2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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
36 * This module implements IPI message queueing and the MI portion of IPI
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
46 #include <sys/rtprio.h>
47 #include <sys/queue.h>
48 #include <sys/thread2.h>
49 #include <sys/sysctl.h>
51 #include <sys/kthread.h>
52 #include <machine/cpu.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>
65 #include <machine/stdarg.h>
66 #include <machine/smp.h>
67 #include <machine/atomic.h>
70 static __int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */
71 static __int64_t ipiq_fifofull; /* number of fifo full conditions detected */
72 static __int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */
73 static __int64_t ipiq_passive; /* passive IPI messages */
74 static __int64_t ipiq_cscount; /* number of cpu synchronizations */
75 static int ipiq_debug; /* set to 1 for debug */
77 static int panic_ipiq_cpu = -1;
78 static int panic_ipiq_count = 100;
83 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0,
84 "Number of IPI's sent");
85 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0,
86 "Number of fifo full conditions detected");
87 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0,
88 "Number of IPI's avoided by interlock with target cpu");
89 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0,
90 "Number of passive IPI messages sent");
91 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0,
92 "Number of cpu synchronizations");
93 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_debug, CTLFLAG_RW, &ipiq_debug, 0,
96 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, "");
97 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, "");
100 #define IPIQ_STRING "func=%p arg1=%p arg2=%d scpu=%d dcpu=%d"
101 #define IPIQ_ARGS void *func, void *arg1, int arg2, int scpu, int dcpu
103 #if !defined(KTR_IPIQ)
104 #define KTR_IPIQ KTR_ALL
106 KTR_INFO_MASTER(ipiq);
107 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARGS);
108 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARGS);
109 KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARGS);
110 KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARGS);
111 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARGS);
112 KTR_INFO(KTR_IPIQ, ipiq, sync_start, 5, "cpumask=%08lx", unsigned long mask);
113 KTR_INFO(KTR_IPIQ, ipiq, sync_end, 6, "cpumask=%08lx", unsigned long mask);
114 KTR_INFO(KTR_IPIQ, ipiq, cpu_send, 7, IPIQ_STRING, IPIQ_ARGS);
115 KTR_INFO(KTR_IPIQ, ipiq, send_end, 8, IPIQ_STRING, IPIQ_ARGS);
117 #define logipiq(name, func, arg1, arg2, sgd, dgd) \
118 KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid)
119 #define logipiq2(name, arg) \
120 KTR_LOG(ipiq_ ## name, arg)
126 static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
127 struct intrframe *frame);
128 static void lwkt_cpusync_remote1(lwkt_cpusync_t cs);
129 static void lwkt_cpusync_remote2(lwkt_cpusync_t cs);
132 * Send a function execution request to another cpu. The request is queued
133 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
134 * possible target cpu. The FIFO can be written.
136 * If the FIFO fills up we have to enable interrupts to avoid an APIC
137 * deadlock and process pending IPIQs while waiting for it to empty.
138 * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
140 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
141 * end will take care of any pending interrupts.
143 * The actual hardware IPI is avoided if the target cpu is already processing
144 * the queue from a prior IPI. It is possible to pipeline IPI messages
145 * very quickly between cpus due to the FIFO hysteresis.
147 * Need not be called from a critical section.
150 lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2)
154 struct globaldata *gd = mycpu;
156 logipiq(send_norm, func, arg1, arg2, gd, target);
159 func(arg1, arg2, NULL);
160 logipiq(send_end, func, arg1, arg2, gd, target);
164 ++gd->gd_intr_nesting_level;
166 if (gd->gd_intr_nesting_level > 20)
167 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
169 KKASSERT(curthread->td_critcount);
171 ip = &gd->gd_ipiq[target->gd_cpuid];
174 * Do not allow the FIFO to become full. Interrupts must be physically
175 * enabled while we liveloop to avoid deadlocking the APIC.
177 * The target ipiq may have gotten filled up due to passive IPIs and thus
178 * not be aware that its queue is too full, so be sure to issue an
179 * ipiq interrupt to the target cpu.
181 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
182 #if defined(__i386__)
183 unsigned int eflags = read_eflags();
184 #elif defined(__x86_64__)
185 unsigned long rflags = read_rflags();
190 DEBUG_PUSH_INFO("send_ipiq3");
191 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
192 if (atomic_poll_acquire_int(&target->gd_npoll)) {
193 logipiq(cpu_send, func, arg1, arg2, gd, target);
194 cpu_send_ipiq(target->gd_cpuid);
196 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
201 #if defined(__i386__)
202 write_eflags(eflags);
203 #elif defined(__x86_64__)
204 write_rflags(rflags);
209 * Queue the new message
211 windex = ip->ip_windex & MAXCPUFIFO_MASK;
212 ip->ip_info[windex].func = func;
213 ip->ip_info[windex].arg1 = arg1;
214 ip->ip_info[windex].arg2 = arg2;
217 atomic_set_cpumask(&target->gd_ipimask, gd->gd_cpumask);
220 * signal the target cpu that there is work pending.
222 if (atomic_poll_acquire_int(&target->gd_npoll)) {
223 logipiq(cpu_send, func, arg1, arg2, gd, target);
224 cpu_send_ipiq(target->gd_cpuid);
228 --gd->gd_intr_nesting_level;
230 logipiq(send_end, func, arg1, arg2, gd, target);
232 return(ip->ip_windex);
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
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.
244 * Need not be called from a critical section.
247 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
248 void *arg1, int arg2)
252 struct globaldata *gd = mycpu;
254 KKASSERT(target != gd);
256 ++gd->gd_intr_nesting_level;
257 logipiq(send_pasv, func, arg1, arg2, gd, target);
259 if (gd->gd_intr_nesting_level > 20)
260 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
262 KKASSERT(curthread->td_critcount);
265 ip = &gd->gd_ipiq[target->gd_cpuid];
268 * Do not allow the FIFO to become full. Interrupts must be physically
269 * enabled while we liveloop to avoid deadlocking the APIC.
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();
280 DEBUG_PUSH_INFO("send_ipiq3_passive");
281 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
282 if (atomic_poll_acquire_int(&target->gd_npoll)) {
283 logipiq(cpu_send, func, arg1, arg2, gd, target);
284 cpu_send_ipiq(target->gd_cpuid);
286 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
291 #if defined(__i386__)
292 write_eflags(eflags);
293 #elif defined(__x86_64__)
294 write_rflags(rflags);
299 * Queue the new message
301 windex = ip->ip_windex & MAXCPUFIFO_MASK;
302 ip->ip_info[windex].func = func;
303 ip->ip_info[windex].arg1 = arg1;
304 ip->ip_info[windex].arg2 = arg2;
307 atomic_set_cpumask(&target->gd_ipimask, gd->gd_cpumask);
308 --gd->gd_intr_nesting_level;
311 * Do not signal the target cpu, it will pick up the IPI when it next
312 * polls (typically on the next tick).
315 logipiq(send_end, func, arg1, arg2, gd, target);
317 return(ip->ip_windex);
321 * Send an IPI request without blocking, return 0 on success, ENOENT on
322 * failure. The actual queueing of the hardware IPI may still force us
323 * to spin and process incoming IPIs but that will eventually go away
324 * when we've gotten rid of the other general IPIs.
327 lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func,
328 void *arg1, int arg2)
332 struct globaldata *gd = mycpu;
334 logipiq(send_nbio, func, arg1, arg2, gd, target);
335 KKASSERT(curthread->td_critcount);
337 func(arg1, arg2, NULL);
338 logipiq(send_end, func, arg1, arg2, gd, target);
342 ++gd->gd_intr_nesting_level;
344 ip = &gd->gd_ipiq[target->gd_cpuid];
346 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
347 logipiq(send_fail, func, arg1, arg2, gd, target);
348 --gd->gd_intr_nesting_level;
352 windex = ip->ip_windex & MAXCPUFIFO_MASK;
353 ip->ip_info[windex].func = func;
354 ip->ip_info[windex].arg1 = arg1;
355 ip->ip_info[windex].arg2 = arg2;
358 atomic_set_cpumask(&target->gd_ipimask, gd->gd_cpumask);
361 * This isn't a passive IPI, we still have to signal the target cpu.
363 if (atomic_poll_acquire_int(&target->gd_npoll)) {
364 logipiq(cpu_send, func, arg1, arg2, gd, target);
365 cpu_send_ipiq(target->gd_cpuid);
369 --gd->gd_intr_nesting_level;
372 logipiq(send_end, func, arg1, arg2, gd, target);
377 * deprecated, used only by fast int forwarding.
380 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
382 return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
386 * Send a message to several target cpus. Typically used for scheduling.
387 * The message will not be sent to stopped cpus.
390 lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2)
395 mask &= ~stopped_cpus;
397 cpuid = BSFCPUMASK(mask);
398 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
399 mask &= ~CPUMASK(cpuid);
406 * Wait for the remote cpu to finish processing a function.
408 * YYY we have to enable interrupts and process the IPIQ while waiting
409 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
410 * function to do this! YYY we really should 'block' here.
412 * MUST be called from a critical section. This routine may be called
413 * from an interrupt (for example, if an interrupt wakes a foreign thread
417 lwkt_wait_ipiq(globaldata_t target, int seq)
420 int maxc = 100000000;
422 if (target != mycpu) {
423 ip = &mycpu->gd_ipiq[target->gd_cpuid];
424 if ((int)(ip->ip_xindex - seq) < 0) {
425 #if defined(__i386__)
426 unsigned int eflags = read_eflags();
427 #elif defined(__x86_64__)
428 unsigned long rflags = read_rflags();
431 DEBUG_PUSH_INFO("wait_ipiq");
432 while ((int)(ip->ip_xindex - seq) < 0) {
437 kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
439 panic("LWKT_WAIT_IPIQ");
441 * xindex may be modified by another cpu, use a load fence
442 * to ensure that the loop does not use a speculative value
443 * (which may improve performance).
448 #if defined(__i386__)
449 write_eflags(eflags);
450 #elif defined(__x86_64__)
451 write_rflags(rflags);
458 lwkt_seq_ipiq(globaldata_t target)
462 ip = &mycpu->gd_ipiq[target->gd_cpuid];
463 return(ip->ip_windex);
467 * Called from IPI interrupt (like a fast interrupt), which has placed
468 * us in a critical section. The MP lock may or may not be held.
469 * May also be called from doreti or splz, or be reentrantly called
470 * indirectly through the ip_info[].func we run.
472 * There are two versions, one where no interrupt frame is available (when
473 * called from the send code and from splz, and one where an interrupt
474 * frame is available.
476 * When the current cpu is mastering a cpusync we do NOT internally loop
477 * on the cpusyncq poll. We also do not re-flag a pending ipi due to
478 * the cpusyncq poll because this can cause doreti/splz to loop internally.
479 * The cpusync master's own loop must be allowed to run to avoid a deadlock.
482 lwkt_process_ipiq(void)
484 globaldata_t gd = mycpu;
490 ++gd->gd_processing_ipiq;
493 mask = gd->gd_ipimask;
494 atomic_clear_cpumask(&gd->gd_ipimask, mask);
496 n = BSFCPUMASK(mask);
497 if (n != gd->gd_cpuid) {
498 sgd = globaldata_find(n);
501 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL))
509 * Process pending cpusyncs. If the current thread has a cpusync
510 * active cpusync we only run the list once and do not re-flag
511 * as the thread itself is processing its interlock.
513 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) {
514 if (gd->gd_curthread->td_cscount == 0)
516 /* need_ipiq(); do not reflag */
520 * Interlock to allow more IPI interrupts. Recheck ipimask after
521 * releasing gd_npoll.
525 atomic_poll_release_int(&gd->gd_npoll);
529 --gd->gd_processing_ipiq;
533 lwkt_process_ipiq_frame(struct intrframe *frame)
535 globaldata_t gd = mycpu;
543 mask = gd->gd_ipimask;
544 atomic_clear_cpumask(&gd->gd_ipimask, mask);
546 n = BSFCPUMASK(mask);
547 if (n != gd->gd_cpuid) {
548 sgd = globaldata_find(n);
551 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], frame))
557 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
558 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame)) {
559 if (gd->gd_curthread->td_cscount == 0)
561 /* need_ipiq(); do not reflag */
566 * Interlock to allow more IPI interrupts. Recheck ipimask after
567 * releasing gd_npoll.
571 atomic_poll_release_int(&gd->gd_npoll);
578 static int iqticks[SMP_MAXCPU];
579 static int iqcount[SMP_MAXCPU];
582 static int iqterm[SMP_MAXCPU];
586 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
587 struct intrframe *frame)
589 globaldata_t mygd = mycpu;
592 ipifunc3_t copy_func;
597 if (iqticks[mygd->gd_cpuid] != ticks) {
598 iqticks[mygd->gd_cpuid] = ticks;
599 iqcount[mygd->gd_cpuid] = 0;
601 if (++iqcount[mygd->gd_cpuid] > 3000000) {
602 kprintf("cpu %d ipiq maxed cscount %d spin %d\n",
604 mygd->gd_curthread->td_cscount,
606 iqcount[mygd->gd_cpuid] = 0;
608 if (++iqterm[mygd->gd_cpuid] > 10)
609 panic("cpu %d ipiq maxed", mygd->gd_cpuid);
612 for (i = 0; i < ncpus; ++i) {
613 if (globaldata_find(i)->gd_infomsg)
614 kprintf(" %s", globaldata_find(i)->gd_infomsg);
621 * Clear the originating core from our ipimask, we will process all
624 * Obtain the current write index, which is modified by a remote cpu.
625 * Issue a load fence to prevent speculative reads of e.g. data written
626 * by the other cpu prior to it updating the index.
628 KKASSERT(curthread->td_critcount);
631 ++mygd->gd_intr_nesting_level;
634 * NOTE: xindex is only updated after we are sure the function has
635 * finished execution. Beware lwkt_process_ipiq() reentrancy!
636 * The function may send an IPI which may block/drain.
638 * NOTE: Due to additional IPI operations that the callback function
639 * may make, it is possible for both rindex and windex to advance and
640 * thus for rindex to advance passed our cached windex.
642 * NOTE: A load fence is required to prevent speculative loads prior
643 * to the loading of ip_rindex. Even though stores might be
644 * ordered, loads are probably not. A memory fence is required
645 * to prevent reordering of the loads after the ip_rindex update.
647 * NOTE: Single pass only. Returns non-zero if the queue is not empty
650 while (wi - (ri = ip->ip_rindex) > 0) {
651 ri &= MAXCPUFIFO_MASK;
653 copy_func = ip->ip_info[ri].func;
654 copy_arg1 = ip->ip_info[ri].arg1;
655 copy_arg2 = ip->ip_info[ri].arg2;
658 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) ==
659 ((ri + 1) & MAXCPUFIFO_MASK));
660 logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
662 if (ipiq_debug && (ip->ip_rindex & 0xFFFFFF) == 0) {
663 kprintf("cpu %d ipifunc %p %p %d (frame %p)\n",
665 copy_func, copy_arg1, copy_arg2,
666 #if defined(__i386__)
667 (frame ? (void *)frame->if_eip : NULL));
668 #elif defined(__amd64__)
669 (frame ? (void *)frame->if_rip : NULL));
675 copy_func(copy_arg1, copy_arg2, frame);
677 ip->ip_xindex = ip->ip_rindex;
681 * Simulate panics during the processing of an IPI
683 if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
684 if (--panic_ipiq_count == 0) {
686 Debugger("PANIC_DEBUG");
688 panic("PANIC_DEBUG");
694 --mygd->gd_intr_nesting_level;
697 * Return non-zero if there is still more in the queue.
700 return (ip->ip_rindex != ip->ip_windex);
704 lwkt_sync_ipiq(void *arg)
706 volatile cpumask_t *cpumask = arg;
708 atomic_clear_cpumask(cpumask, mycpu->gd_cpumask);
714 lwkt_synchronize_ipiqs(const char *wmesg)
716 volatile cpumask_t other_cpumask;
718 other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
719 lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq,
720 __DEVOLATILE(void *, &other_cpumask));
722 while (other_cpumask != 0) {
723 tsleep_interlock(&other_cpumask, 0);
724 if (other_cpumask != 0)
725 tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0);
732 * CPU Synchronization Support
734 * lwkt_cpusync_interlock() - Place specified cpus in a quiescent state.
735 * The current cpu is placed in a hard critical
738 * lwkt_cpusync_deinterlock() - Execute cs_func on specified cpus, including
739 * current cpu if specified, then return.
742 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *arg)
744 struct lwkt_cpusync cs;
746 lwkt_cpusync_init(&cs, mask, func, arg);
747 lwkt_cpusync_interlock(&cs);
748 lwkt_cpusync_deinterlock(&cs);
753 lwkt_cpusync_interlock(lwkt_cpusync_t cs)
757 const char *smsg = "SMPSYNL";
759 globaldata_t gd = mycpu;
763 * mask acknowledge (cs_mack): 0->mask for stage 1
765 * mack does not include the current cpu.
767 mask = cs->cs_mask & gd->gd_other_cpus & smp_active_mask;
769 crit_enter_id("cpusync");
771 DEBUG_PUSH_INFO("cpusync_interlock");
773 ++gd->gd_curthread->td_cscount;
774 lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote1, cs);
775 logipiq2(sync_start, (long)mask);
777 if (gd->gd_curthread->td_wmesg == NULL)
778 gd->gd_curthread->td_wmesg = smsg;
780 while (cs->cs_mack != mask) {
785 if (gd->gd_curthread->td_wmesg == smsg)
786 gd->gd_curthread->td_wmesg = NULL;
796 * Interlocked cpus have executed remote1 and are polling in remote2.
797 * To deinterlock we clear cs_mack and wait for the cpus to execute
798 * the func and set their bit in cs_mack again.
802 lwkt_cpusync_deinterlock(lwkt_cpusync_t cs)
804 globaldata_t gd = mycpu;
807 const char *smsg = "SMPSYNU";
812 * mask acknowledge (cs_mack): mack->0->mack for stage 2
814 * Clearing cpu bits for polling cpus in cs_mack will cause them to
815 * execute stage 2, which executes the cs_func(cs_data) and then sets
816 * their bit in cs_mack again.
818 * mack does not include the current cpu.
824 if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
825 cs->cs_func(cs->cs_data);
827 DEBUG_PUSH_INFO("cpusync_deinterlock");
829 if (gd->gd_curthread->td_wmesg == NULL)
830 gd->gd_curthread->td_wmesg = smsg;
832 while (cs->cs_mack != mask) {
837 if (gd->gd_curthread->td_wmesg == smsg)
838 gd->gd_curthread->td_wmesg = NULL;
842 * cpusyncq ipis may be left queued without the RQF flag set due to
843 * a non-zero td_cscount, so be sure to process any laggards after
844 * decrementing td_cscount.
846 --gd->gd_curthread->td_cscount;
848 logipiq2(sync_end, (long)mask);
850 crit_exit_id("cpusync");
852 if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
853 cs->cs_func(cs->cs_data);
860 * helper IPI remote messaging function.
862 * Called on remote cpu when a new cpu synchronization request has been
863 * sent to us. Execute the run function and adjust cs_count, then requeue
864 * the request so we spin on it.
867 lwkt_cpusync_remote1(lwkt_cpusync_t cs)
869 globaldata_t gd = mycpu;
871 atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
872 lwkt_cpusync_remote2(cs);
876 * helper IPI remote messaging function.
878 * Poll for the originator telling us to finish. If it hasn't, requeue
879 * our request so we spin on it.
882 lwkt_cpusync_remote2(lwkt_cpusync_t cs)
884 globaldata_t gd = mycpu;
886 if ((cs->cs_mack & gd->gd_cpumask) == 0) {
888 cs->cs_func(cs->cs_data);
889 atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
890 /* cs can be ripped out at this point */
895 ip = &gd->gd_cpusyncq;
896 wi = ip->ip_windex & MAXCPUFIFO_MASK;
897 ip->ip_info[wi].func = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
898 ip->ip_info[wi].arg1 = cs;
899 ip->ip_info[wi].arg2 = 0;
901 KKASSERT(ip->ip_windex - ip->ip_rindex < MAXCPUFIFO);
903 if (ipiq_debug && (ip->ip_windex & 0xFFFFFF) == 0) {
904 kprintf("cpu %d cm=%016jx %016jx f=%p\n",
906 (intmax_t)cs->cs_mask, (intmax_t)cs->cs_mack,