vkernel - do not set SA_NODEFER for SIGIO and SIGUSR1
[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.
1b1e83e2
MD
182 *
183 * The target ipiq may have gotten filled up due to passive IPIs and thus
184 * not be aware that its queue is too full, so be sure to issue an
185 * ipiq interrupt to the target cpu.
4c9f5a7f
MD
186 */
187 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
46d4e165 188#if defined(__i386__)
4c9f5a7f 189 unsigned int eflags = read_eflags();
b2b3ffcd 190#elif defined(__x86_64__)
46d4e165
JG
191 unsigned long rflags = read_rflags();
192#endif
4c9f5a7f 193
866b61fb
SZ
194 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0) {
195 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 196 cpu_send_ipiq(target->gd_cpuid);
866b61fb 197 }
4c9f5a7f
MD
198 cpu_enable_intr();
199 ++ipiq_fifofull;
1b1e83e2 200 cpu_send_ipiq(target->gd_cpuid);
cfaeae2a 201 DEBUG_PUSH_INFO("send_ipiq3");
4c9f5a7f
MD
202 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
203 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
204 lwkt_process_ipiq();
205 }
cfaeae2a 206 DEBUG_POP_INFO();
46d4e165 207#if defined(__i386__)
4c9f5a7f 208 write_eflags(eflags);
b2b3ffcd 209#elif defined(__x86_64__)
46d4e165
JG
210 write_rflags(rflags);
211#endif
4c9f5a7f
MD
212 }
213
214 /*
215 * Queue the new message
3b6b7bd1 216 */
3b6b7bd1 217 windex = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473
MD
218 ip->ip_func[windex] = func;
219 ip->ip_arg1[windex] = arg1;
220 ip->ip_arg2[windex] = arg2;
35238fa5 221 cpu_sfence();
3b6b7bd1 222 ++ip->ip_windex;
4c9f5a7f
MD
223 --gd->gd_intr_nesting_level;
224
225 /*
226 * signal the target cpu that there is work pending.
227 */
228 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
866b61fb 229 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f
MD
230 cpu_send_ipiq(target->gd_cpuid);
231 } else {
866b61fb
SZ
232 if (ipiq_optimized == 0) {
233 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 234 cpu_send_ipiq(target->gd_cpuid);
866b61fb
SZ
235 } else {
236 ++ipiq_avoided;
237 }
4c9f5a7f
MD
238 }
239 crit_exit();
c92e86f1
SZ
240
241 logipiq(send_end, func, arg1, arg2, gd, target);
4c9f5a7f
MD
242 return(ip->ip_windex);
243}
244
245/*
246 * Similar to lwkt_send_ipiq() but this function does not actually initiate
247 * the IPI to the target cpu unless the FIFO has become too full, so it is
248 * very fast.
249 *
250 * This function is used for non-critical IPI messages, such as memory
251 * deallocations. The queue will typically be flushed by the target cpu at
252 * the next clock interrupt.
253 *
254 * Need not be called from a critical section.
255 */
256int
b8a98473
MD
257lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
258 void *arg1, int arg2)
4c9f5a7f
MD
259{
260 lwkt_ipiq_t ip;
261 int windex;
262 struct globaldata *gd = mycpu;
263
264 KKASSERT(target != gd);
265 crit_enter();
a7adb95a 266 logipiq(send_pasv, func, arg1, arg2, gd, target);
4c9f5a7f
MD
267 ++gd->gd_intr_nesting_level;
268#ifdef INVARIANTS
269 if (gd->gd_intr_nesting_level > 20)
270 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
271#endif
f9235b6d 272 KKASSERT(curthread->td_critcount);
4c9f5a7f
MD
273 ++ipiq_count;
274 ++ipiq_passive;
275 ip = &gd->gd_ipiq[target->gd_cpuid];
276
277 /*
278 * Do not allow the FIFO to become full. Interrupts must be physically
279 * enabled while we liveloop to avoid deadlocking the APIC.
280 */
3b6b7bd1 281 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
46d4e165 282#if defined(__i386__)
3b6b7bd1 283 unsigned int eflags = read_eflags();
b2b3ffcd 284#elif defined(__x86_64__)
46d4e165
JG
285 unsigned long rflags = read_rflags();
286#endif
4c9f5a7f 287
866b61fb
SZ
288 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0) {
289 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 290 cpu_send_ipiq(target->gd_cpuid);
866b61fb 291 }
3b6b7bd1
MD
292 cpu_enable_intr();
293 ++ipiq_fifofull;
1b1e83e2 294 cpu_send_ipiq(target->gd_cpuid);
cfaeae2a 295 DEBUG_PUSH_INFO("send_ipiq3_passive");
3b6b7bd1
MD
296 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
297 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
298 lwkt_process_ipiq();
299 }
cfaeae2a 300 DEBUG_POP_INFO();
46d4e165 301#if defined(__i386__)
3b6b7bd1 302 write_eflags(eflags);
b2b3ffcd 303#elif defined(__x86_64__)
46d4e165
JG
304 write_rflags(rflags);
305#endif
3b6b7bd1 306 }
4c9f5a7f
MD
307
308 /*
309 * Queue the new message
310 */
311 windex = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473
MD
312 ip->ip_func[windex] = func;
313 ip->ip_arg1[windex] = arg1;
314 ip->ip_arg2[windex] = arg2;
35238fa5 315 cpu_sfence();
4c9f5a7f 316 ++ip->ip_windex;
3b6b7bd1 317 --gd->gd_intr_nesting_level;
4c9f5a7f
MD
318
319 /*
320 * Do not signal the target cpu, it will pick up the IPI when it next
321 * polls (typically on the next tick).
322 */
3b6b7bd1 323 crit_exit();
c92e86f1
SZ
324
325 logipiq(send_end, func, arg1, arg2, gd, target);
3b6b7bd1
MD
326 return(ip->ip_windex);
327}
328
41a01a4d 329/*
4c9f5a7f
MD
330 * Send an IPI request without blocking, return 0 on success, ENOENT on
331 * failure. The actual queueing of the hardware IPI may still force us
332 * to spin and process incoming IPIs but that will eventually go away
333 * when we've gotten rid of the other general IPIs.
41a01a4d
MD
334 */
335int
b8a98473
MD
336lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func,
337 void *arg1, int arg2)
41a01a4d
MD
338{
339 lwkt_ipiq_t ip;
340 int windex;
341 struct globaldata *gd = mycpu;
342
a7adb95a 343 logipiq(send_nbio, func, arg1, arg2, gd, target);
f9235b6d 344 KKASSERT(curthread->td_critcount);
41a01a4d 345 if (target == gd) {
b8a98473 346 func(arg1, arg2, NULL);
c92e86f1 347 logipiq(send_end, func, arg1, arg2, gd, target);
41a01a4d
MD
348 return(0);
349 }
350 ++ipiq_count;
351 ip = &gd->gd_ipiq[target->gd_cpuid];
352
ac72c7f4 353 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
a7adb95a 354 logipiq(send_fail, func, arg1, arg2, gd, target);
41a01a4d 355 return(ENOENT);
ac72c7f4 356 }
41a01a4d 357 windex = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473
MD
358 ip->ip_func[windex] = func;
359 ip->ip_arg1[windex] = arg1;
360 ip->ip_arg2[windex] = arg2;
35238fa5 361 cpu_sfence();
41a01a4d 362 ++ip->ip_windex;
4c9f5a7f 363
41a01a4d 364 /*
4c9f5a7f 365 * This isn't a passive IPI, we still have to signal the target cpu.
41a01a4d 366 */
4c9f5a7f 367 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
866b61fb 368 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f
MD
369 cpu_send_ipiq(target->gd_cpuid);
370 } else {
866b61fb
SZ
371 if (ipiq_optimized == 0) {
372 logipiq(cpu_send, func, arg1, arg2, gd, target);
4c9f5a7f 373 cpu_send_ipiq(target->gd_cpuid);
866b61fb 374 } else {
728f6208 375 ++ipiq_avoided;
866b61fb 376 }
4c9f5a7f 377 }
c92e86f1
SZ
378
379 logipiq(send_end, func, arg1, arg2, gd, target);
41a01a4d
MD
380 return(0);
381}
382
3b6b7bd1
MD
383/*
384 * deprecated, used only by fast int forwarding.
385 */
386int
b8a98473 387lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
3b6b7bd1 388{
b8a98473 389 return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
3b6b7bd1
MD
390}
391
392/*
393 * Send a message to several target cpus. Typically used for scheduling.
394 * The message will not be sent to stopped cpus.
395 */
396int
da23a592 397lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2)
3b6b7bd1
MD
398{
399 int cpuid;
400 int count = 0;
401
402 mask &= ~stopped_cpus;
403 while (mask) {
da23a592 404 cpuid = BSFCPUMASK(mask);
b8a98473 405 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
da23a592 406 mask &= ~CPUMASK(cpuid);
3b6b7bd1
MD
407 ++count;
408 }
409 return(count);
410}
411
412/*
413 * Wait for the remote cpu to finish processing a function.
414 *
415 * YYY we have to enable interrupts and process the IPIQ while waiting
416 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
417 * function to do this! YYY we really should 'block' here.
418 *
419 * MUST be called from a critical section. This routine may be called
420 * from an interrupt (for example, if an interrupt wakes a foreign thread
421 * up).
422 */
423void
424lwkt_wait_ipiq(globaldata_t target, int seq)
425{
426 lwkt_ipiq_t ip;
427 int maxc = 100000000;
428
429 if (target != mycpu) {
430 ip = &mycpu->gd_ipiq[target->gd_cpuid];
431 if ((int)(ip->ip_xindex - seq) < 0) {
46d4e165 432#if defined(__i386__)
3b6b7bd1 433 unsigned int eflags = read_eflags();
b2b3ffcd 434#elif defined(__x86_64__)
46d4e165
JG
435 unsigned long rflags = read_rflags();
436#endif
3b6b7bd1 437 cpu_enable_intr();
cfaeae2a 438 DEBUG_PUSH_INFO("wait_ipiq");
3b6b7bd1 439 while ((int)(ip->ip_xindex - seq) < 0) {
41a01a4d 440 crit_enter();
3b6b7bd1 441 lwkt_process_ipiq();
41a01a4d 442 crit_exit();
3b6b7bd1 443 if (--maxc == 0)
6ea70f76 444 kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
3b6b7bd1
MD
445 if (maxc < -1000000)
446 panic("LWKT_WAIT_IPIQ");
35238fa5
MD
447 /*
448 * xindex may be modified by another cpu, use a load fence
449 * to ensure that the loop does not use a speculative value
450 * (which may improve performance).
451 */
452 cpu_lfence();
3b6b7bd1 453 }
cfaeae2a 454 DEBUG_POP_INFO();
46d4e165 455#if defined(__i386__)
3b6b7bd1 456 write_eflags(eflags);
b2b3ffcd 457#elif defined(__x86_64__)
46d4e165
JG
458 write_rflags(rflags);
459#endif
3b6b7bd1
MD
460 }
461 }
462}
463
41a01a4d
MD
464int
465lwkt_seq_ipiq(globaldata_t target)
466{
467 lwkt_ipiq_t ip;
468
469 ip = &mycpu->gd_ipiq[target->gd_cpuid];
470 return(ip->ip_windex);
471}
472
3b6b7bd1
MD
473/*
474 * Called from IPI interrupt (like a fast interrupt), which has placed
475 * us in a critical section. The MP lock may or may not be held.
476 * May also be called from doreti or splz, or be reentrantly called
477 * indirectly through the ip_func[] we run.
478 *
479 * There are two versions, one where no interrupt frame is available (when
480 * called from the send code and from splz, and one where an interrupt
481 * frame is available.
d5b2d319
MD
482 *
483 * When the current cpu is mastering a cpusync we do NOT internally loop
484 * on the cpusyncq poll. We also do not re-flag a pending ipi due to
485 * the cpusyncq poll because this can cause doreti/splz to loop internally.
486 * The cpusync master's own loop must be allowed to run to avoid a deadlock.
3b6b7bd1
MD
487 */
488void
489lwkt_process_ipiq(void)
490{
491 globaldata_t gd = mycpu;
ac72c7f4 492 globaldata_t sgd;
3b6b7bd1
MD
493 lwkt_ipiq_t ip;
494 int n;
495
496again:
497 for (n = 0; n < ncpus; ++n) {
498 if (n != gd->gd_cpuid) {
ac72c7f4
MD
499 sgd = globaldata_find(n);
500 ip = sgd->gd_ipiq;
3b6b7bd1 501 if (ip != NULL) {
b8a98473 502 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL))
3b6b7bd1
MD
503 ;
504 }
505 }
506 }
507 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
b8a98473 508 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) {
0f7a3396
MD
509 if (gd->gd_curthread->td_cscount == 0)
510 goto again;
0f7a3396 511 }
3b6b7bd1
MD
512 }
513}
514
3b6b7bd1 515void
c7eb0589 516lwkt_process_ipiq_frame(struct intrframe *frame)
3b6b7bd1
MD
517{
518 globaldata_t gd = mycpu;
ac72c7f4 519 globaldata_t sgd;
3b6b7bd1
MD
520 lwkt_ipiq_t ip;
521 int n;
522
523again:
524 for (n = 0; n < ncpus; ++n) {
525 if (n != gd->gd_cpuid) {
ac72c7f4
MD
526 sgd = globaldata_find(n);
527 ip = sgd->gd_ipiq;
3b6b7bd1 528 if (ip != NULL) {
c7eb0589 529 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], frame))
3b6b7bd1
MD
530 ;
531 }
532 }
533 }
534 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
c7eb0589 535 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame)) {
0f7a3396
MD
536 if (gd->gd_curthread->td_cscount == 0)
537 goto again;
0f7a3396 538 }
3b6b7bd1
MD
539 }
540}
3b6b7bd1 541
cfaeae2a
MD
542#if 0
543static int iqticks[SMP_MAXCPU];
544static int iqcount[SMP_MAXCPU];
545#endif
546#if 0
547static int iqterm[SMP_MAXCPU];
548#endif
549
3b6b7bd1 550static int
b8a98473
MD
551lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
552 struct intrframe *frame)
3b6b7bd1 553{
2de4f77e 554 globaldata_t mygd = mycpu;
3b6b7bd1 555 int ri;
35238fa5 556 int wi;
b8a98473
MD
557 ipifunc3_t copy_func;
558 void *copy_arg1;
559 int copy_arg2;
35238fa5 560
cfaeae2a
MD
561#if 0
562 if (iqticks[mygd->gd_cpuid] != ticks) {
563 iqticks[mygd->gd_cpuid] = ticks;
564 iqcount[mygd->gd_cpuid] = 0;
565 }
566 if (++iqcount[mygd->gd_cpuid] > 3000000) {
567 kprintf("cpu %d ipiq maxed cscount %d spin %d\n",
568 mygd->gd_cpuid,
569 mygd->gd_curthread->td_cscount,
570 mygd->gd_spinlocks_wr);
571 iqcount[mygd->gd_cpuid] = 0;
572#if 0
573 if (++iqterm[mygd->gd_cpuid] > 10)
574 panic("cpu %d ipiq maxed", mygd->gd_cpuid);
575#endif
576 int i;
577 for (i = 0; i < ncpus; ++i) {
578 if (globaldata_find(i)->gd_infomsg)
579 kprintf(" %s", globaldata_find(i)->gd_infomsg);
580 }
581 kprintf("\n");
582 }
583#endif
584
35238fa5
MD
585 /*
586 * Obtain the current write index, which is modified by a remote cpu.
587 * Issue a load fence to prevent speculative reads of e.g. data written
588 * by the other cpu prior to it updating the index.
589 */
f9235b6d 590 KKASSERT(curthread->td_critcount);
35238fa5
MD
591 wi = ip->ip_windex;
592 cpu_lfence();
2de4f77e 593 ++mygd->gd_intr_nesting_level;
35238fa5 594
3b6b7bd1 595 /*
562273ea
MD
596 * NOTE: xindex is only updated after we are sure the function has
597 * finished execution. Beware lwkt_process_ipiq() reentrancy!
598 * The function may send an IPI which may block/drain.
d64a7617 599 *
562273ea
MD
600 * NOTE: Due to additional IPI operations that the callback function
601 * may make, it is possible for both rindex and windex to advance and
602 * thus for rindex to advance passed our cached windex.
603 *
d5b2d319 604 * NOTE: A load fence is required to prevent speculative loads prior
562273ea 605 * to the loading of ip_rindex. Even though stores might be
d5b2d319
MD
606 * ordered, loads are probably not. A memory fence is required
607 * to prevent reordering of the loads after the ip_rindex update.
3b6b7bd1 608 */
d64a7617 609 while (wi - (ri = ip->ip_rindex) > 0) {
3b6b7bd1 610 ri &= MAXCPUFIFO_MASK;
d5b2d319 611 cpu_lfence();
728f6208 612 copy_func = ip->ip_func[ri];
b8a98473
MD
613 copy_arg1 = ip->ip_arg1[ri];
614 copy_arg2 = ip->ip_arg2[ri];
d5b2d319 615 cpu_mfence();
728f6208 616 ++ip->ip_rindex;
562273ea
MD
617 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) ==
618 ((ri + 1) & MAXCPUFIFO_MASK));
a7adb95a 619 logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
d5b2d319
MD
620#ifdef INVARIANTS
621 if (ipiq_debug && (ip->ip_rindex & 0xFFFFFF) == 0) {
622 kprintf("cpu %d ipifunc %p %p %d (frame %p)\n",
623 mycpu->gd_cpuid,
624 copy_func, copy_arg1, copy_arg2,
625#if defined(__i386__)
626 (frame ? (void *)frame->if_eip : NULL));
627#elif defined(__amd64__)
628 (frame ? (void *)frame->if_rip : NULL));
629#else
630 NULL);
631#endif
632 }
633#endif
b8a98473 634 copy_func(copy_arg1, copy_arg2, frame);
35238fa5 635 cpu_sfence();
3b6b7bd1 636 ip->ip_xindex = ip->ip_rindex;
e8f15168
MD
637
638#ifdef PANIC_DEBUG
639 /*
640 * Simulate panics during the processing of an IPI
641 */
642 if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
643 if (--panic_ipiq_count == 0) {
644#ifdef DDB
645 Debugger("PANIC_DEBUG");
646#else
647 panic("PANIC_DEBUG");
648#endif
649 }
650 }
651#endif
3b6b7bd1 652 }
2de4f77e 653 --mygd->gd_intr_nesting_level;
4c9f5a7f
MD
654
655 /*
656 * Return non-zero if there are more IPI messages pending on this
657 * ipiq. ip_npoll is left set as long as possible to reduce the
658 * number of IPIs queued by the originating cpu, but must be cleared
659 * *BEFORE* checking windex.
660 */
661 atomic_poll_release_int(&ip->ip_npoll);
3b6b7bd1
MD
662 return(wi != ip->ip_windex);
663}
664
6c92c1f2
SZ
665static void
666lwkt_sync_ipiq(void *arg)
667{
5a1a2253 668 volatile cpumask_t *cpumask = arg;
6c92c1f2 669
da23a592 670 atomic_clear_cpumask(cpumask, mycpu->gd_cpumask);
6c92c1f2
SZ
671 if (*cpumask == 0)
672 wakeup(cpumask);
673}
674
675void
676lwkt_synchronize_ipiqs(const char *wmesg)
677{
5a1a2253 678 volatile cpumask_t other_cpumask;
6c92c1f2
SZ
679
680 other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
5a1a2253
SZ
681 lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq,
682 __DEVOLATILE(void *, &other_cpumask));
6c92c1f2 683
6c92c1f2 684 while (other_cpumask != 0) {
ae8e83e6 685 tsleep_interlock(&other_cpumask, 0);
6c92c1f2 686 if (other_cpumask != 0)
d9345d3a 687 tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0);
6c92c1f2 688 }
6c92c1f2
SZ
689}
690
0f7a3396
MD
691#endif
692
3b6b7bd1
MD
693/*
694 * CPU Synchronization Support
5c71a36a 695 *
d5b2d319
MD
696 * lwkt_cpusync_interlock() - Place specified cpus in a quiescent state.
697 * The current cpu is placed in a hard critical
698 * section.
5c71a36a 699 *
d5b2d319
MD
700 * lwkt_cpusync_deinterlock() - Execute cs_func on specified cpus, including
701 * current cpu if specified, then return.
3b6b7bd1 702 */
3b6b7bd1 703void
d5b2d319 704lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *arg)
5c71a36a 705{
d5b2d319 706 struct lwkt_cpusync cs;
5c71a36a 707
d5b2d319
MD
708 lwkt_cpusync_init(&cs, mask, func, arg);
709 lwkt_cpusync_interlock(&cs);
710 lwkt_cpusync_deinterlock(&cs);
3b6b7bd1
MD
711}
712
d5b2d319 713
5c71a36a 714void
d5b2d319 715lwkt_cpusync_interlock(lwkt_cpusync_t cs)
3b6b7bd1 716{
d5b2d319 717#ifdef SMP
0f7a3396 718 globaldata_t gd = mycpu;
d5b2d319 719 cpumask_t mask;
0f7a3396 720
d5b2d319
MD
721 /*
722 * mask acknowledge (cs_mack): 0->mask for stage 1
723 *
724 * mack does not include the current cpu.
725 */
726 mask = cs->cs_mask & gd->gd_other_cpus & smp_active_mask;
727 cs->cs_mack = 0;
728 crit_enter_id("cpusync");
729 if (mask) {
cfaeae2a 730 DEBUG_PUSH_INFO("cpusync_interlock");
0f7a3396
MD
731 ++ipiq_cscount;
732 ++gd->gd_curthread->td_cscount;
d5b2d319
MD
733 lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote1, cs);
734 logipiq2(sync_start, mask);
735 while (cs->cs_mack != mask) {
0f7a3396 736 lwkt_process_ipiq();
d5b2d319 737 cpu_pause();
0f7a3396 738 }
cfaeae2a 739 DEBUG_POP_INFO();
3b6b7bd1 740 }
d5b2d319
MD
741#else
742 cs->cs_mack = 0;
0f7a3396 743#endif
3b6b7bd1
MD
744}
745
746/*
d5b2d319
MD
747 * Interlocked cpus have executed remote1 and are polling in remote2.
748 * To deinterlock we clear cs_mack and wait for the cpus to execute
749 * the func and set their bit in cs_mack again.
0f7a3396 750 *
3b6b7bd1
MD
751 */
752void
d5b2d319 753lwkt_cpusync_deinterlock(lwkt_cpusync_t cs)
3b6b7bd1 754{
0f7a3396 755 globaldata_t gd = mycpu;
0f7a3396 756#ifdef SMP
d5b2d319
MD
757 cpumask_t mask;
758
759 /*
760 * mask acknowledge (cs_mack): mack->0->mack for stage 2
761 *
762 * Clearing cpu bits for polling cpus in cs_mack will cause them to
763 * execute stage 2, which executes the cs_func(cs_data) and then sets
764 * their bit in cs_mack again.
765 *
766 * mack does not include the current cpu.
767 */
768 mask = cs->cs_mack;
769 cpu_ccfence();
770 cs->cs_mack = 0;
771 if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
772 cs->cs_func(cs->cs_data);
773 if (mask) {
cfaeae2a 774 DEBUG_PUSH_INFO("cpusync_deinterlock");
d5b2d319 775 while (cs->cs_mack != mask) {
0f7a3396 776 lwkt_process_ipiq();
d5b2d319 777 cpu_pause();
0f7a3396 778 }
cfaeae2a
MD
779 DEBUG_POP_INFO();
780 /*
781 * cpusyncq ipis may be left queued without the RQF flag set due to
782 * a non-zero td_cscount, so be sure to process any laggards after
783 * decrementing td_cscount.
784 */
0f7a3396 785 --gd->gd_curthread->td_cscount;
d5b2d319
MD
786 lwkt_process_ipiq();
787 logipiq2(sync_end, mask);
3b6b7bd1 788 }
d5b2d319
MD
789 crit_exit_id("cpusync");
790#else
791 if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
792 cs->cs_func(cs->cs_data);
0f7a3396 793#endif
3b6b7bd1
MD
794}
795
0f7a3396
MD
796#ifdef SMP
797
3b6b7bd1
MD
798/*
799 * helper IPI remote messaging function.
800 *
801 * Called on remote cpu when a new cpu synchronization request has been
802 * sent to us. Execute the run function and adjust cs_count, then requeue
803 * the request so we spin on it.
804 */
805static void
d5b2d319 806lwkt_cpusync_remote1(lwkt_cpusync_t cs)
3b6b7bd1 807{
d5b2d319
MD
808 globaldata_t gd = mycpu;
809
810 atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
811 lwkt_cpusync_remote2(cs);
3b6b7bd1
MD
812}
813
814/*
815 * helper IPI remote messaging function.
816 *
817 * Poll for the originator telling us to finish. If it hasn't, requeue
d5b2d319 818 * our request so we spin on it.
3b6b7bd1
MD
819 */
820static void
d5b2d319 821lwkt_cpusync_remote2(lwkt_cpusync_t cs)
3b6b7bd1 822{
d5b2d319
MD
823 globaldata_t gd = mycpu;
824
825 if ((cs->cs_mack & gd->gd_cpumask) == 0) {
826 if (cs->cs_func)
827 cs->cs_func(cs->cs_data);
828 atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
3b6b7bd1 829 } else {
3b6b7bd1
MD
830 lwkt_ipiq_t ip;
831 int wi;
832
833 ip = &gd->gd_cpusyncq;
834 wi = ip->ip_windex & MAXCPUFIFO_MASK;
b8a98473 835 ip->ip_func[wi] = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
d5b2d319 836 ip->ip_arg1[wi] = cs;
b8a98473 837 ip->ip_arg2[wi] = 0;
35238fa5 838 cpu_sfence();
3b6b7bd1 839 ++ip->ip_windex;
37494a7a 840 if (ipiq_debug && (ip->ip_windex & 0xFFFFFF) == 0) {
cfaeae2a
MD
841 kprintf("cpu %d cm=%016jx %016jx f=%p\n",
842 gd->gd_cpuid,
843 (intmax_t)cs->cs_mask, (intmax_t)cs->cs_mack,
844 cs->cs_func);
37494a7a 845 }
3b6b7bd1
MD
846 }
847}
848
3b6b7bd1 849#endif