tcplay(8): Fix SYNOPSIS & sync to usage()
[dragonfly.git] / sys / kern / lwkt_ipiq.c
CommitLineData
3b6b7bd1 1/*
8c10bfcf
MD
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 *
3b6b7bd1
MD
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
8c10bfcf 10 *
3b6b7bd1
MD
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
8c10bfcf
MD
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
3b6b7bd1 32 * SUCH DAMAGE.
8c10bfcf 33 *
546f2c66 34 * $DragonFly: src/sys/kern/lwkt_ipiq.c,v 1.27 2008/05/18 20:57:56 nth Exp $
3b6b7bd1
MD
35 */
36
37/*
38 * This module implements IPI message queueing and the MI portion of IPI
39 * message processing.
40 */
41
e8f15168
MD
42#include "opt_ddb.h"
43
3b6b7bd1
MD
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>
ac72c7f4 52#include <sys/ktr.h>
3b6b7bd1
MD
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>
3b6b7bd1
MD
69#include <machine/smp.h>
70#include <machine/atomic.h>
71
3b6b7bd1 72#ifdef SMP
4c9f5a7f
MD
73static __int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */
74static __int64_t ipiq_fifofull; /* number of fifo full conditions detected */
75static __int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */
76static __int64_t ipiq_passive; /* passive IPI messages */
77static __int64_t ipiq_cscount; /* number of cpu synchronizations */
78static int ipiq_optimized = 1; /* XXX temporary sysctl */
d5b2d319 79static int ipiq_debug; /* set to 1 for debug */
e8f15168
MD
80#ifdef PANIC_DEBUG
81static int panic_ipiq_cpu = -1;
82static int panic_ipiq_count = 100;
83#endif
3b6b7bd1
MD
84#endif
85
3b6b7bd1 86#ifdef SMP
0c52fa62
SG
87SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0,
88 "Number of IPI's sent");
89SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0,
90 "Number of fifo full conditions detected");
91SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0,
92 "Number of IPI's avoided by interlock with target cpu");
93SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0,
94 "Number of passive IPI messages sent");
95SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0,
96 "Number of cpu synchronizations");
97SYSCTL_INT(_lwkt, OID_AUTO, ipiq_optimized, CTLFLAG_RW, &ipiq_optimized, 0,
98 "");
d5b2d319
MD
99SYSCTL_INT(_lwkt, OID_AUTO, ipiq_debug, CTLFLAG_RW, &ipiq_debug, 0,
100 "");
e8f15168
MD
101#ifdef PANIC_DEBUG
102SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, "");
103SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, "");
104#endif
3b6b7bd1 105
a7adb95a 106#define IPIQ_STRING "func=%p arg1=%p arg2=%d scpu=%d dcpu=%d"
5118bbc4 107#define IPIQ_ARG_SIZE (sizeof(void *) * 2 + sizeof(int) * 3)
ac72c7f4
MD
108
109#if !defined(KTR_IPIQ)
110#define KTR_IPIQ KTR_ALL
3b6b7bd1 111#endif
ac72c7f4
MD
112KTR_INFO_MASTER(ipiq);
113KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARG_SIZE);
114KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARG_SIZE);
115KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARG_SIZE);
116KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARG_SIZE);
117KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARG_SIZE);
d7ed9e5e 118KTR_INFO(KTR_IPIQ, ipiq, sync_start, 5, "cpumask=%08x", sizeof(cpumask_t));
d5b2d319 119KTR_INFO(KTR_IPIQ, ipiq, sync_end, 6, "cpumask=%08x", sizeof(cpumask_t));
866b61fb 120KTR_INFO(KTR_IPIQ, ipiq, cpu_send, 7, IPIQ_STRING, IPIQ_ARG_SIZE);
c92e86f1 121KTR_INFO(KTR_IPIQ, ipiq, send_end, 8, IPIQ_STRING, IPIQ_ARG_SIZE);
ac72c7f4 122
a7adb95a
SZ
123#define logipiq(name, func, arg1, arg2, sgd, dgd) \
124 KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid)
d7ed9e5e
MD
125#define logipiq2(name, arg) \
126 KTR_LOG(ipiq_ ## name, arg)
ac72c7f4
MD
127
128#endif /* SMP */
3b6b7bd1
MD
129
130#ifdef SMP
131
b8a98473
MD
132static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
133 struct intrframe *frame);
d5b2d319
MD
134static void lwkt_cpusync_remote1(lwkt_cpusync_t cs);
135static void lwkt_cpusync_remote2(lwkt_cpusync_t cs);
3b6b7bd1
MD
136
137/*
138 * Send a function execution request to another cpu. The request is queued
139 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
140 * possible target cpu. The FIFO can be written.
141 *
4c9f5a7f
MD
142 * If the FIFO fills up we have to enable interrupts to avoid an APIC
143 * deadlock and process pending IPIQs while waiting for it to empty.
144 * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
3b6b7bd1
MD
145 *
146 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
147 * end will take care of any pending interrupts.
148 *
4c9f5a7f
MD
149 * The actual hardware IPI is avoided if the target cpu is already processing
150 * the queue from a prior IPI. It is possible to pipeline IPI messages
151 * very quickly between cpus due to the FIFO hysteresis.
152 *
153 * Need not be called from a critical section.
3b6b7bd1
MD
154 */
155int
b8a98473 156lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2)
3b6b7bd1
MD
157{
158 lwkt_ipiq_t ip;
159 int windex;
160 struct globaldata *gd = mycpu;
161
a7adb95a 162 logipiq(send_norm, func, arg1, arg2, gd, target);
ac72c7f4 163
3b6b7bd1 164 if (target == gd) {
b8a98473 165 func(arg1, arg2, NULL);
c92e86f1 166 logipiq(send_end, func, arg1, arg2, gd, target);
3b6b7bd1
MD
167 return(0);
168 }
169 crit_enter();
170 ++gd->gd_intr_nesting_level;
171#ifdef INVARIANTS
172 if (gd->gd_intr_nesting_level > 20)
173 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
174#endif
f9235b6d 175 KKASSERT(curthread->td_critcount);
3b6b7bd1
MD
176 ++ipiq_count;
177 ip = &gd->gd_ipiq[target->gd_cpuid];
178
179 /*
4c9f5a7f
MD
180 * Do not allow the FIFO to become full. Interrupts must be physically
181 * enabled while we liveloop to avoid deadlocking the APIC.
182 */
183 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
46d4e165 184#if defined(__i386__)
4c9f5a7f 185 unsigned int eflags = read_eflags();
b2b3ffcd 186#elif defined(__x86_64__)
46d4e165
JG
187 unsigned long rflags = read_rflags();
188#endif
4c9f5a7f 189
866b61fb
SZ
190 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0) {
191 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 192 cpu_send_ipiq(target->gd_cpuid);
866b61fb 193 }
4c9f5a7f
MD
194 cpu_enable_intr();
195 ++ipiq_fifofull;
cfaeae2a 196 DEBUG_PUSH_INFO("send_ipiq3");
4c9f5a7f
MD
197 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
198 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
199 lwkt_process_ipiq();
200 }
cfaeae2a 201 DEBUG_POP_INFO();
46d4e165 202#if defined(__i386__)
4c9f5a7f 203 write_eflags(eflags);
b2b3ffcd 204#elif defined(__x86_64__)
46d4e165
JG
205 write_rflags(rflags);
206#endif
4c9f5a7f
MD
207 }
208
209 /*
210 * Queue the new message
3b6b7bd1 211 */
3b6b7bd1 212 windex = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473
MD
213 ip->ip_func[windex] = func;
214 ip->ip_arg1[windex] = arg1;
215 ip->ip_arg2[windex] = arg2;
35238fa5 216 cpu_sfence();
3b6b7bd1 217 ++ip->ip_windex;
4c9f5a7f
MD
218 --gd->gd_intr_nesting_level;
219
220 /*
221 * signal the target cpu that there is work pending.
222 */
223 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
866b61fb 224 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f
MD
225 cpu_send_ipiq(target->gd_cpuid);
226 } else {
866b61fb
SZ
227 if (ipiq_optimized == 0) {
228 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 229 cpu_send_ipiq(target->gd_cpuid);
866b61fb
SZ
230 } else {
231 ++ipiq_avoided;
232 }
4c9f5a7f
MD
233 }
234 crit_exit();
c92e86f1
SZ
235
236 logipiq(send_end, func, arg1, arg2, gd, target);
4c9f5a7f
MD
237 return(ip->ip_windex);
238}
239
240/*
241 * Similar to lwkt_send_ipiq() but this function does not actually initiate
242 * the IPI to the target cpu unless the FIFO has become too full, so it is
243 * very fast.
244 *
245 * This function is used for non-critical IPI messages, such as memory
246 * deallocations. The queue will typically be flushed by the target cpu at
247 * the next clock interrupt.
248 *
249 * Need not be called from a critical section.
250 */
251int
b8a98473
MD
252lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
253 void *arg1, int arg2)
4c9f5a7f
MD
254{
255 lwkt_ipiq_t ip;
256 int windex;
257 struct globaldata *gd = mycpu;
258
259 KKASSERT(target != gd);
260 crit_enter();
a7adb95a 261 logipiq(send_pasv, func, arg1, arg2, gd, target);
4c9f5a7f
MD
262 ++gd->gd_intr_nesting_level;
263#ifdef INVARIANTS
264 if (gd->gd_intr_nesting_level > 20)
265 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
266#endif
f9235b6d 267 KKASSERT(curthread->td_critcount);
4c9f5a7f
MD
268 ++ipiq_count;
269 ++ipiq_passive;
270 ip = &gd->gd_ipiq[target->gd_cpuid];
271
272 /*
273 * Do not allow the FIFO to become full. Interrupts must be physically
274 * enabled while we liveloop to avoid deadlocking the APIC.
275 */
3b6b7bd1 276 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
46d4e165 277#if defined(__i386__)
3b6b7bd1 278 unsigned int eflags = read_eflags();
b2b3ffcd 279#elif defined(__x86_64__)
46d4e165
JG
280 unsigned long rflags = read_rflags();
281#endif
4c9f5a7f 282
866b61fb
SZ
283 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0) {
284 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 285 cpu_send_ipiq(target->gd_cpuid);
866b61fb 286 }
3b6b7bd1
MD
287 cpu_enable_intr();
288 ++ipiq_fifofull;
cfaeae2a 289 DEBUG_PUSH_INFO("send_ipiq3_passive");
3b6b7bd1
MD
290 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
291 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
292 lwkt_process_ipiq();
293 }
cfaeae2a 294 DEBUG_POP_INFO();
46d4e165 295#if defined(__i386__)
3b6b7bd1 296 write_eflags(eflags);
b2b3ffcd 297#elif defined(__x86_64__)
46d4e165
JG
298 write_rflags(rflags);
299#endif
3b6b7bd1 300 }
4c9f5a7f
MD
301
302 /*
303 * Queue the new message
304 */
305 windex = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473
MD
306 ip->ip_func[windex] = func;
307 ip->ip_arg1[windex] = arg1;
308 ip->ip_arg2[windex] = arg2;
35238fa5 309 cpu_sfence();
4c9f5a7f 310 ++ip->ip_windex;
3b6b7bd1 311 --gd->gd_intr_nesting_level;
4c9f5a7f
MD
312
313 /*
314 * Do not signal the target cpu, it will pick up the IPI when it next
315 * polls (typically on the next tick).
316 */
3b6b7bd1 317 crit_exit();
c92e86f1
SZ
318
319 logipiq(send_end, func, arg1, arg2, gd, target);
3b6b7bd1
MD
320 return(ip->ip_windex);
321}
322
41a01a4d 323/*
4c9f5a7f
MD
324 * Send an IPI request without blocking, return 0 on success, ENOENT on
325 * failure. The actual queueing of the hardware IPI may still force us
326 * to spin and process incoming IPIs but that will eventually go away
327 * when we've gotten rid of the other general IPIs.
41a01a4d
MD
328 */
329int
b8a98473
MD
330lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func,
331 void *arg1, int arg2)
41a01a4d
MD
332{
333 lwkt_ipiq_t ip;
334 int windex;
335 struct globaldata *gd = mycpu;
336
a7adb95a 337 logipiq(send_nbio, func, arg1, arg2, gd, target);
f9235b6d 338 KKASSERT(curthread->td_critcount);
41a01a4d 339 if (target == gd) {
b8a98473 340 func(arg1, arg2, NULL);
c92e86f1 341 logipiq(send_end, func, arg1, arg2, gd, target);
41a01a4d
MD
342 return(0);
343 }
344 ++ipiq_count;
345 ip = &gd->gd_ipiq[target->gd_cpuid];
346
ac72c7f4 347 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
a7adb95a 348 logipiq(send_fail, func, arg1, arg2, gd, target);
41a01a4d 349 return(ENOENT);
ac72c7f4 350 }
41a01a4d 351 windex = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473
MD
352 ip->ip_func[windex] = func;
353 ip->ip_arg1[windex] = arg1;
354 ip->ip_arg2[windex] = arg2;
35238fa5 355 cpu_sfence();
41a01a4d 356 ++ip->ip_windex;
4c9f5a7f 357
41a01a4d 358 /*
4c9f5a7f 359 * This isn't a passive IPI, we still have to signal the target cpu.
41a01a4d 360 */
4c9f5a7f 361 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
866b61fb 362 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f
MD
363 cpu_send_ipiq(target->gd_cpuid);
364 } else {
866b61fb
SZ
365 if (ipiq_optimized == 0) {
366 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 367 cpu_send_ipiq(target->gd_cpuid);
866b61fb 368 } else {
728f6208 369 ++ipiq_avoided;
866b61fb 370 }
4c9f5a7f 371 }
c92e86f1
SZ
372
373 logipiq(send_end, func, arg1, arg2, gd, target);
41a01a4d
MD
374 return(0);
375}
376
3b6b7bd1
MD
377/*
378 * deprecated, used only by fast int forwarding.
379 */
380int
b8a98473 381lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
3b6b7bd1 382{
b8a98473 383 return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
3b6b7bd1
MD
384}
385
386/*
387 * Send a message to several target cpus. Typically used for scheduling.
388 * The message will not be sent to stopped cpus.
389 */
390int
da23a592 391lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2)
3b6b7bd1
MD
392{
393 int cpuid;
394 int count = 0;
395
396 mask &= ~stopped_cpus;
397 while (mask) {
da23a592 398 cpuid = BSFCPUMASK(mask);
b8a98473 399 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
da23a592 400 mask &= ~CPUMASK(cpuid);
3b6b7bd1
MD
401 ++count;
402 }
403 return(count);
404}
405
406/*
407 * Wait for the remote cpu to finish processing a function.
408 *
409 * YYY we have to enable interrupts and process the IPIQ while waiting
410 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
411 * function to do this! YYY we really should 'block' here.
412 *
413 * MUST be called from a critical section. This routine may be called
414 * from an interrupt (for example, if an interrupt wakes a foreign thread
415 * up).
416 */
417void
418lwkt_wait_ipiq(globaldata_t target, int seq)
419{
420 lwkt_ipiq_t ip;
421 int maxc = 100000000;
422
423 if (target != mycpu) {
424 ip = &mycpu->gd_ipiq[target->gd_cpuid];
425 if ((int)(ip->ip_xindex - seq) < 0) {
46d4e165 426#if defined(__i386__)
3b6b7bd1 427 unsigned int eflags = read_eflags();
b2b3ffcd 428#elif defined(__x86_64__)
46d4e165
JG
429 unsigned long rflags = read_rflags();
430#endif
3b6b7bd1 431 cpu_enable_intr();
cfaeae2a 432 DEBUG_PUSH_INFO("wait_ipiq");
3b6b7bd1 433 while ((int)(ip->ip_xindex - seq) < 0) {
41a01a4d 434 crit_enter();
3b6b7bd1 435 lwkt_process_ipiq();
41a01a4d 436 crit_exit();
3b6b7bd1 437 if (--maxc == 0)
6ea70f76 438 kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
3b6b7bd1
MD
439 if (maxc < -1000000)
440 panic("LWKT_WAIT_IPIQ");
35238fa5
MD
441 /*
442 * xindex may be modified by another cpu, use a load fence
443 * to ensure that the loop does not use a speculative value
444 * (which may improve performance).
445 */
446 cpu_lfence();
3b6b7bd1 447 }
cfaeae2a 448 DEBUG_POP_INFO();
46d4e165 449#if defined(__i386__)
3b6b7bd1 450 write_eflags(eflags);
b2b3ffcd 451#elif defined(__x86_64__)
46d4e165
JG
452 write_rflags(rflags);
453#endif
3b6b7bd1
MD
454 }
455 }
456}
457
41a01a4d
MD
458int
459lwkt_seq_ipiq(globaldata_t target)
460{
461 lwkt_ipiq_t ip;
462
463 ip = &mycpu->gd_ipiq[target->gd_cpuid];
464 return(ip->ip_windex);
465}
466
3b6b7bd1
MD
467/*
468 * Called from IPI interrupt (like a fast interrupt), which has placed
469 * us in a critical section. The MP lock may or may not be held.
470 * May also be called from doreti or splz, or be reentrantly called
471 * indirectly through the ip_func[] we run.
472 *
473 * There are two versions, one where no interrupt frame is available (when
474 * called from the send code and from splz, and one where an interrupt
475 * frame is available.
d5b2d319
MD
476 *
477 * When the current cpu is mastering a cpusync we do NOT internally loop
478 * on the cpusyncq poll. We also do not re-flag a pending ipi due to
479 * the cpusyncq poll because this can cause doreti/splz to loop internally.
480 * The cpusync master's own loop must be allowed to run to avoid a deadlock.
3b6b7bd1
MD
481 */
482void
483lwkt_process_ipiq(void)
484{
485 globaldata_t gd = mycpu;
ac72c7f4 486 globaldata_t sgd;
3b6b7bd1
MD
487 lwkt_ipiq_t ip;
488 int n;
489
490again:
491 for (n = 0; n < ncpus; ++n) {
492 if (n != gd->gd_cpuid) {
ac72c7f4
MD
493 sgd = globaldata_find(n);
494 ip = sgd->gd_ipiq;
3b6b7bd1 495 if (ip != NULL) {
b8a98473 496 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL))
3b6b7bd1
MD
497 ;
498 }
499 }
500 }
501 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
b8a98473 502 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) {
0f7a3396
MD
503 if (gd->gd_curthread->td_cscount == 0)
504 goto again;
0f7a3396 505 }
3b6b7bd1
MD
506 }
507}
508
3b6b7bd1 509void
c7eb0589 510lwkt_process_ipiq_frame(struct intrframe *frame)
3b6b7bd1
MD
511{
512 globaldata_t gd = mycpu;
ac72c7f4 513 globaldata_t sgd;
3b6b7bd1
MD
514 lwkt_ipiq_t ip;
515 int n;
516
517again:
518 for (n = 0; n < ncpus; ++n) {
519 if (n != gd->gd_cpuid) {
ac72c7f4
MD
520 sgd = globaldata_find(n);
521 ip = sgd->gd_ipiq;
3b6b7bd1 522 if (ip != NULL) {
c7eb0589 523 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], frame))
3b6b7bd1
MD
524 ;
525 }
526 }
527 }
528 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
c7eb0589 529 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame)) {
0f7a3396
MD
530 if (gd->gd_curthread->td_cscount == 0)
531 goto again;
0f7a3396 532 }
3b6b7bd1
MD
533 }
534}
3b6b7bd1 535
cfaeae2a
MD
536#if 0
537static int iqticks[SMP_MAXCPU];
538static int iqcount[SMP_MAXCPU];
539#endif
540#if 0
541static int iqterm[SMP_MAXCPU];
542#endif
543
3b6b7bd1 544static int
b8a98473
MD
545lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
546 struct intrframe *frame)
3b6b7bd1 547{
2de4f77e 548 globaldata_t mygd = mycpu;
3b6b7bd1 549 int ri;
35238fa5 550 int wi;
b8a98473
MD
551 ipifunc3_t copy_func;
552 void *copy_arg1;
553 int copy_arg2;
35238fa5 554
cfaeae2a
MD
555#if 0
556 if (iqticks[mygd->gd_cpuid] != ticks) {
557 iqticks[mygd->gd_cpuid] = ticks;
558 iqcount[mygd->gd_cpuid] = 0;
559 }
560 if (++iqcount[mygd->gd_cpuid] > 3000000) {
561 kprintf("cpu %d ipiq maxed cscount %d spin %d\n",
562 mygd->gd_cpuid,
563 mygd->gd_curthread->td_cscount,
564 mygd->gd_spinlocks_wr);
565 iqcount[mygd->gd_cpuid] = 0;
566#if 0
567 if (++iqterm[mygd->gd_cpuid] > 10)
568 panic("cpu %d ipiq maxed", mygd->gd_cpuid);
569#endif
570 int i;
571 for (i = 0; i < ncpus; ++i) {
572 if (globaldata_find(i)->gd_infomsg)
573 kprintf(" %s", globaldata_find(i)->gd_infomsg);
574 }
575 kprintf("\n");
576 }
577#endif
578
35238fa5
MD
579 /*
580 * Obtain the current write index, which is modified by a remote cpu.
581 * Issue a load fence to prevent speculative reads of e.g. data written
582 * by the other cpu prior to it updating the index.
583 */
f9235b6d 584 KKASSERT(curthread->td_critcount);
35238fa5
MD
585 wi = ip->ip_windex;
586 cpu_lfence();
2de4f77e 587 ++mygd->gd_intr_nesting_level;
35238fa5 588
3b6b7bd1 589 /*
562273ea
MD
590 * NOTE: xindex is only updated after we are sure the function has
591 * finished execution. Beware lwkt_process_ipiq() reentrancy!
592 * The function may send an IPI which may block/drain.
d64a7617 593 *
562273ea
MD
594 * NOTE: Due to additional IPI operations that the callback function
595 * may make, it is possible for both rindex and windex to advance and
596 * thus for rindex to advance passed our cached windex.
597 *
d5b2d319 598 * NOTE: A load fence is required to prevent speculative loads prior
562273ea 599 * to the loading of ip_rindex. Even though stores might be
d5b2d319
MD
600 * ordered, loads are probably not. A memory fence is required
601 * to prevent reordering of the loads after the ip_rindex update.
3b6b7bd1 602 */
d64a7617 603 while (wi - (ri = ip->ip_rindex) > 0) {
3b6b7bd1 604 ri &= MAXCPUFIFO_MASK;
d5b2d319 605 cpu_lfence();
728f6208 606 copy_func = ip->ip_func[ri];
b8a98473
MD
607 copy_arg1 = ip->ip_arg1[ri];
608 copy_arg2 = ip->ip_arg2[ri];
d5b2d319 609 cpu_mfence();
728f6208 610 ++ip->ip_rindex;
562273ea
MD
611 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) ==
612 ((ri + 1) & MAXCPUFIFO_MASK));
a7adb95a 613 logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
d5b2d319
MD
614#ifdef INVARIANTS
615 if (ipiq_debug && (ip->ip_rindex & 0xFFFFFF) == 0) {
616 kprintf("cpu %d ipifunc %p %p %d (frame %p)\n",
617 mycpu->gd_cpuid,
618 copy_func, copy_arg1, copy_arg2,
619#if defined(__i386__)
620 (frame ? (void *)frame->if_eip : NULL));
621#elif defined(__amd64__)
622 (frame ? (void *)frame->if_rip : NULL));
623#else
624 NULL);
625#endif
626 }
627#endif
b8a98473 628 copy_func(copy_arg1, copy_arg2, frame);
35238fa5 629 cpu_sfence();
3b6b7bd1 630 ip->ip_xindex = ip->ip_rindex;
e8f15168
MD
631
632#ifdef PANIC_DEBUG
633 /*
634 * Simulate panics during the processing of an IPI
635 */
636 if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
637 if (--panic_ipiq_count == 0) {
638#ifdef DDB
639 Debugger("PANIC_DEBUG");
640#else
641 panic("PANIC_DEBUG");
642#endif
643 }
644 }
645#endif
3b6b7bd1 646 }
2de4f77e 647 --mygd->gd_intr_nesting_level;
4c9f5a7f
MD
648
649 /*
650 * Return non-zero if there are more IPI messages pending on this
651 * ipiq. ip_npoll is left set as long as possible to reduce the
652 * number of IPIs queued by the originating cpu, but must be cleared
653 * *BEFORE* checking windex.
654 */
655 atomic_poll_release_int(&ip->ip_npoll);
3b6b7bd1
MD
656 return(wi != ip->ip_windex);
657}
658
6c92c1f2
SZ
659static void
660lwkt_sync_ipiq(void *arg)
661{
5a1a2253 662 volatile cpumask_t *cpumask = arg;
6c92c1f2 663
da23a592 664 atomic_clear_cpumask(cpumask, mycpu->gd_cpumask);
6c92c1f2
SZ
665 if (*cpumask == 0)
666 wakeup(cpumask);
667}
668
669void
670lwkt_synchronize_ipiqs(const char *wmesg)
671{
5a1a2253 672 volatile cpumask_t other_cpumask;
6c92c1f2
SZ
673
674 other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
5a1a2253
SZ
675 lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq,
676 __DEVOLATILE(void *, &other_cpumask));
6c92c1f2 677
6c92c1f2 678 while (other_cpumask != 0) {
ae8e83e6 679 tsleep_interlock(&other_cpumask, 0);
6c92c1f2 680 if (other_cpumask != 0)
d9345d3a 681 tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0);
6c92c1f2 682 }
6c92c1f2
SZ
683}
684
0f7a3396
MD
685#endif
686
3b6b7bd1
MD
687/*
688 * CPU Synchronization Support
5c71a36a 689 *
d5b2d319
MD
690 * lwkt_cpusync_interlock() - Place specified cpus in a quiescent state.
691 * The current cpu is placed in a hard critical
692 * section.
5c71a36a 693 *
d5b2d319
MD
694 * lwkt_cpusync_deinterlock() - Execute cs_func on specified cpus, including
695 * current cpu if specified, then return.
3b6b7bd1 696 */
3b6b7bd1 697void
d5b2d319 698lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *arg)
5c71a36a 699{
d5b2d319 700 struct lwkt_cpusync cs;
5c71a36a 701
d5b2d319
MD
702 lwkt_cpusync_init(&cs, mask, func, arg);
703 lwkt_cpusync_interlock(&cs);
704 lwkt_cpusync_deinterlock(&cs);
3b6b7bd1
MD
705}
706
d5b2d319 707
5c71a36a 708void
d5b2d319 709lwkt_cpusync_interlock(lwkt_cpusync_t cs)
3b6b7bd1 710{
d5b2d319 711#ifdef SMP
0f7a3396 712 globaldata_t gd = mycpu;
d5b2d319 713 cpumask_t mask;
0f7a3396 714
d5b2d319
MD
715 /*
716 * mask acknowledge (cs_mack): 0->mask for stage 1
717 *
718 * mack does not include the current cpu.
719 */
720 mask = cs->cs_mask & gd->gd_other_cpus & smp_active_mask;
721 cs->cs_mack = 0;
722 crit_enter_id("cpusync");
723 if (mask) {
cfaeae2a 724 DEBUG_PUSH_INFO("cpusync_interlock");
0f7a3396
MD
725 ++ipiq_cscount;
726 ++gd->gd_curthread->td_cscount;
d5b2d319
MD
727 lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote1, cs);
728 logipiq2(sync_start, mask);
729 while (cs->cs_mack != mask) {
0f7a3396 730 lwkt_process_ipiq();
d5b2d319 731 cpu_pause();
0f7a3396 732 }
cfaeae2a 733 DEBUG_POP_INFO();
3b6b7bd1 734 }
d5b2d319
MD
735#else
736 cs->cs_mack = 0;
0f7a3396 737#endif
3b6b7bd1
MD
738}
739
740/*
d5b2d319
MD
741 * Interlocked cpus have executed remote1 and are polling in remote2.
742 * To deinterlock we clear cs_mack and wait for the cpus to execute
743 * the func and set their bit in cs_mack again.
0f7a3396 744 *
3b6b7bd1
MD
745 */
746void
d5b2d319 747lwkt_cpusync_deinterlock(lwkt_cpusync_t cs)
3b6b7bd1 748{
0f7a3396 749 globaldata_t gd = mycpu;
0f7a3396 750#ifdef SMP
d5b2d319
MD
751 cpumask_t mask;
752
753 /*
754 * mask acknowledge (cs_mack): mack->0->mack for stage 2
755 *
756 * Clearing cpu bits for polling cpus in cs_mack will cause them to
757 * execute stage 2, which executes the cs_func(cs_data) and then sets
758 * their bit in cs_mack again.
759 *
760 * mack does not include the current cpu.
761 */
762 mask = cs->cs_mack;
763 cpu_ccfence();
764 cs->cs_mack = 0;
765 if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
766 cs->cs_func(cs->cs_data);
767 if (mask) {
cfaeae2a 768 DEBUG_PUSH_INFO("cpusync_deinterlock");
d5b2d319 769 while (cs->cs_mack != mask) {
0f7a3396 770 lwkt_process_ipiq();
d5b2d319 771 cpu_pause();
0f7a3396 772 }
cfaeae2a
MD
773 DEBUG_POP_INFO();
774 /*
775 * cpusyncq ipis may be left queued without the RQF flag set due to
776 * a non-zero td_cscount, so be sure to process any laggards after
777 * decrementing td_cscount.
778 */
0f7a3396 779 --gd->gd_curthread->td_cscount;
d5b2d319
MD
780 lwkt_process_ipiq();
781 logipiq2(sync_end, mask);
3b6b7bd1 782 }
d5b2d319
MD
783 crit_exit_id("cpusync");
784#else
785 if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
786 cs->cs_func(cs->cs_data);
0f7a3396 787#endif
3b6b7bd1
MD
788}
789
0f7a3396
MD
790#ifdef SMP
791
3b6b7bd1
MD
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 */
799static void
d5b2d319 800lwkt_cpusync_remote1(lwkt_cpusync_t cs)
3b6b7bd1 801{
d5b2d319
MD
802 globaldata_t gd = mycpu;
803
804 atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
805 lwkt_cpusync_remote2(cs);
3b6b7bd1
MD
806}
807
808/*
809 * helper IPI remote messaging function.
810 *
811 * Poll for the originator telling us to finish. If it hasn't, requeue
d5b2d319 812 * our request so we spin on it.
3b6b7bd1
MD
813 */
814static void
d5b2d319 815lwkt_cpusync_remote2(lwkt_cpusync_t cs)
3b6b7bd1 816{
d5b2d319
MD
817 globaldata_t gd = mycpu;
818
819 if ((cs->cs_mack & gd->gd_cpumask) == 0) {
820 if (cs->cs_func)
821 cs->cs_func(cs->cs_data);
822 atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
3b6b7bd1 823 } else {
3b6b7bd1
MD
824 lwkt_ipiq_t ip;
825 int wi;
826
827 ip = &gd->gd_cpusyncq;
828 wi = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473 829 ip->ip_func[wi] = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
d5b2d319 830 ip->ip_arg1[wi] = cs;
b8a98473 831 ip->ip_arg2[wi] = 0;
35238fa5 832 cpu_sfence();
3b6b7bd1 833 ++ip->ip_windex;
37494a7a 834 if (ipiq_debug && (ip->ip_windex & 0xFFFFFF) == 0) {
cfaeae2a
MD
835 kprintf("cpu %d cm=%016jx %016jx f=%p\n",
836 gd->gd_cpuid,
837 (intmax_t)cs->cs_mask, (intmax_t)cs->cs_mack,
838 cs->cs_func);
37494a7a 839 }
3b6b7bd1
MD
840 }
841}
842
3b6b7bd1 843#endif