1 /* 2 * Copyright (c) 2003-2016 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 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 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 /* 36 * This module implements IPI message queueing and the MI portion of IPI 37 * message processing. 38 */ 39 40 #include "opt_ddb.h" 41 42 #include <sys/param.h> 43 #include <sys/systm.h> 44 #include <sys/kernel.h> 45 #include <sys/proc.h> 46 #include <sys/rtprio.h> 47 #include <sys/queue.h> 48 #include <sys/thread2.h> 49 #include <sys/sysctl.h> 50 #include <sys/ktr.h> 51 #include <sys/kthread.h> 52 #include <machine/cpu.h> 53 #include <sys/lock.h> 54 55 #include <vm/vm.h> 56 #include <vm/vm_param.h> 57 #include <vm/vm_kern.h> 58 #include <vm/vm_object.h> 59 #include <vm/vm_page.h> 60 #include <vm/vm_map.h> 61 #include <vm/vm_pager.h> 62 #include <vm/vm_extern.h> 63 #include <vm/vm_zone.h> 64 65 #include <machine/stdarg.h> 66 #include <machine/smp.h> 67 #include <machine/clock.h> 68 #include <machine/atomic.h> 69 70 #ifdef _KERNEL_VIRTUAL 71 #include <pthread.h> 72 #endif 73 74 struct ipiq_stats { 75 int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */ 76 int64_t ipiq_fifofull; /* number of fifo full conditions detected */ 77 int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */ 78 int64_t ipiq_passive; /* passive IPI messages */ 79 int64_t ipiq_cscount; /* number of cpu synchronizations */ 80 } __cachealign; 81 82 static struct ipiq_stats ipiq_stats_percpu[MAXCPU]; 83 #define ipiq_stat(gd) ipiq_stats_percpu[(gd)->gd_cpuid] 84 85 static int ipiq_debug; /* set to 1 for debug */ 86 #ifdef PANIC_DEBUG 87 static int panic_ipiq_cpu = -1; 88 static int panic_ipiq_count = 100; 89 #endif 90 91 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_debug, CTLFLAG_RW, &ipiq_debug, 0, 92 ""); 93 #ifdef PANIC_DEBUG 94 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, ""); 95 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, ""); 96 #endif 97 98 #define IPIQ_STRING "func=%p arg1=%p arg2=%d scpu=%d dcpu=%d" 99 #define IPIQ_ARGS void *func, void *arg1, int arg2, int scpu, int dcpu 100 101 #if !defined(KTR_IPIQ) 102 #define KTR_IPIQ KTR_ALL 103 #endif 104 KTR_INFO_MASTER(ipiq); 105 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARGS); 106 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARGS); 107 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARGS); 108 KTR_INFO(KTR_IPIQ, ipiq, sync_start, 5, "cpumask=%08lx", unsigned long mask); 109 KTR_INFO(KTR_IPIQ, ipiq, sync_end, 6, "cpumask=%08lx", unsigned long mask); 110 KTR_INFO(KTR_IPIQ, ipiq, cpu_send, 7, IPIQ_STRING, IPIQ_ARGS); 111 KTR_INFO(KTR_IPIQ, ipiq, send_end, 8, IPIQ_STRING, IPIQ_ARGS); 112 KTR_INFO(KTR_IPIQ, ipiq, sync_quick, 9, "cpumask=%08lx", unsigned long mask); 113 114 #define logipiq(name, func, arg1, arg2, sgd, dgd) \ 115 KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid) 116 #define logipiq2(name, arg) \ 117 KTR_LOG(ipiq_ ## name, arg) 118 119 static void lwkt_process_ipiq_nested(void); 120 static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip, 121 struct intrframe *frame, int limit); 122 static void lwkt_cpusync_remote1(lwkt_cpusync_t cs); 123 static void lwkt_cpusync_remote2(lwkt_cpusync_t cs); 124 125 #define IPIQ_SYSCTL(name) \ 126 static int \ 127 sysctl_##name(SYSCTL_HANDLER_ARGS) \ 128 { \ 129 int64_t val = 0; \ 130 int cpu, error; \ 131 \ 132 for (cpu = 0; cpu < ncpus; ++cpu) \ 133 val += ipiq_stats_percpu[cpu].name; \ 134 \ 135 error = sysctl_handle_quad(oidp, &val, 0, req); \ 136 if (error || req->newptr == NULL) \ 137 return error; \ 138 \ 139 for (cpu = 0; cpu < ncpus; ++cpu) \ 140 ipiq_stats_percpu[cpu].name = val; \ 141 \ 142 return 0; \ 143 } 144 145 IPIQ_SYSCTL(ipiq_count); 146 IPIQ_SYSCTL(ipiq_fifofull); 147 IPIQ_SYSCTL(ipiq_avoided); 148 IPIQ_SYSCTL(ipiq_passive); 149 IPIQ_SYSCTL(ipiq_cscount); 150 151 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_count, (CTLTYPE_QUAD | CTLFLAG_RW), 152 0, 0, sysctl_ipiq_count, "Q", "Number of IPI's sent"); 153 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_fifofull, (CTLTYPE_QUAD | CTLFLAG_RW), 154 0, 0, sysctl_ipiq_fifofull, "Q", 155 "Number of fifo full conditions detected"); 156 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_avoided, (CTLTYPE_QUAD | CTLFLAG_RW), 157 0, 0, sysctl_ipiq_avoided, "Q", 158 "Number of IPI's avoided by interlock with target cpu"); 159 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_passive, (CTLTYPE_QUAD | CTLFLAG_RW), 160 0, 0, sysctl_ipiq_passive, "Q", 161 "Number of passive IPI messages sent"); 162 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_cscount, (CTLTYPE_QUAD | CTLFLAG_RW), 163 0, 0, sysctl_ipiq_cscount, "Q", 164 "Number of cpu synchronizations"); 165 166 /* 167 * Send a function execution request to another cpu. The request is queued 168 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every 169 * possible target cpu. The FIFO can be written. 170 * 171 * If the FIFO fills up we have to enable interrupts to avoid an APIC 172 * deadlock and process pending IPIQs while waiting for it to empty. 173 * Otherwise we may soft-deadlock with another cpu whos FIFO is also full. 174 * 175 * We can safely bump gd_intr_nesting_level because our crit_exit() at the 176 * end will take care of any pending interrupts. 177 * 178 * The actual hardware IPI is avoided if the target cpu is already processing 179 * the queue from a prior IPI. It is possible to pipeline IPI messages 180 * very quickly between cpus due to the FIFO hysteresis. 181 * 182 * Need not be called from a critical section. 183 */ 184 int 185 lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2) 186 { 187 lwkt_ipiq_t ip; 188 int windex; 189 int level1; 190 int level2; 191 long rflags; 192 struct globaldata *gd = mycpu; 193 194 logipiq(send_norm, func, arg1, arg2, gd, target); 195 196 if (target == gd) { 197 func(arg1, arg2, NULL); 198 logipiq(send_end, func, arg1, arg2, gd, target); 199 return(0); 200 } 201 crit_enter(); 202 ++gd->gd_intr_nesting_level; 203 #ifdef INVARIANTS 204 if (gd->gd_intr_nesting_level > 20) 205 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!"); 206 #endif 207 KKASSERT(curthread->td_critcount); 208 ++ipiq_stat(gd).ipiq_count; 209 ip = &gd->gd_ipiq[target->gd_cpuid]; 210 211 /* 212 * Do not allow the FIFO to become full. Interrupts must be physically 213 * enabled while we liveloop to avoid deadlocking the APIC. 214 * 215 * When we are not nested inside a processing loop we allow the FIFO 216 * to get 1/2 full. Once it exceeds 1/2 full we must wait for it to 217 * drain, executing any incoming IPIs while we wait. 218 * 219 * When we are nested we allow the FIFO to get almost completely full. 220 * This allows us to queue IPIs sent from IPI callbacks. The processing 221 * code will only process incoming FIFOs that are trying to drain while 222 * we wait, and only to the only-slightly-less-full point, to avoid a 223 * deadlock. 224 * 225 * We are guaranteed 226 */ 227 228 if (gd->gd_processing_ipiq == 0) { 229 level1 = MAXCPUFIFO / 2; 230 level2 = MAXCPUFIFO / 4; 231 } else { 232 level1 = MAXCPUFIFO - 3; 233 level2 = MAXCPUFIFO - 5; 234 } 235 236 if (ip->ip_windex - ip->ip_rindex > level1) { 237 #ifndef _KERNEL_VIRTUAL 238 uint64_t tsc_base = rdtsc(); 239 #endif 240 int repeating = 0; 241 int olimit; 242 243 rflags = read_rflags(); 244 cpu_enable_intr(); 245 ++ipiq_stat(gd).ipiq_fifofull; 246 DEBUG_PUSH_INFO("send_ipiq3"); 247 olimit = atomic_swap_int(&ip->ip_drain, level2); 248 while (ip->ip_windex - ip->ip_rindex > level2) { 249 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1); 250 lwkt_process_ipiq_nested(); 251 cpu_pause(); 252 253 /* 254 * Check for target not draining issue. This should be fixed but 255 * leave the code in-place anyway as it can recover an otherwise 256 * dead system. 257 */ 258 #ifdef _KERNEL_VIRTUAL 259 if (repeating++ > 10) 260 pthread_yield(); 261 #else 262 if (rdtsc() - tsc_base > tsc_frequency) { 263 ++repeating; 264 if (repeating > 10) { 265 kprintf("send_ipiq %d->%d tgt not draining (%d) sniff=%p,%p\n", 266 gd->gd_cpuid, target->gd_cpuid, repeating, 267 target->gd_sample_pc, target->gd_sample_sp); 268 smp_sniff(); 269 ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid); 270 cpu_send_ipiq(target->gd_cpuid); 271 } else { 272 kprintf("send_ipiq %d->%d tgt not draining (%d)\n", 273 gd->gd_cpuid, target->gd_cpuid, repeating); 274 smp_sniff(); 275 } 276 tsc_base = rdtsc(); 277 } 278 #endif 279 } 280 atomic_swap_int(&ip->ip_drain, olimit); 281 DEBUG_POP_INFO(); 282 #if defined(__x86_64__) 283 write_rflags(rflags); 284 #else 285 #error "no write_*flags" 286 #endif 287 } 288 289 /* 290 * Queue the new message and signal the target cpu. For now we need to 291 * physically disable interrupts because the target will not get signalled 292 * by other cpus once we set target->gd_npoll and we don't want to get 293 * interrupted. 294 * 295 * XXX not sure why this is a problem, the critical section should prevent 296 * any stalls (incoming interrupts except Xinvltlb and Xsnoop will 297 * just be made pending). 298 */ 299 rflags = read_rflags(); 300 #ifndef _KERNEL_VIRTUAL 301 cpu_disable_intr(); 302 #endif 303 304 windex = ip->ip_windex & MAXCPUFIFO_MASK; 305 ip->ip_info[windex].func = func; 306 ip->ip_info[windex].arg1 = arg1; 307 ip->ip_info[windex].arg2 = arg2; 308 cpu_sfence(); 309 ++ip->ip_windex; 310 ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid); 311 312 /* 313 * signal the target cpu that there is work pending. 314 */ 315 if (atomic_swap_int(&target->gd_npoll, 1) == 0) { 316 logipiq(cpu_send, func, arg1, arg2, gd, target); 317 cpu_send_ipiq(target->gd_cpuid); 318 } else { 319 ++ipiq_stat(gd).ipiq_avoided; 320 } 321 write_rflags(rflags); 322 323 --gd->gd_intr_nesting_level; 324 crit_exit(); 325 logipiq(send_end, func, arg1, arg2, gd, target); 326 327 return(ip->ip_windex); 328 } 329 330 /* 331 * Similar to lwkt_send_ipiq() but this function does not actually initiate 332 * the IPI to the target cpu unless the FIFO is greater than 1/4 full. 333 * This function is usually very fast. 334 * 335 * This function is used for non-critical IPI messages, such as memory 336 * deallocations. The queue will typically be flushed by the target cpu at 337 * the next clock interrupt. 338 * 339 * Need not be called from a critical section. 340 */ 341 int 342 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func, 343 void *arg1, int arg2) 344 { 345 lwkt_ipiq_t ip; 346 int windex; 347 struct globaldata *gd = mycpu; 348 349 KKASSERT(target != gd); 350 crit_enter_gd(gd); 351 ++gd->gd_intr_nesting_level; 352 ip = &gd->gd_ipiq[target->gd_cpuid]; 353 354 /* 355 * If the FIFO is too full send the IPI actively. 356 * 357 * WARNING! This level must be low enough not to trigger a wait loop 358 * in the active sending code since we are not signalling the 359 * target cpu. 360 */ 361 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO / 4) { 362 --gd->gd_intr_nesting_level; 363 crit_exit_gd(gd); 364 return lwkt_send_ipiq3(target, func, arg1, arg2); 365 } 366 367 /* 368 * Else we can do it passively. 369 */ 370 logipiq(send_pasv, func, arg1, arg2, gd, target); 371 ++ipiq_stat(gd).ipiq_count; 372 ++ipiq_stat(gd).ipiq_passive; 373 374 /* 375 * Queue the new message 376 */ 377 windex = ip->ip_windex & MAXCPUFIFO_MASK; 378 ip->ip_info[windex].func = func; 379 ip->ip_info[windex].arg1 = arg1; 380 ip->ip_info[windex].arg2 = arg2; 381 cpu_sfence(); 382 ++ip->ip_windex; 383 ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid); 384 --gd->gd_intr_nesting_level; 385 386 /* 387 * Do not signal the target cpu, it will pick up the IPI when it next 388 * polls (typically on the next tick). 389 */ 390 crit_exit(); 391 logipiq(send_end, func, arg1, arg2, gd, target); 392 393 return(ip->ip_windex); 394 } 395 396 /* 397 * deprecated, used only by fast int forwarding. 398 */ 399 int 400 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2) 401 { 402 return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2)); 403 } 404 405 /* 406 * Send a message to several target cpus. Typically used for scheduling. 407 * The message will not be sent to stopped cpus. 408 * 409 * To prevent treating low-numbered cpus as favored sons, the IPIs are 410 * issued in order starting at mycpu upward, then from 0 through mycpu. 411 * This is particularly important to prevent random scheduler pickups 412 * from favoring cpu 0. 413 */ 414 int 415 lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2) 416 { 417 int cpuid; 418 int count = 0; 419 cpumask_t amask; 420 421 CPUMASK_NANDMASK(mask, stopped_cpus); 422 423 /* 424 * All cpus in mask which are >= mycpu 425 */ 426 CPUMASK_ASSBMASK(amask, mycpu->gd_cpuid); 427 CPUMASK_INVMASK(amask); 428 CPUMASK_ANDMASK(amask, mask); 429 while (CPUMASK_TESTNZERO(amask)) { 430 cpuid = BSFCPUMASK(amask); 431 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2); 432 CPUMASK_NANDBIT(amask, cpuid); 433 ++count; 434 } 435 436 /* 437 * All cpus in mask which are < mycpu 438 */ 439 CPUMASK_ASSBMASK(amask, mycpu->gd_cpuid); 440 CPUMASK_ANDMASK(amask, mask); 441 while (CPUMASK_TESTNZERO(amask)) { 442 cpuid = BSFCPUMASK(amask); 443 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2); 444 CPUMASK_NANDBIT(amask, cpuid); 445 ++count; 446 } 447 return(count); 448 } 449 450 /* 451 * Wait for the remote cpu to finish processing a function. 452 * 453 * YYY we have to enable interrupts and process the IPIQ while waiting 454 * for it to empty or we may deadlock with another cpu. Create a CPU_*() 455 * function to do this! YYY we really should 'block' here. 456 * 457 * MUST be called from a critical section. This routine may be called 458 * from an interrupt (for example, if an interrupt wakes a foreign thread 459 * up). 460 */ 461 void 462 lwkt_wait_ipiq(globaldata_t target, int seq) 463 { 464 lwkt_ipiq_t ip; 465 466 if (target != mycpu) { 467 ip = &mycpu->gd_ipiq[target->gd_cpuid]; 468 if ((int)(ip->ip_xindex - seq) < 0) { 469 #if defined(__x86_64__) 470 unsigned long rflags = read_rflags(); 471 #else 472 #error "no read_*flags" 473 #endif 474 int64_t time_tgt = tsc_get_target(1000000000LL); 475 int time_loops = 10; 476 int benice = 0; 477 #ifdef _KERNEL_VIRTUAL 478 int repeating = 0; 479 #endif 480 481 cpu_enable_intr(); 482 DEBUG_PUSH_INFO("wait_ipiq"); 483 while ((int)(ip->ip_xindex - seq) < 0) { 484 crit_enter(); 485 lwkt_process_ipiq(); 486 crit_exit(); 487 #ifdef _KERNEL_VIRTUAL 488 if (repeating++ > 10) 489 pthread_yield(); 490 #endif 491 492 /* 493 * IPIQs must be handled within 10 seconds and this code 494 * will warn after one second. 495 */ 496 if ((benice & 255) == 0 && tsc_test_target(time_tgt) > 0) { 497 kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", 498 mycpu->gd_cpuid, target->gd_cpuid, 499 ip->ip_xindex - seq); 500 if (--time_loops == 0) 501 panic("LWKT_WAIT_IPIQ"); 502 time_tgt = tsc_get_target(1000000000LL); 503 } 504 ++benice; 505 506 /* 507 * xindex may be modified by another cpu, use a load fence 508 * to ensure that the loop does not use a speculative value 509 * (which may improve performance). 510 */ 511 cpu_pause(); 512 cpu_lfence(); 513 } 514 DEBUG_POP_INFO(); 515 #if defined(__x86_64__) 516 write_rflags(rflags); 517 #else 518 #error "no write_*flags" 519 #endif 520 } 521 } 522 } 523 524 /* 525 * Called from IPI interrupt (like a fast interrupt), which has placed 526 * us in a critical section. The MP lock may or may not be held. 527 * May also be called from doreti or splz, or be reentrantly called 528 * indirectly through the ip_info[].func we run. 529 * 530 * There are two versions, one where no interrupt frame is available (when 531 * called from the send code and from splz, and one where an interrupt 532 * frame is available. 533 * 534 * When the current cpu is mastering a cpusync we do NOT internally loop 535 * on the cpusyncq poll. We also do not re-flag a pending ipi due to 536 * the cpusyncq poll because this can cause doreti/splz to loop internally. 537 * The cpusync master's own loop must be allowed to run to avoid a deadlock. 538 */ 539 void 540 lwkt_process_ipiq(void) 541 { 542 globaldata_t gd = mycpu; 543 globaldata_t sgd; 544 lwkt_ipiq_t ip; 545 cpumask_t mask; 546 int n; 547 548 ++gd->gd_processing_ipiq; 549 again: 550 mask = gd->gd_ipimask; 551 cpu_ccfence(); 552 while (CPUMASK_TESTNZERO(mask)) { 553 n = BSFCPUMASK(mask); 554 if (n != gd->gd_cpuid) { 555 sgd = globaldata_find(n); 556 ip = sgd->gd_ipiq; 557 if (ip != NULL) { 558 ip += gd->gd_cpuid; 559 while (lwkt_process_ipiq_core(sgd, ip, NULL, 0)) 560 ; 561 ATOMIC_CPUMASK_NANDBIT(gd->gd_ipimask, n); 562 if (ip->ip_rindex != ip->ip_windex) 563 ATOMIC_CPUMASK_ORBIT(gd->gd_ipimask, n); 564 } 565 } 566 CPUMASK_NANDBIT(mask, n); 567 } 568 569 /* 570 * Process pending cpusyncs. If the current thread has a cpusync 571 * active cpusync we only run the list once and do not re-flag 572 * as the thread itself is processing its interlock. 573 */ 574 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL, 0)) { 575 if (gd->gd_curthread->td_cscount == 0) 576 goto again; 577 /* need_ipiq(); do not reflag */ 578 } 579 580 /* 581 * Interlock to allow more IPI interrupts. 582 */ 583 --gd->gd_processing_ipiq; 584 } 585 586 void 587 lwkt_process_ipiq_frame(struct intrframe *frame) 588 { 589 globaldata_t gd = mycpu; 590 globaldata_t sgd; 591 lwkt_ipiq_t ip; 592 cpumask_t mask; 593 int n; 594 595 ++gd->gd_processing_ipiq; 596 again: 597 mask = gd->gd_ipimask; 598 cpu_ccfence(); 599 while (CPUMASK_TESTNZERO(mask)) { 600 n = BSFCPUMASK(mask); 601 if (n != gd->gd_cpuid) { 602 sgd = globaldata_find(n); 603 ip = sgd->gd_ipiq; 604 if (ip != NULL) { 605 ip += gd->gd_cpuid; 606 while (lwkt_process_ipiq_core(sgd, ip, frame, 0)) 607 ; 608 ATOMIC_CPUMASK_NANDBIT(gd->gd_ipimask, n); 609 if (ip->ip_rindex != ip->ip_windex) 610 ATOMIC_CPUMASK_ORBIT(gd->gd_ipimask, n); 611 } 612 } 613 CPUMASK_NANDBIT(mask, n); 614 } 615 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) { 616 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame, 0)) { 617 if (gd->gd_curthread->td_cscount == 0) 618 goto again; 619 /* need_ipiq(); do not reflag */ 620 } 621 } 622 --gd->gd_processing_ipiq; 623 } 624 625 /* 626 * Only process incoming IPIQs from draining senders and only process them 627 * to the point where the draining sender is able to continue. This is 628 * necessary to avoid deadlocking the IPI subsystem because we are acting on 629 * incoming messages and the callback may queue additional messages. 630 * 631 * We only want to have to act on senders that are blocked to limit the 632 * number of additional messages sent. At the same time, recipients are 633 * trying to drain our own queue. Theoretically this create a pipeline that 634 * cannot deadlock. 635 */ 636 static void 637 lwkt_process_ipiq_nested(void) 638 { 639 globaldata_t gd = mycpu; 640 globaldata_t sgd; 641 lwkt_ipiq_t ip; 642 cpumask_t mask; 643 int n; 644 int limit; 645 646 ++gd->gd_processing_ipiq; 647 again: 648 mask = gd->gd_ipimask; 649 cpu_ccfence(); 650 while (CPUMASK_TESTNZERO(mask)) { 651 n = BSFCPUMASK(mask); 652 if (n != gd->gd_cpuid) { 653 sgd = globaldata_find(n); 654 ip = sgd->gd_ipiq; 655 656 /* 657 * NOTE: We do not mess with the cpumask at all, instead we allow 658 * the top-level ipiq processor deal with it. 659 */ 660 if (ip != NULL) { 661 ip += gd->gd_cpuid; 662 if ((limit = ip->ip_drain) != 0) { 663 lwkt_process_ipiq_core(sgd, ip, NULL, limit); 664 /* no gd_ipimask when doing limited processing */ 665 } 666 } 667 } 668 CPUMASK_NANDBIT(mask, n); 669 } 670 671 /* 672 * Process pending cpusyncs. If the current thread has a cpusync 673 * active cpusync we only run the list once and do not re-flag 674 * as the thread itself is processing its interlock. 675 */ 676 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL, 0)) { 677 if (gd->gd_curthread->td_cscount == 0) 678 goto again; 679 /* need_ipiq(); do not reflag */ 680 } 681 --gd->gd_processing_ipiq; 682 } 683 684 /* 685 * Process incoming IPI requests until only <limit> are left (0 to exhaust 686 * all incoming IPI requests). 687 */ 688 static int 689 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip, 690 struct intrframe *frame, int limit) 691 { 692 globaldata_t mygd = mycpu; 693 int ri; 694 int wi; 695 ipifunc3_t copy_func; 696 void *copy_arg1; 697 int copy_arg2; 698 699 /* 700 * Clear the originating core from our ipimask, we will process all 701 * incoming messages. 702 * 703 * Obtain the current write index, which is modified by a remote cpu. 704 * Issue a load fence to prevent speculative reads of e.g. data written 705 * by the other cpu prior to it updating the index. 706 */ 707 KKASSERT(curthread->td_critcount); 708 wi = ip->ip_windex; 709 cpu_lfence(); 710 ++mygd->gd_intr_nesting_level; 711 712 /* 713 * NOTE: xindex is only updated after we are sure the function has 714 * finished execution. Beware lwkt_process_ipiq() reentrancy! 715 * The function may send an IPI which may block/drain. 716 * 717 * NOTE: Due to additional IPI operations that the callback function 718 * may make, it is possible for both rindex and windex to advance and 719 * thus for rindex to advance passed our cached windex. 720 * 721 * NOTE: A load fence is required to prevent speculative loads prior 722 * to the loading of ip_rindex. Even though stores might be 723 * ordered, loads are probably not. A memory fence is required 724 * to prevent reordering of the loads after the ip_rindex update. 725 * 726 * NOTE: Single pass only. Returns non-zero if the queue is not empty 727 * on return. 728 */ 729 while (wi - (ri = ip->ip_rindex) > limit) { 730 ri &= MAXCPUFIFO_MASK; 731 cpu_lfence(); 732 copy_func = ip->ip_info[ri].func; 733 copy_arg1 = ip->ip_info[ri].arg1; 734 copy_arg2 = ip->ip_info[ri].arg2; 735 cpu_mfence(); 736 ++ip->ip_rindex; 737 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) == 738 ((ri + 1) & MAXCPUFIFO_MASK)); 739 logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu); 740 #ifdef INVARIANTS 741 if (ipiq_debug && (ip->ip_rindex & 0xFFFFFF) == 0) { 742 kprintf("cpu %d ipifunc %p %p %d (frame %p)\n", 743 mycpu->gd_cpuid, 744 copy_func, copy_arg1, copy_arg2, 745 #if defined(__x86_64__) 746 (frame ? (void *)frame->if_rip : NULL)); 747 #else 748 NULL); 749 #endif 750 } 751 #endif 752 copy_func(copy_arg1, copy_arg2, frame); 753 cpu_sfence(); 754 ip->ip_xindex = ip->ip_rindex; 755 756 #ifdef PANIC_DEBUG 757 /* 758 * Simulate panics during the processing of an IPI 759 */ 760 if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) { 761 if (--panic_ipiq_count == 0) { 762 #ifdef DDB 763 Debugger("PANIC_DEBUG"); 764 #else 765 panic("PANIC_DEBUG"); 766 #endif 767 } 768 } 769 #endif 770 } 771 --mygd->gd_intr_nesting_level; 772 773 /* 774 * Return non-zero if there is still more in the queue. Don't worry 775 * about fencing, we will get another interrupt if necessary. 776 */ 777 return (ip->ip_rindex != ip->ip_windex); 778 } 779 780 static void 781 lwkt_sync_ipiq(void *arg) 782 { 783 volatile cpumask_t *cpumask = arg; 784 785 ATOMIC_CPUMASK_NANDBIT(*cpumask, mycpu->gd_cpuid); 786 if (CPUMASK_TESTZERO(*cpumask)) 787 wakeup(cpumask); 788 } 789 790 void 791 lwkt_synchronize_ipiqs(const char *wmesg) 792 { 793 volatile cpumask_t other_cpumask; 794 795 other_cpumask = smp_active_mask; 796 CPUMASK_ANDMASK(other_cpumask, mycpu->gd_other_cpus); 797 lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq, 798 __DEVOLATILE(void *, &other_cpumask)); 799 800 while (CPUMASK_TESTNZERO(other_cpumask)) { 801 tsleep_interlock(&other_cpumask, 0); 802 if (CPUMASK_TESTNZERO(other_cpumask)) 803 tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0); 804 } 805 } 806 807 /* 808 * CPU Synchronization Support 809 * 810 * lwkt_cpusync_interlock() - Place specified cpus in a quiescent state. 811 * The current cpu is placed in a hard critical 812 * section. 813 * 814 * lwkt_cpusync_deinterlock() - Execute cs_func on specified cpus, including 815 * current cpu if specified, then return. 816 */ 817 void 818 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *arg) 819 { 820 struct lwkt_cpusync cs; 821 822 lwkt_cpusync_init(&cs, mask, func, arg); 823 lwkt_cpusync_interlock(&cs); 824 lwkt_cpusync_deinterlock(&cs); 825 } 826 827 828 void 829 lwkt_cpusync_interlock(lwkt_cpusync_t cs) 830 { 831 globaldata_t gd = mycpu; 832 cpumask_t mask; 833 834 /* 835 * mask acknowledge (cs_mack): 0->mask for stage 1 836 * 837 * mack does not include the current cpu. 838 */ 839 mask = cs->cs_mask; 840 CPUMASK_ANDMASK(mask, gd->gd_other_cpus); 841 CPUMASK_ANDMASK(mask, smp_active_mask); 842 CPUMASK_ASSZERO(cs->cs_mack); 843 844 crit_enter_id("cpusync"); 845 if (CPUMASK_TESTNZERO(mask)) { 846 DEBUG_PUSH_INFO("cpusync_interlock"); 847 ++ipiq_stat(gd).ipiq_cscount; 848 ++gd->gd_curthread->td_cscount; 849 lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote1, cs); 850 logipiq2(sync_start, (long)CPUMASK_LOWMASK(mask)); 851 while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) { 852 lwkt_process_ipiq(); 853 cpu_pause(); 854 #ifdef _KERNEL_VIRTUAL 855 pthread_yield(); 856 #endif 857 } 858 DEBUG_POP_INFO(); 859 } 860 } 861 862 /* 863 * Interlocked cpus have executed remote1 and are polling in remote2. 864 * To deinterlock we clear cs_mack and wait for the cpus to execute 865 * the func and set their bit in cs_mack again. 866 * 867 */ 868 void 869 lwkt_cpusync_deinterlock(lwkt_cpusync_t cs) 870 { 871 globaldata_t gd = mycpu; 872 cpumask_t mask; 873 874 /* 875 * mask acknowledge (cs_mack): mack->0->mack for stage 2 876 * 877 * Clearing cpu bits for polling cpus in cs_mack will cause them to 878 * execute stage 2, which executes the cs_func(cs_data) and then sets 879 * their bit in cs_mack again. 880 * 881 * mack does not include the current cpu. 882 */ 883 mask = cs->cs_mack; 884 cpu_ccfence(); 885 CPUMASK_ASSZERO(cs->cs_mack); 886 cpu_ccfence(); 887 if (cs->cs_func && CPUMASK_TESTBIT(cs->cs_mask, gd->gd_cpuid)) 888 cs->cs_func(cs->cs_data); 889 if (CPUMASK_TESTNZERO(mask)) { 890 DEBUG_PUSH_INFO("cpusync_deinterlock"); 891 while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) { 892 lwkt_process_ipiq(); 893 cpu_pause(); 894 #ifdef _KERNEL_VIRTUAL 895 pthread_yield(); 896 #endif 897 } 898 DEBUG_POP_INFO(); 899 /* 900 * cpusyncq ipis may be left queued without the RQF flag set due to 901 * a non-zero td_cscount, so be sure to process any laggards after 902 * decrementing td_cscount. 903 */ 904 --gd->gd_curthread->td_cscount; 905 lwkt_process_ipiq(); 906 logipiq2(sync_end, (long)CPUMASK_LOWMASK(mask)); 907 } 908 crit_exit_id("cpusync"); 909 } 910 911 /* 912 * The quick version does not quiesce the target cpu(s) but instead executes 913 * the function on the target cpu(s) and waits for all to acknowledge. This 914 * avoids spinning on the target cpus. 915 * 916 * This function is typically only used for kernel_pmap updates. User pmaps 917 * have to be quiesced. 918 */ 919 void 920 lwkt_cpusync_quick(lwkt_cpusync_t cs) 921 { 922 globaldata_t gd = mycpu; 923 cpumask_t mask; 924 925 /* 926 * stage-2 cs_mack only. 927 */ 928 mask = cs->cs_mask; 929 CPUMASK_ANDMASK(mask, gd->gd_other_cpus); 930 CPUMASK_ANDMASK(mask, smp_active_mask); 931 CPUMASK_ASSZERO(cs->cs_mack); 932 933 crit_enter_id("cpusync"); 934 if (CPUMASK_TESTNZERO(mask)) { 935 DEBUG_PUSH_INFO("cpusync_interlock"); 936 ++ipiq_stat(gd).ipiq_cscount; 937 ++gd->gd_curthread->td_cscount; 938 lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote2, cs); 939 logipiq2(sync_quick, (long)CPUMASK_LOWMASK(mask)); 940 while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) { 941 lwkt_process_ipiq(); 942 cpu_pause(); 943 #ifdef _KERNEL_VIRTUAL 944 pthread_yield(); 945 #endif 946 } 947 948 /* 949 * cpusyncq ipis may be left queued without the RQF flag set due to 950 * a non-zero td_cscount, so be sure to process any laggards after 951 * decrementing td_cscount. 952 */ 953 DEBUG_POP_INFO(); 954 --gd->gd_curthread->td_cscount; 955 lwkt_process_ipiq(); 956 } 957 if (cs->cs_func && CPUMASK_TESTBIT(cs->cs_mask, gd->gd_cpuid)) 958 cs->cs_func(cs->cs_data); 959 crit_exit_id("cpusync"); 960 } 961 962 /* 963 * helper IPI remote messaging function. 964 * 965 * Called on remote cpu when a new cpu synchronization request has been 966 * sent to us. Execute the run function and adjust cs_count, then requeue 967 * the request so we spin on it. 968 */ 969 static void 970 lwkt_cpusync_remote1(lwkt_cpusync_t cs) 971 { 972 globaldata_t gd = mycpu; 973 974 ATOMIC_CPUMASK_ORBIT(cs->cs_mack, gd->gd_cpuid); 975 lwkt_cpusync_remote2(cs); 976 } 977 978 /* 979 * helper IPI remote messaging function. 980 * 981 * Poll for the originator telling us to finish. If it hasn't, requeue 982 * our request so we spin on it. 983 */ 984 static void 985 lwkt_cpusync_remote2(lwkt_cpusync_t cs) 986 { 987 globaldata_t gd = mycpu; 988 989 if (CPUMASK_TESTMASK(cs->cs_mack, gd->gd_cpumask) == 0) { 990 if (cs->cs_func) 991 cs->cs_func(cs->cs_data); 992 ATOMIC_CPUMASK_ORBIT(cs->cs_mack, gd->gd_cpuid); 993 /* cs can be ripped out at this point */ 994 } else { 995 lwkt_ipiq_t ip; 996 int wi; 997 998 cpu_pause(); 999 #ifdef _KERNEL_VIRTUAL 1000 pthread_yield(); 1001 #endif 1002 cpu_lfence(); 1003 1004 /* 1005 * Requeue our IPI to avoid a deep stack recursion. If no other 1006 * IPIs are pending we can just loop up, which should help VMs 1007 * better-detect spin loops. 1008 */ 1009 ip = &gd->gd_cpusyncq; 1010 1011 wi = ip->ip_windex & MAXCPUFIFO_MASK; 1012 ip->ip_info[wi].func = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2; 1013 ip->ip_info[wi].arg1 = cs; 1014 ip->ip_info[wi].arg2 = 0; 1015 cpu_sfence(); 1016 KKASSERT(ip->ip_windex - ip->ip_rindex < MAXCPUFIFO); 1017 ++ip->ip_windex; 1018 if (ipiq_debug && (ip->ip_windex & 0xFFFFFF) == 0) { 1019 kprintf("cpu %d cm=%016jx %016jx f=%p\n", 1020 gd->gd_cpuid, 1021 (intmax_t)CPUMASK_LOWMASK(cs->cs_mask), 1022 (intmax_t)CPUMASK_LOWMASK(cs->cs_mack), 1023 cs->cs_func); 1024 } 1025 } 1026 } 1027 1028 #define LWKT_IPIQ_NLATENCY 8 1029 #define LWKT_IPIQ_NLATENCY_MASK (LWKT_IPIQ_NLATENCY - 1) 1030 1031 struct lwkt_ipiq_latency_log { 1032 int idx; /* unmasked index */ 1033 int pad; 1034 uint64_t latency[LWKT_IPIQ_NLATENCY]; 1035 }; 1036 1037 static struct lwkt_ipiq_latency_log lwkt_ipiq_latency_logs[MAXCPU]; 1038 static uint64_t save_tsc; 1039 1040 /* 1041 * IPI callback (already in a critical section) 1042 */ 1043 static void 1044 lwkt_ipiq_latency_testfunc(void *arg __unused) 1045 { 1046 uint64_t delta_tsc; 1047 struct globaldata *gd; 1048 struct lwkt_ipiq_latency_log *lat; 1049 1050 /* 1051 * Get delta TSC (assume TSCs are synchronized) as quickly as 1052 * possible and then convert to nanoseconds. 1053 */ 1054 delta_tsc = rdtsc_ordered() - save_tsc; 1055 delta_tsc = delta_tsc * 1000000000LU / tsc_frequency; 1056 1057 /* 1058 * Record in our save array. 1059 */ 1060 gd = mycpu; 1061 lat = &lwkt_ipiq_latency_logs[gd->gd_cpuid]; 1062 lat->latency[lat->idx & LWKT_IPIQ_NLATENCY_MASK] = delta_tsc; 1063 ++lat->idx; 1064 } 1065 1066 /* 1067 * Send IPI from cpu0 to other cpus 1068 * 1069 * NOTE: Machine must be idle for test to run dependably, and also probably 1070 * a good idea not to be running powerd. 1071 * 1072 * NOTE: Caller should use 'usched :1 <command>' to lock itself to cpu 0. 1073 * See 'ipitest' script in /usr/src/test/sysperf/ipitest 1074 */ 1075 static int 1076 lwkt_ipiq_latency_test(SYSCTL_HANDLER_ARGS) 1077 { 1078 struct globaldata *gd; 1079 int cpu = 0, orig_cpu, error; 1080 1081 error = sysctl_handle_int(oidp, &cpu, arg2, req); 1082 if (error || req->newptr == NULL) 1083 return error; 1084 1085 if (cpu == 0) 1086 return 0; 1087 else if (cpu >= ncpus || cpu < 0) 1088 return EINVAL; 1089 1090 orig_cpu = mycpuid; 1091 lwkt_migratecpu(0); 1092 1093 gd = globaldata_find(cpu); 1094 1095 save_tsc = rdtsc_ordered(); 1096 lwkt_send_ipiq(gd, lwkt_ipiq_latency_testfunc, NULL); 1097 1098 lwkt_migratecpu(orig_cpu); 1099 return 0; 1100 } 1101 1102 SYSCTL_NODE(_debug, OID_AUTO, ipiq, CTLFLAG_RW, 0, ""); 1103 SYSCTL_PROC(_debug_ipiq, OID_AUTO, latency_test, CTLTYPE_INT | CTLFLAG_RW, 1104 NULL, 0, lwkt_ipiq_latency_test, "I", 1105 "ipi latency test, arg: remote cpuid"); 1106 1107 static int 1108 lwkt_ipiq_latency(SYSCTL_HANDLER_ARGS) 1109 { 1110 struct lwkt_ipiq_latency_log *latency = arg1; 1111 uint64_t lat[LWKT_IPIQ_NLATENCY]; 1112 int i; 1113 1114 for (i = 0; i < LWKT_IPIQ_NLATENCY; ++i) 1115 lat[i] = latency->latency[i]; 1116 1117 return sysctl_handle_opaque(oidp, lat, sizeof(lat), req); 1118 } 1119 1120 static void 1121 lwkt_ipiq_latency_init(void *dummy __unused) 1122 { 1123 int cpu; 1124 1125 for (cpu = 0; cpu < ncpus; ++cpu) { 1126 char name[32]; 1127 1128 ksnprintf(name, sizeof(name), "latency%d", cpu); 1129 SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_debug_ipiq), 1130 OID_AUTO, name, CTLTYPE_OPAQUE | CTLFLAG_RD, 1131 &lwkt_ipiq_latency_logs[cpu], 0, lwkt_ipiq_latency, 1132 "LU", "7 latest ipi latency measurement results"); 1133 } 1134 } 1135 SYSINIT(lwkt_ipiq_latency, SI_SUB_CONFIGURE, SI_ORDER_ANY, 1136 lwkt_ipiq_latency_init, NULL); 1137