | * our request so we spin on it. When the originator requests that we |
| 457 | * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data) |
| 458 | * in tandem with the release. |
| 459 | */ |
| 460 | static void |
| 461 | lwkt_cpusync_remote2(lwkt_cpusync_t poll) |
| 462 | { |
| 463 | if (poll->cs_count < 0) { |
| 464 | cpusync_func2_t savef; |
| 465 | void *saved; |
| 466 | |
| 467 | if (poll->cs_fin1_func) |
| 468 | poll->cs_fin1_func(poll); |
| 469 | if (poll->cs_fin2_func) { |
| 470 | savef = poll->cs_fin2_func; |
| 471 | saved = poll->cs_data; |
| 472 | atomic_add_int(&poll->cs_count, -1); |
| 473 | savef(saved); |
| 474 | } else { |
| 475 | atomic_add_int(&poll->cs_count, -1); |
| 476 | } |
| 477 | } else { |
| 478 | globaldata_t gd = mycpu; |
| 479 | lwkt_ipiq_t ip; |
| 480 | int wi; |
| 481 | |
| 482 | ip = &gd->gd_cpusyncq; |
| 483 | wi = ip->ip_windex & MAXCPUFIFO_MASK; |
| 484 | ip->ip_func[wi] = (ipifunc2_t)lwkt_cpusync_remote2; |
| 485 | ip->ip_arg[wi] = poll; |
| 486 | ++ip->ip_windex; |
| 487 | } |
| 488 | } |
| 489 | |
| 490 | #else |
| 491 | |
| 492 | /* |
| 493 | * !SMP dummy routines |
| 494 | */ |
| 495 | |
| 496 | int |
| 497 | lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg) |
| 498 | { |
| 499 | panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", target->gd_cpuid, func, arg); |
| 500 | return(0); /* NOT REACHED */ |
| 501 | } |
| 502 | |
| 503 | void |
| 504 | lwkt_wait_ipiq(globaldata_t target, int seq) |
| 505 | { |
| 506 | panic("lwkt_wait_ipiq: UP box! (%d,%d)", target->gd_cpuid, seq); |
| 507 | } |
| 508 | |
| 509 | #endif |