| 1 | /* |
| 2 | * Copyright (c) 2003,2004 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 | * $DragonFly: src/sys/kern/lwkt_ipiq.c,v 1.27 2008/05/18 20:57:56 nth Exp $ |
| 35 | */ |
| 36 | |
| 37 | /* |
| 38 | * This module implements IPI message queueing and the MI portion of IPI |
| 39 | * message processing. |
| 40 | */ |
| 41 | |
| 42 | #include "opt_ddb.h" |
| 43 | |
| 44 | #include <sys/param.h> |
| 45 | #include <sys/systm.h> |
| 46 | #include <sys/kernel.h> |
| 47 | #include <sys/proc.h> |
| 48 | #include <sys/rtprio.h> |
| 49 | #include <sys/queue.h> |
| 50 | #include <sys/thread2.h> |
| 51 | #include <sys/sysctl.h> |
| 52 | #include <sys/ktr.h> |
| 53 | #include <sys/kthread.h> |
| 54 | #include <machine/cpu.h> |
| 55 | #include <sys/lock.h> |
| 56 | #include <sys/caps.h> |
| 57 | |
| 58 | #include <vm/vm.h> |
| 59 | #include <vm/vm_param.h> |
| 60 | #include <vm/vm_kern.h> |
| 61 | #include <vm/vm_object.h> |
| 62 | #include <vm/vm_page.h> |
| 63 | #include <vm/vm_map.h> |
| 64 | #include <vm/vm_pager.h> |
| 65 | #include <vm/vm_extern.h> |
| 66 | #include <vm/vm_zone.h> |
| 67 | |
| 68 | #include <machine/stdarg.h> |
| 69 | #include <machine/smp.h> |
| 70 | #include <machine/atomic.h> |
| 71 | |
| 72 | #ifdef SMP |
| 73 | static __int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */ |
| 74 | static __int64_t ipiq_fifofull; /* number of fifo full conditions detected */ |
| 75 | static __int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */ |
| 76 | static __int64_t ipiq_passive; /* passive IPI messages */ |
| 77 | static __int64_t ipiq_cscount; /* number of cpu synchronizations */ |
| 78 | static int ipiq_optimized = 1; /* XXX temporary sysctl */ |
| 79 | #ifdef PANIC_DEBUG |
| 80 | static int panic_ipiq_cpu = -1; |
| 81 | static int panic_ipiq_count = 100; |
| 82 | #endif |
| 83 | #endif |
| 84 | |
| 85 | #ifdef SMP |
| 86 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, |
| 87 | "Number of IPI's sent"); |
| 88 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, |
| 89 | "Number of fifo full conditions detected"); |
| 90 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0, |
| 91 | "Number of IPI's avoided by interlock with target cpu"); |
| 92 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0, |
| 93 | "Number of passive IPI messages sent"); |
| 94 | SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0, |
| 95 | "Number of cpu synchronizations"); |
| 96 | SYSCTL_INT(_lwkt, OID_AUTO, ipiq_optimized, CTLFLAG_RW, &ipiq_optimized, 0, |
| 97 | ""); |
| 98 | #ifdef PANIC_DEBUG |
| 99 | SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, ""); |
| 100 | SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, ""); |
| 101 | #endif |
| 102 | |
| 103 | #define IPIQ_STRING "func=%p arg1=%p arg2=%d scpu=%d dcpu=%d" |
| 104 | #define IPIQ_ARG_SIZE (sizeof(void *) * 2 + sizeof(int) * 3) |
| 105 | |
| 106 | #if !defined(KTR_IPIQ) |
| 107 | #define KTR_IPIQ KTR_ALL |
| 108 | #endif |
| 109 | KTR_INFO_MASTER(ipiq); |
| 110 | KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARG_SIZE); |
| 111 | KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARG_SIZE); |
| 112 | KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARG_SIZE); |
| 113 | KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARG_SIZE); |
| 114 | KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARG_SIZE); |
| 115 | KTR_INFO(KTR_IPIQ, ipiq, sync_start, 5, "cpumask=%08x", sizeof(cpumask_t)); |
| 116 | KTR_INFO(KTR_IPIQ, ipiq, sync_add, 6, "cpumask=%08x", sizeof(cpumask_t)); |
| 117 | KTR_INFO(KTR_IPIQ, ipiq, cpu_send, 7, IPIQ_STRING, IPIQ_ARG_SIZE); |
| 118 | KTR_INFO(KTR_IPIQ, ipiq, send_end, 8, IPIQ_STRING, IPIQ_ARG_SIZE); |
| 119 | |
| 120 | #define logipiq(name, func, arg1, arg2, sgd, dgd) \ |
| 121 | KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid) |
| 122 | #define logipiq2(name, arg) \ |
| 123 | KTR_LOG(ipiq_ ## name, arg) |
| 124 | |
| 125 | #endif /* SMP */ |
| 126 | |
| 127 | #ifdef SMP |
| 128 | |
| 129 | static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip, |
| 130 | struct intrframe *frame); |
| 131 | static void lwkt_cpusync_remote1(lwkt_cpusync_t poll); |
| 132 | static void lwkt_cpusync_remote2(lwkt_cpusync_t poll); |
| 133 | |
| 134 | /* |
| 135 | * Send a function execution request to another cpu. The request is queued |
| 136 | * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every |
| 137 | * possible target cpu. The FIFO can be written. |
| 138 | * |
| 139 | * If the FIFO fills up we have to enable interrupts to avoid an APIC |
| 140 | * deadlock and process pending IPIQs while waiting for it to empty. |
| 141 | * Otherwise we may soft-deadlock with another cpu whos FIFO is also full. |
| 142 | * |
| 143 | * We can safely bump gd_intr_nesting_level because our crit_exit() at the |
| 144 | * end will take care of any pending interrupts. |
| 145 | * |
| 146 | * The actual hardware IPI is avoided if the target cpu is already processing |
| 147 | * the queue from a prior IPI. It is possible to pipeline IPI messages |
| 148 | * very quickly between cpus due to the FIFO hysteresis. |
| 149 | * |
| 150 | * Need not be called from a critical section. |
| 151 | */ |
| 152 | int |
| 153 | lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2) |
| 154 | { |
| 155 | lwkt_ipiq_t ip; |
| 156 | int windex; |
| 157 | struct globaldata *gd = mycpu; |
| 158 | |
| 159 | logipiq(send_norm, func, arg1, arg2, gd, target); |
| 160 | |
| 161 | if (target == gd) { |
| 162 | func(arg1, arg2, NULL); |
| 163 | logipiq(send_end, func, arg1, arg2, gd, target); |
| 164 | return(0); |
| 165 | } |
| 166 | crit_enter(); |
| 167 | ++gd->gd_intr_nesting_level; |
| 168 | #ifdef INVARIANTS |
| 169 | if (gd->gd_intr_nesting_level > 20) |
| 170 | panic("lwkt_send_ipiq: TOO HEAVILY NESTED!"); |
| 171 | #endif |
| 172 | KKASSERT(curthread->td_critcount); |
| 173 | ++ipiq_count; |
| 174 | ip = &gd->gd_ipiq[target->gd_cpuid]; |
| 175 | |
| 176 | /* |
| 177 | * Do not allow the FIFO to become full. Interrupts must be physically |
| 178 | * enabled while we liveloop to avoid deadlocking the APIC. |
| 179 | */ |
| 180 | if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) { |
| 181 | #if defined(__i386__) |
| 182 | unsigned int eflags = read_eflags(); |
| 183 | #elif defined(__x86_64__) |
| 184 | unsigned long rflags = read_rflags(); |
| 185 | #endif |
| 186 | |
| 187 | if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0) { |
| 188 | logipiq(cpu_send, func, arg1, arg2, gd, target); |
| 189 | cpu_send_ipiq(target->gd_cpuid); |
| 190 | } |
| 191 | cpu_enable_intr(); |
| 192 | ++ipiq_fifofull; |
| 193 | while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) { |
| 194 | KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1); |
| 195 | lwkt_process_ipiq(); |
| 196 | } |
| 197 | #if defined(__i386__) |
| 198 | write_eflags(eflags); |
| 199 | #elif defined(__x86_64__) |
| 200 | write_rflags(rflags); |
| 201 | #endif |
| 202 | } |
| 203 | |
| 204 | /* |
| 205 | * Queue the new message |
| 206 | */ |
| 207 | windex = ip->ip_windex & MAXCPUFIFO_MASK; |
| 208 | ip->ip_func[windex] = func; |
| 209 | ip->ip_arg1[windex] = arg1; |
| 210 | ip->ip_arg2[windex] = arg2; |
| 211 | cpu_sfence(); |
| 212 | ++ip->ip_windex; |
| 213 | --gd->gd_intr_nesting_level; |
| 214 | |
| 215 | /* |
| 216 | * signal the target cpu that there is work pending. |
| 217 | */ |
| 218 | if (atomic_poll_acquire_int(&ip->ip_npoll)) { |
| 219 | logipiq(cpu_send, func, arg1, arg2, gd, target); |
| 220 | cpu_send_ipiq(target->gd_cpuid); |
| 221 | } else { |
| 222 | if (ipiq_optimized == 0) { |
| 223 | logipiq(cpu_send, func, arg1, arg2, gd, target); |
| 224 | cpu_send_ipiq(target->gd_cpuid); |
| 225 | } else { |
| 226 | ++ipiq_avoided; |
| 227 | } |
| 228 | } |
| 229 | crit_exit(); |
| 230 | |
| 231 | logipiq(send_end, func, arg1, arg2, gd, target); |
| 232 | return(ip->ip_windex); |
| 233 | } |
| 234 | |
| 235 | /* |
| 236 | * Similar to lwkt_send_ipiq() but this function does not actually initiate |
| 237 | * the IPI to the target cpu unless the FIFO has become too full, so it is |
| 238 | * very fast. |
| 239 | * |
| 240 | * This function is used for non-critical IPI messages, such as memory |
| 241 | * deallocations. The queue will typically be flushed by the target cpu at |
| 242 | * the next clock interrupt. |
| 243 | * |
| 244 | * Need not be called from a critical section. |
| 245 | */ |
| 246 | int |
| 247 | lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func, |
| 248 | void *arg1, int arg2) |
| 249 | { |
| 250 | lwkt_ipiq_t ip; |
| 251 | int windex; |
| 252 | struct globaldata *gd = mycpu; |
| 253 | |
| 254 | KKASSERT(target != gd); |
| 255 | crit_enter(); |
| 256 | logipiq(send_pasv, func, arg1, arg2, gd, target); |
| 257 | ++gd->gd_intr_nesting_level; |
| 258 | #ifdef INVARIANTS |
| 259 | if (gd->gd_intr_nesting_level > 20) |
| 260 | panic("lwkt_send_ipiq: TOO HEAVILY NESTED!"); |
| 261 | #endif |
| 262 | KKASSERT(curthread->td_critcount); |
| 263 | ++ipiq_count; |
| 264 | ++ipiq_passive; |
| 265 | ip = &gd->gd_ipiq[target->gd_cpuid]; |
| 266 | |
| 267 | /* |
| 268 | * Do not allow the FIFO to become full. Interrupts must be physically |
| 269 | * enabled while we liveloop to avoid deadlocking the APIC. |
| 270 | */ |
| 271 | if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) { |
| 272 | #if defined(__i386__) |
| 273 | unsigned int eflags = read_eflags(); |
| 274 | #elif defined(__x86_64__) |
| 275 | unsigned long rflags = read_rflags(); |
| 276 | #endif |
| 277 | |
| 278 | if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0) { |
| 279 | logipiq(cpu_send, func, arg1, arg2, gd, target); |
| 280 | cpu_send_ipiq(target->gd_cpuid); |
| 281 | } |
| 282 | cpu_enable_intr(); |
| 283 | ++ipiq_fifofull; |
| 284 | while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) { |
| 285 | KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1); |
| 286 | lwkt_process_ipiq(); |
| 287 | } |
| 288 | #if defined(__i386__) |
| 289 | write_eflags(eflags); |
| 290 | #elif defined(__x86_64__) |
| 291 | write_rflags(rflags); |
| 292 | #endif |
| 293 | } |
| 294 | |
| 295 | /* |
| 296 | * Queue the new message |
| 297 | */ |
| 298 | windex = ip->ip_windex & MAXCPUFIFO_MASK; |
| 299 | ip->ip_func[windex] = func; |
| 300 | ip->ip_arg1[windex] = arg1; |
| 301 | ip->ip_arg2[windex] = arg2; |
| 302 | cpu_sfence(); |
| 303 | ++ip->ip_windex; |
| 304 | --gd->gd_intr_nesting_level; |
| 305 | |
| 306 | /* |
| 307 | * Do not signal the target cpu, it will pick up the IPI when it next |
| 308 | * polls (typically on the next tick). |
| 309 | */ |
| 310 | crit_exit(); |
| 311 | |
| 312 | logipiq(send_end, func, arg1, arg2, gd, target); |
| 313 | return(ip->ip_windex); |
| 314 | } |
| 315 | |
| 316 | /* |
| 317 | * Send an IPI request without blocking, return 0 on success, ENOENT on |
| 318 | * failure. The actual queueing of the hardware IPI may still force us |
| 319 | * to spin and process incoming IPIs but that will eventually go away |
| 320 | * when we've gotten rid of the other general IPIs. |
| 321 | */ |
| 322 | int |
| 323 | lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func, |
| 324 | void *arg1, int arg2) |
| 325 | { |
| 326 | lwkt_ipiq_t ip; |
| 327 | int windex; |
| 328 | struct globaldata *gd = mycpu; |
| 329 | |
| 330 | logipiq(send_nbio, func, arg1, arg2, gd, target); |
| 331 | KKASSERT(curthread->td_critcount); |
| 332 | if (target == gd) { |
| 333 | func(arg1, arg2, NULL); |
| 334 | logipiq(send_end, func, arg1, arg2, gd, target); |
| 335 | return(0); |
| 336 | } |
| 337 | ++ipiq_count; |
| 338 | ip = &gd->gd_ipiq[target->gd_cpuid]; |
| 339 | |
| 340 | if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) { |
| 341 | logipiq(send_fail, func, arg1, arg2, gd, target); |
| 342 | return(ENOENT); |
| 343 | } |
| 344 | windex = ip->ip_windex & MAXCPUFIFO_MASK; |
| 345 | ip->ip_func[windex] = func; |
| 346 | ip->ip_arg1[windex] = arg1; |
| 347 | ip->ip_arg2[windex] = arg2; |
| 348 | cpu_sfence(); |
| 349 | ++ip->ip_windex; |
| 350 | |
| 351 | /* |
| 352 | * This isn't a passive IPI, we still have to signal the target cpu. |
| 353 | */ |
| 354 | if (atomic_poll_acquire_int(&ip->ip_npoll)) { |
| 355 | logipiq(cpu_send, func, arg1, arg2, gd, target); |
| 356 | cpu_send_ipiq(target->gd_cpuid); |
| 357 | } else { |
| 358 | if (ipiq_optimized == 0) { |
| 359 | logipiq(cpu_send, func, arg1, arg2, gd, target); |
| 360 | cpu_send_ipiq(target->gd_cpuid); |
| 361 | } else { |
| 362 | ++ipiq_avoided; |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | logipiq(send_end, func, arg1, arg2, gd, target); |
| 367 | return(0); |
| 368 | } |
| 369 | |
| 370 | /* |
| 371 | * deprecated, used only by fast int forwarding. |
| 372 | */ |
| 373 | int |
| 374 | lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2) |
| 375 | { |
| 376 | return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2)); |
| 377 | } |
| 378 | |
| 379 | /* |
| 380 | * Send a message to several target cpus. Typically used for scheduling. |
| 381 | * The message will not be sent to stopped cpus. |
| 382 | */ |
| 383 | int |
| 384 | lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2) |
| 385 | { |
| 386 | int cpuid; |
| 387 | int count = 0; |
| 388 | |
| 389 | mask &= ~stopped_cpus; |
| 390 | while (mask) { |
| 391 | cpuid = BSFCPUMASK(mask); |
| 392 | lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2); |
| 393 | mask &= ~CPUMASK(cpuid); |
| 394 | ++count; |
| 395 | } |
| 396 | return(count); |
| 397 | } |
| 398 | |
| 399 | /* |
| 400 | * Wait for the remote cpu to finish processing a function. |
| 401 | * |
| 402 | * YYY we have to enable interrupts and process the IPIQ while waiting |
| 403 | * for it to empty or we may deadlock with another cpu. Create a CPU_*() |
| 404 | * function to do this! YYY we really should 'block' here. |
| 405 | * |
| 406 | * MUST be called from a critical section. This routine may be called |
| 407 | * from an interrupt (for example, if an interrupt wakes a foreign thread |
| 408 | * up). |
| 409 | */ |
| 410 | void |
| 411 | lwkt_wait_ipiq(globaldata_t target, int seq) |
| 412 | { |
| 413 | lwkt_ipiq_t ip; |
| 414 | int maxc = 100000000; |
| 415 | |
| 416 | if (target != mycpu) { |
| 417 | ip = &mycpu->gd_ipiq[target->gd_cpuid]; |
| 418 | if ((int)(ip->ip_xindex - seq) < 0) { |
| 419 | #if defined(__i386__) |
| 420 | unsigned int eflags = read_eflags(); |
| 421 | #elif defined(__x86_64__) |
| 422 | unsigned long rflags = read_rflags(); |
| 423 | #endif |
| 424 | cpu_enable_intr(); |
| 425 | while ((int)(ip->ip_xindex - seq) < 0) { |
| 426 | crit_enter(); |
| 427 | lwkt_process_ipiq(); |
| 428 | crit_exit(); |
| 429 | if (--maxc == 0) |
| 430 | kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq); |
| 431 | if (maxc < -1000000) |
| 432 | panic("LWKT_WAIT_IPIQ"); |
| 433 | /* |
| 434 | * xindex may be modified by another cpu, use a load fence |
| 435 | * to ensure that the loop does not use a speculative value |
| 436 | * (which may improve performance). |
| 437 | */ |
| 438 | cpu_lfence(); |
| 439 | } |
| 440 | #if defined(__i386__) |
| 441 | write_eflags(eflags); |
| 442 | #elif defined(__x86_64__) |
| 443 | write_rflags(rflags); |
| 444 | #endif |
| 445 | } |
| 446 | } |
| 447 | } |
| 448 | |
| 449 | int |
| 450 | lwkt_seq_ipiq(globaldata_t target) |
| 451 | { |
| 452 | lwkt_ipiq_t ip; |
| 453 | |
| 454 | ip = &mycpu->gd_ipiq[target->gd_cpuid]; |
| 455 | return(ip->ip_windex); |
| 456 | } |
| 457 | |
| 458 | /* |
| 459 | * Called from IPI interrupt (like a fast interrupt), which has placed |
| 460 | * us in a critical section. The MP lock may or may not be held. |
| 461 | * May also be called from doreti or splz, or be reentrantly called |
| 462 | * indirectly through the ip_func[] we run. |
| 463 | * |
| 464 | * There are two versions, one where no interrupt frame is available (when |
| 465 | * called from the send code and from splz, and one where an interrupt |
| 466 | * frame is available. |
| 467 | */ |
| 468 | void |
| 469 | lwkt_process_ipiq(void) |
| 470 | { |
| 471 | globaldata_t gd = mycpu; |
| 472 | globaldata_t sgd; |
| 473 | lwkt_ipiq_t ip; |
| 474 | int n; |
| 475 | |
| 476 | again: |
| 477 | for (n = 0; n < ncpus; ++n) { |
| 478 | if (n != gd->gd_cpuid) { |
| 479 | sgd = globaldata_find(n); |
| 480 | ip = sgd->gd_ipiq; |
| 481 | if (ip != NULL) { |
| 482 | while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL)) |
| 483 | ; |
| 484 | } |
| 485 | } |
| 486 | } |
| 487 | if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) { |
| 488 | if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) { |
| 489 | if (gd->gd_curthread->td_cscount == 0) |
| 490 | goto again; |
| 491 | need_ipiq(); |
| 492 | } |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | void |
| 497 | lwkt_process_ipiq_frame(struct intrframe *frame) |
| 498 | { |
| 499 | globaldata_t gd = mycpu; |
| 500 | globaldata_t sgd; |
| 501 | lwkt_ipiq_t ip; |
| 502 | int n; |
| 503 | |
| 504 | again: |
| 505 | for (n = 0; n < ncpus; ++n) { |
| 506 | if (n != gd->gd_cpuid) { |
| 507 | sgd = globaldata_find(n); |
| 508 | ip = sgd->gd_ipiq; |
| 509 | if (ip != NULL) { |
| 510 | while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], frame)) |
| 511 | ; |
| 512 | } |
| 513 | } |
| 514 | } |
| 515 | if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) { |
| 516 | if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame)) { |
| 517 | if (gd->gd_curthread->td_cscount == 0) |
| 518 | goto again; |
| 519 | need_ipiq(); |
| 520 | } |
| 521 | } |
| 522 | } |
| 523 | |
| 524 | static int |
| 525 | lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip, |
| 526 | struct intrframe *frame) |
| 527 | { |
| 528 | globaldata_t mygd = mycpu; |
| 529 | int ri; |
| 530 | int wi; |
| 531 | ipifunc3_t copy_func; |
| 532 | void *copy_arg1; |
| 533 | int copy_arg2; |
| 534 | |
| 535 | /* |
| 536 | * Obtain the current write index, which is modified by a remote cpu. |
| 537 | * Issue a load fence to prevent speculative reads of e.g. data written |
| 538 | * by the other cpu prior to it updating the index. |
| 539 | */ |
| 540 | KKASSERT(curthread->td_critcount); |
| 541 | wi = ip->ip_windex; |
| 542 | cpu_lfence(); |
| 543 | ++mygd->gd_intr_nesting_level; |
| 544 | |
| 545 | /* |
| 546 | * NOTE: xindex is only updated after we are sure the function has |
| 547 | * finished execution. Beware lwkt_process_ipiq() reentrancy! |
| 548 | * The function may send an IPI which may block/drain. |
| 549 | * |
| 550 | * NOTE: Due to additional IPI operations that the callback function |
| 551 | * may make, it is possible for both rindex and windex to advance and |
| 552 | * thus for rindex to advance passed our cached windex. |
| 553 | * |
| 554 | * NOTE: A memory fence is required to prevent speculative loads prior |
| 555 | * to the loading of ip_rindex. Even though stores might be |
| 556 | * ordered, loads are probably not. |
| 557 | */ |
| 558 | while (wi - (ri = ip->ip_rindex) > 0) { |
| 559 | ri &= MAXCPUFIFO_MASK; |
| 560 | cpu_mfence(); |
| 561 | copy_func = ip->ip_func[ri]; |
| 562 | copy_arg1 = ip->ip_arg1[ri]; |
| 563 | copy_arg2 = ip->ip_arg2[ri]; |
| 564 | ++ip->ip_rindex; |
| 565 | KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) == |
| 566 | ((ri + 1) & MAXCPUFIFO_MASK)); |
| 567 | logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu); |
| 568 | copy_func(copy_arg1, copy_arg2, frame); |
| 569 | cpu_sfence(); |
| 570 | ip->ip_xindex = ip->ip_rindex; |
| 571 | |
| 572 | #ifdef PANIC_DEBUG |
| 573 | /* |
| 574 | * Simulate panics during the processing of an IPI |
| 575 | */ |
| 576 | if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) { |
| 577 | if (--panic_ipiq_count == 0) { |
| 578 | #ifdef DDB |
| 579 | Debugger("PANIC_DEBUG"); |
| 580 | #else |
| 581 | panic("PANIC_DEBUG"); |
| 582 | #endif |
| 583 | } |
| 584 | } |
| 585 | #endif |
| 586 | } |
| 587 | --mygd->gd_intr_nesting_level; |
| 588 | |
| 589 | /* |
| 590 | * Return non-zero if there are more IPI messages pending on this |
| 591 | * ipiq. ip_npoll is left set as long as possible to reduce the |
| 592 | * number of IPIs queued by the originating cpu, but must be cleared |
| 593 | * *BEFORE* checking windex. |
| 594 | */ |
| 595 | atomic_poll_release_int(&ip->ip_npoll); |
| 596 | return(wi != ip->ip_windex); |
| 597 | } |
| 598 | |
| 599 | static void |
| 600 | lwkt_sync_ipiq(void *arg) |
| 601 | { |
| 602 | cpumask_t *cpumask = arg; |
| 603 | |
| 604 | atomic_clear_cpumask(cpumask, mycpu->gd_cpumask); |
| 605 | if (*cpumask == 0) |
| 606 | wakeup(cpumask); |
| 607 | } |
| 608 | |
| 609 | void |
| 610 | lwkt_synchronize_ipiqs(const char *wmesg) |
| 611 | { |
| 612 | cpumask_t other_cpumask; |
| 613 | |
| 614 | other_cpumask = mycpu->gd_other_cpus & smp_active_mask; |
| 615 | lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq, &other_cpumask); |
| 616 | |
| 617 | while (other_cpumask != 0) { |
| 618 | tsleep_interlock(&other_cpumask, 0); |
| 619 | if (other_cpumask != 0) |
| 620 | tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0); |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | #endif |
| 625 | |
| 626 | /* |
| 627 | * CPU Synchronization Support |
| 628 | * |
| 629 | * lwkt_cpusync_simple() |
| 630 | * |
| 631 | * The function is executed synchronously before return on remote cpus. |
| 632 | * A lwkt_cpusync_t pointer is passed as an argument. The data can |
| 633 | * be accessed via arg->cs_data. |
| 634 | * |
| 635 | * XXX should I just pass the data as an argument to be consistent? |
| 636 | */ |
| 637 | |
| 638 | void |
| 639 | lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data) |
| 640 | { |
| 641 | struct lwkt_cpusync cmd; |
| 642 | |
| 643 | cmd.cs_run_func = NULL; |
| 644 | cmd.cs_fin1_func = func; |
| 645 | cmd.cs_fin2_func = NULL; |
| 646 | cmd.cs_data = data; |
| 647 | lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd); |
| 648 | if (mask & CPUMASK(mycpu->gd_cpuid)) |
| 649 | func(&cmd); |
| 650 | lwkt_cpusync_finish(&cmd); |
| 651 | } |
| 652 | |
| 653 | /* |
| 654 | * lwkt_cpusync_fastdata() |
| 655 | * |
| 656 | * The function is executed in tandem with return on remote cpus. |
| 657 | * The data is directly passed as an argument. Do not pass pointers to |
| 658 | * temporary storage as the storage might have |
| 659 | * gone poof by the time the target cpu executes |
| 660 | * the function. |
| 661 | * |
| 662 | * At the moment lwkt_cpusync is declared on the stack and we must wait |
| 663 | * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future |
| 664 | * optimization we should be able to put a counter in the globaldata |
| 665 | * structure (if it is not otherwise being used) and just poke it and |
| 666 | * return without waiting. XXX |
| 667 | */ |
| 668 | void |
| 669 | lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data) |
| 670 | { |
| 671 | struct lwkt_cpusync cmd; |
| 672 | |
| 673 | cmd.cs_run_func = NULL; |
| 674 | cmd.cs_fin1_func = NULL; |
| 675 | cmd.cs_fin2_func = func; |
| 676 | cmd.cs_data = NULL; |
| 677 | lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd); |
| 678 | if (mask & CPUMASK(mycpu->gd_cpuid)) |
| 679 | func(data); |
| 680 | lwkt_cpusync_finish(&cmd); |
| 681 | } |
| 682 | |
| 683 | /* |
| 684 | * lwkt_cpusync_start() |
| 685 | * |
| 686 | * Start synchronization with a set of target cpus, return once they are |
| 687 | * known to be in a synchronization loop. The target cpus will execute |
| 688 | * poll->cs_run_func() IN TANDEM WITH THE RETURN. |
| 689 | * |
| 690 | * XXX future: add lwkt_cpusync_start_quick() and require a call to |
| 691 | * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to |
| 692 | * potentially absorb the IPI latency doing something useful. |
| 693 | */ |
| 694 | void |
| 695 | lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll) |
| 696 | { |
| 697 | globaldata_t gd = mycpu; |
| 698 | |
| 699 | poll->cs_count = 0; |
| 700 | poll->cs_mask = mask; |
| 701 | #ifdef SMP |
| 702 | logipiq2(sync_start, mask & gd->gd_other_cpus); |
| 703 | poll->cs_maxcount = lwkt_send_ipiq_mask( |
| 704 | mask & gd->gd_other_cpus & smp_active_mask, |
| 705 | (ipifunc1_t)lwkt_cpusync_remote1, poll); |
| 706 | #endif |
| 707 | if (mask & gd->gd_cpumask) { |
| 708 | if (poll->cs_run_func) |
| 709 | poll->cs_run_func(poll); |
| 710 | } |
| 711 | #ifdef SMP |
| 712 | if (poll->cs_maxcount) { |
| 713 | ++ipiq_cscount; |
| 714 | ++gd->gd_curthread->td_cscount; |
| 715 | while (poll->cs_count != poll->cs_maxcount) { |
| 716 | crit_enter(); |
| 717 | lwkt_process_ipiq(); |
| 718 | crit_exit(); |
| 719 | } |
| 720 | } |
| 721 | #endif |
| 722 | } |
| 723 | |
| 724 | void |
| 725 | lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll) |
| 726 | { |
| 727 | globaldata_t gd = mycpu; |
| 728 | #ifdef SMP |
| 729 | int count; |
| 730 | #endif |
| 731 | |
| 732 | mask &= ~poll->cs_mask; |
| 733 | poll->cs_mask |= mask; |
| 734 | #ifdef SMP |
| 735 | logipiq2(sync_add, mask & gd->gd_other_cpus); |
| 736 | count = lwkt_send_ipiq_mask( |
| 737 | mask & gd->gd_other_cpus & smp_active_mask, |
| 738 | (ipifunc1_t)lwkt_cpusync_remote1, poll); |
| 739 | #endif |
| 740 | if (mask & gd->gd_cpumask) { |
| 741 | if (poll->cs_run_func) |
| 742 | poll->cs_run_func(poll); |
| 743 | } |
| 744 | #ifdef SMP |
| 745 | poll->cs_maxcount += count; |
| 746 | if (poll->cs_maxcount) { |
| 747 | if (poll->cs_maxcount == count) |
| 748 | ++gd->gd_curthread->td_cscount; |
| 749 | while (poll->cs_count != poll->cs_maxcount) { |
| 750 | crit_enter(); |
| 751 | lwkt_process_ipiq(); |
| 752 | crit_exit(); |
| 753 | } |
| 754 | } |
| 755 | #endif |
| 756 | } |
| 757 | |
| 758 | /* |
| 759 | * Finish synchronization with a set of target cpus. The target cpus will |
| 760 | * execute cs_fin1_func(poll) prior to this function returning, and will |
| 761 | * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN. |
| 762 | * |
| 763 | * If cs_maxcount is non-zero then we are mastering a cpusync with one or |
| 764 | * more remote cpus and must account for it in our thread structure. |
| 765 | */ |
| 766 | void |
| 767 | lwkt_cpusync_finish(lwkt_cpusync_t poll) |
| 768 | { |
| 769 | globaldata_t gd = mycpu; |
| 770 | |
| 771 | poll->cs_count = -1; |
| 772 | if (poll->cs_mask & gd->gd_cpumask) { |
| 773 | if (poll->cs_fin1_func) |
| 774 | poll->cs_fin1_func(poll); |
| 775 | if (poll->cs_fin2_func) |
| 776 | poll->cs_fin2_func(poll->cs_data); |
| 777 | } |
| 778 | #ifdef SMP |
| 779 | if (poll->cs_maxcount) { |
| 780 | while (poll->cs_count != -(poll->cs_maxcount + 1)) { |
| 781 | crit_enter(); |
| 782 | lwkt_process_ipiq(); |
| 783 | crit_exit(); |
| 784 | } |
| 785 | --gd->gd_curthread->td_cscount; |
| 786 | } |
| 787 | #endif |
| 788 | } |
| 789 | |
| 790 | #ifdef SMP |
| 791 | |
| 792 | /* |
| 793 | * helper IPI remote messaging function. |
| 794 | * |
| 795 | * Called on remote cpu when a new cpu synchronization request has been |
| 796 | * sent to us. Execute the run function and adjust cs_count, then requeue |
| 797 | * the request so we spin on it. |
| 798 | */ |
| 799 | static void |
| 800 | lwkt_cpusync_remote1(lwkt_cpusync_t poll) |
| 801 | { |
| 802 | atomic_add_int(&poll->cs_count, 1); |
| 803 | if (poll->cs_run_func) |
| 804 | poll->cs_run_func(poll); |
| 805 | lwkt_cpusync_remote2(poll); |
| 806 | } |
| 807 | |
| 808 | /* |
| 809 | * helper IPI remote messaging function. |
| 810 | * |
| 811 | * Poll for the originator telling us to finish. If it hasn't, requeue |
| 812 | * our request so we spin on it. When the originator requests that we |
| 813 | * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data) |
| 814 | * in tandem with the release. |
| 815 | */ |
| 816 | static void |
| 817 | lwkt_cpusync_remote2(lwkt_cpusync_t poll) |
| 818 | { |
| 819 | if (poll->cs_count < 0) { |
| 820 | cpusync_func2_t savef; |
| 821 | void *saved; |
| 822 | |
| 823 | if (poll->cs_fin1_func) |
| 824 | poll->cs_fin1_func(poll); |
| 825 | if (poll->cs_fin2_func) { |
| 826 | savef = poll->cs_fin2_func; |
| 827 | saved = poll->cs_data; |
| 828 | cpu_ccfence(); /* required ordering for MP operation */ |
| 829 | atomic_add_int(&poll->cs_count, -1); |
| 830 | savef(saved); |
| 831 | } else { |
| 832 | atomic_add_int(&poll->cs_count, -1); |
| 833 | } |
| 834 | } else { |
| 835 | globaldata_t gd = mycpu; |
| 836 | lwkt_ipiq_t ip; |
| 837 | int wi; |
| 838 | |
| 839 | ip = &gd->gd_cpusyncq; |
| 840 | wi = ip->ip_windex & MAXCPUFIFO_MASK; |
| 841 | ip->ip_func[wi] = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2; |
| 842 | ip->ip_arg1[wi] = poll; |
| 843 | ip->ip_arg2[wi] = 0; |
| 844 | cpu_sfence(); |
| 845 | ++ip->ip_windex; |
| 846 | } |
| 847 | } |
| 848 | |
| 849 | #endif |