xref: /dflybsd-src/sys/kern/lwkt_ipiq.c (revision dae741e33c840b92a8a53bf9f01157ede145e256)
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 
35 /*
36  * This module implements IPI message queueing and the MI portion of IPI
37  * message processing.
38  */
39 
40 #include "opt_ddb.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/proc.h>
46 #include <sys/rtprio.h>
47 #include <sys/queue.h>
48 #include <sys/thread2.h>
49 #include <sys/sysctl.h>
50 #include <sys/ktr.h>
51 #include <sys/kthread.h>
52 #include <machine/cpu.h>
53 #include <sys/lock.h>
54 #include <sys/caps.h>
55 
56 #include <vm/vm.h>
57 #include <vm/vm_param.h>
58 #include <vm/vm_kern.h>
59 #include <vm/vm_object.h>
60 #include <vm/vm_page.h>
61 #include <vm/vm_map.h>
62 #include <vm/vm_pager.h>
63 #include <vm/vm_extern.h>
64 #include <vm/vm_zone.h>
65 
66 #include <machine/stdarg.h>
67 #include <machine/smp.h>
68 #include <machine/atomic.h>
69 
70 #ifdef SMP
71 static __int64_t ipiq_count;	/* total calls to lwkt_send_ipiq*() */
72 static __int64_t ipiq_fifofull;	/* number of fifo full conditions detected */
73 static __int64_t ipiq_avoided;	/* interlock with target avoids cpu ipi */
74 static __int64_t ipiq_passive;	/* passive IPI messages */
75 static __int64_t ipiq_cscount;	/* number of cpu synchronizations */
76 static int ipiq_debug;		/* set to 1 for debug */
77 #ifdef PANIC_DEBUG
78 static int	panic_ipiq_cpu = -1;
79 static int	panic_ipiq_count = 100;
80 #endif
81 #endif
82 
83 #ifdef SMP
84 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0,
85     "Number of IPI's sent");
86 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0,
87     "Number of fifo full conditions detected");
88 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0,
89     "Number of IPI's avoided by interlock with target cpu");
90 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0,
91     "Number of passive IPI messages sent");
92 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0,
93     "Number of cpu synchronizations");
94 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_debug, CTLFLAG_RW, &ipiq_debug, 0,
95     "");
96 #ifdef PANIC_DEBUG
97 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, "");
98 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, "");
99 #endif
100 
101 #define IPIQ_STRING	"func=%p arg1=%p arg2=%d scpu=%d dcpu=%d"
102 #define IPIQ_ARG_SIZE	(sizeof(void *) * 2 + sizeof(int) * 3)
103 
104 #if !defined(KTR_IPIQ)
105 #define KTR_IPIQ	KTR_ALL
106 #endif
107 KTR_INFO_MASTER(ipiq);
108 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARG_SIZE);
109 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARG_SIZE);
110 KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARG_SIZE);
111 KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARG_SIZE);
112 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARG_SIZE);
113 KTR_INFO(KTR_IPIQ, ipiq, sync_start, 5, "cpumask=%08x", sizeof(cpumask_t));
114 KTR_INFO(KTR_IPIQ, ipiq, sync_end, 6, "cpumask=%08x", sizeof(cpumask_t));
115 KTR_INFO(KTR_IPIQ, ipiq, cpu_send, 7, IPIQ_STRING, IPIQ_ARG_SIZE);
116 KTR_INFO(KTR_IPIQ, ipiq, send_end, 8, IPIQ_STRING, IPIQ_ARG_SIZE);
117 
118 #define logipiq(name, func, arg1, arg2, sgd, dgd)	\
119 	KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid)
120 #define logipiq2(name, arg)	\
121 	KTR_LOG(ipiq_ ## name, arg)
122 
123 #endif	/* SMP */
124 
125 #ifdef SMP
126 
127 static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
128 				  struct intrframe *frame);
129 static void lwkt_cpusync_remote1(lwkt_cpusync_t cs);
130 static void lwkt_cpusync_remote2(lwkt_cpusync_t cs);
131 
132 /*
133  * Send a function execution request to another cpu.  The request is queued
134  * on the cpu<->cpu ipiq matrix.  Each cpu owns a unique ipiq FIFO for every
135  * possible target cpu.  The FIFO can be written.
136  *
137  * If the FIFO fills up we have to enable interrupts to avoid an APIC
138  * deadlock and process pending IPIQs while waiting for it to empty.
139  * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
140  *
141  * We can safely bump gd_intr_nesting_level because our crit_exit() at the
142  * end will take care of any pending interrupts.
143  *
144  * The actual hardware IPI is avoided if the target cpu is already processing
145  * the queue from a prior IPI.  It is possible to pipeline IPI messages
146  * very quickly between cpus due to the FIFO hysteresis.
147  *
148  * Need not be called from a critical section.
149  */
150 int
151 lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2)
152 {
153     lwkt_ipiq_t ip;
154     int windex;
155     struct globaldata *gd = mycpu;
156 
157     logipiq(send_norm, func, arg1, arg2, gd, target);
158 
159     if (target == gd) {
160 	func(arg1, arg2, NULL);
161 	logipiq(send_end, func, arg1, arg2, gd, target);
162 	return(0);
163     }
164     crit_enter();
165     ++gd->gd_intr_nesting_level;
166 #ifdef INVARIANTS
167     if (gd->gd_intr_nesting_level > 20)
168 	panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
169 #endif
170     KKASSERT(curthread->td_critcount);
171     ++ipiq_count;
172     ip = &gd->gd_ipiq[target->gd_cpuid];
173 
174     /*
175      * Do not allow the FIFO to become full.  Interrupts must be physically
176      * enabled while we liveloop to avoid deadlocking the APIC.
177      *
178      * The target ipiq may have gotten filled up due to passive IPIs and thus
179      * not be aware that its queue is too full, so be sure to issue an
180      * ipiq interrupt to the target cpu.
181      */
182     if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
183 #if defined(__i386__)
184 	unsigned int eflags = read_eflags();
185 #elif defined(__x86_64__)
186 	unsigned long rflags = read_rflags();
187 #endif
188 
189 	cpu_enable_intr();
190 	++ipiq_fifofull;
191 	DEBUG_PUSH_INFO("send_ipiq3");
192 	while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
193 	    if (atomic_poll_acquire_int(&target->gd_npoll)) {
194 		logipiq(cpu_send, func, arg1, arg2, gd, target);
195 		cpu_send_ipiq(target->gd_cpuid);
196 	    }
197 	    KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
198 	    lwkt_process_ipiq();
199 	    cpu_pause();
200 	}
201 	DEBUG_POP_INFO();
202 #if defined(__i386__)
203 	write_eflags(eflags);
204 #elif defined(__x86_64__)
205 	write_rflags(rflags);
206 #endif
207     }
208 
209     /*
210      * Queue the new message
211      */
212     windex = ip->ip_windex & MAXCPUFIFO_MASK;
213     ip->ip_info[windex].func = func;
214     ip->ip_info[windex].arg1 = arg1;
215     ip->ip_info[windex].arg2 = arg2;
216     cpu_sfence();
217     ++ip->ip_windex;
218     atomic_set_cpumask(&target->gd_ipimask, gd->gd_cpumask);
219 
220     /*
221      * signal the target cpu that there is work pending.
222      */
223     if (atomic_poll_acquire_int(&target->gd_npoll)) {
224 	logipiq(cpu_send, func, arg1, arg2, gd, target);
225 	cpu_send_ipiq(target->gd_cpuid);
226     } else {
227 	++ipiq_avoided;
228     }
229     --gd->gd_intr_nesting_level;
230     crit_exit();
231     logipiq(send_end, func, arg1, arg2, gd, target);
232 
233     return(ip->ip_windex);
234 }
235 
236 /*
237  * Similar to lwkt_send_ipiq() but this function does not actually initiate
238  * the IPI to the target cpu unless the FIFO has become too full, so it is
239  * very fast.
240  *
241  * This function is used for non-critical IPI messages, such as memory
242  * deallocations.  The queue will typically be flushed by the target cpu at
243  * the next clock interrupt.
244  *
245  * Need not be called from a critical section.
246  */
247 int
248 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
249 			void *arg1, int arg2)
250 {
251     lwkt_ipiq_t ip;
252     int windex;
253     struct globaldata *gd = mycpu;
254 
255     KKASSERT(target != gd);
256     crit_enter();
257     ++gd->gd_intr_nesting_level;
258     logipiq(send_pasv, func, arg1, arg2, gd, target);
259 #ifdef INVARIANTS
260     if (gd->gd_intr_nesting_level > 20)
261 	panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
262 #endif
263     KKASSERT(curthread->td_critcount);
264     ++ipiq_count;
265     ++ipiq_passive;
266     ip = &gd->gd_ipiq[target->gd_cpuid];
267 
268     /*
269      * Do not allow the FIFO to become full.  Interrupts must be physically
270      * enabled while we liveloop to avoid deadlocking the APIC.
271      */
272     if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
273 #if defined(__i386__)
274 	unsigned int eflags = read_eflags();
275 #elif defined(__x86_64__)
276 	unsigned long rflags = read_rflags();
277 #endif
278 
279 	cpu_enable_intr();
280 	++ipiq_fifofull;
281 	DEBUG_PUSH_INFO("send_ipiq3_passive");
282 	while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
283 	    if (atomic_poll_acquire_int(&target->gd_npoll)) {
284 		logipiq(cpu_send, func, arg1, arg2, gd, target);
285 		cpu_send_ipiq(target->gd_cpuid);
286 	    }
287 	    KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
288 	    lwkt_process_ipiq();
289 	    cpu_pause();
290 	}
291 	DEBUG_POP_INFO();
292 #if defined(__i386__)
293 	write_eflags(eflags);
294 #elif defined(__x86_64__)
295 	write_rflags(rflags);
296 #endif
297     }
298 
299     /*
300      * Queue the new message
301      */
302     windex = ip->ip_windex & MAXCPUFIFO_MASK;
303     ip->ip_info[windex].func = func;
304     ip->ip_info[windex].arg1 = arg1;
305     ip->ip_info[windex].arg2 = arg2;
306     cpu_sfence();
307     ++ip->ip_windex;
308     atomic_set_cpumask(&target->gd_ipimask, gd->gd_cpumask);
309     --gd->gd_intr_nesting_level;
310 
311     /*
312      * Do not signal the target cpu, it will pick up the IPI when it next
313      * polls (typically on the next tick).
314      */
315     crit_exit();
316     logipiq(send_end, func, arg1, arg2, gd, target);
317 
318     return(ip->ip_windex);
319 }
320 
321 /*
322  * Send an IPI request without blocking, return 0 on success, ENOENT on
323  * failure.  The actual queueing of the hardware IPI may still force us
324  * to spin and process incoming IPIs but that will eventually go away
325  * when we've gotten rid of the other general IPIs.
326  */
327 int
328 lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func,
329 		       void *arg1, int arg2)
330 {
331     lwkt_ipiq_t ip;
332     int windex;
333     struct globaldata *gd = mycpu;
334 
335     logipiq(send_nbio, func, arg1, arg2, gd, target);
336     KKASSERT(curthread->td_critcount);
337     if (target == gd) {
338 	func(arg1, arg2, NULL);
339 	logipiq(send_end, func, arg1, arg2, gd, target);
340 	return(0);
341     }
342     crit_enter();
343     ++gd->gd_intr_nesting_level;
344     ++ipiq_count;
345     ip = &gd->gd_ipiq[target->gd_cpuid];
346 
347     if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
348 	logipiq(send_fail, func, arg1, arg2, gd, target);
349 	--gd->gd_intr_nesting_level;
350 	crit_exit();
351 	return(ENOENT);
352     }
353     windex = ip->ip_windex & MAXCPUFIFO_MASK;
354     ip->ip_info[windex].func = func;
355     ip->ip_info[windex].arg1 = arg1;
356     ip->ip_info[windex].arg2 = arg2;
357     cpu_sfence();
358     ++ip->ip_windex;
359     atomic_set_cpumask(&target->gd_ipimask, gd->gd_cpumask);
360 
361     /*
362      * This isn't a passive IPI, we still have to signal the target cpu.
363      */
364     if (atomic_poll_acquire_int(&target->gd_npoll)) {
365 	logipiq(cpu_send, func, arg1, arg2, gd, target);
366 	cpu_send_ipiq(target->gd_cpuid);
367     } else {
368 	++ipiq_avoided;
369     }
370     --gd->gd_intr_nesting_level;
371     crit_exit();
372 
373     logipiq(send_end, func, arg1, arg2, gd, target);
374     return(0);
375 }
376 
377 /*
378  * deprecated, used only by fast int forwarding.
379  */
380 int
381 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
382 {
383     return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
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  */
390 int
391 lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2)
392 {
393     int cpuid;
394     int count = 0;
395 
396     mask &= ~stopped_cpus;
397     while (mask) {
398 	cpuid = BSFCPUMASK(mask);
399 	lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
400 	mask &= ~CPUMASK(cpuid);
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  */
417 void
418 lwkt_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) {
426 #if defined(__i386__)
427 	    unsigned int eflags = read_eflags();
428 #elif defined(__x86_64__)
429 	    unsigned long rflags = read_rflags();
430 #endif
431 	    cpu_enable_intr();
432 	    DEBUG_PUSH_INFO("wait_ipiq");
433 	    while ((int)(ip->ip_xindex - seq) < 0) {
434 		crit_enter();
435 		lwkt_process_ipiq();
436 		crit_exit();
437 		if (--maxc == 0)
438 			kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
439 		if (maxc < -1000000)
440 			panic("LWKT_WAIT_IPIQ");
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();
447 	    }
448 	    DEBUG_POP_INFO();
449 #if defined(__i386__)
450 	    write_eflags(eflags);
451 #elif defined(__x86_64__)
452 	    write_rflags(rflags);
453 #endif
454 	}
455     }
456 }
457 
458 int
459 lwkt_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 
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_info[].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.
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.
481  */
482 void
483 lwkt_process_ipiq(void)
484 {
485     globaldata_t gd = mycpu;
486     globaldata_t sgd;
487     lwkt_ipiq_t ip;
488     cpumask_t mask;
489     int n;
490 
491     ++gd->gd_processing_ipiq;
492 again:
493     cpu_lfence();
494     mask = gd->gd_ipimask;
495     atomic_clear_cpumask(&gd->gd_ipimask, mask);
496     while (mask) {
497 	n = BSFCPUMASK(mask);
498 	if (n != gd->gd_cpuid) {
499 	    sgd = globaldata_find(n);
500 	    ip = sgd->gd_ipiq;
501 	    if (ip != NULL) {
502 		while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL))
503 		    ;
504 	    }
505 	}
506 	mask &= ~CPUMASK(n);
507     }
508     if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) {
509 	if (gd->gd_curthread->td_cscount == 0)
510 	    goto again;
511 	/* need_ipiq(); do not reflag */
512     }
513 
514     /*
515      * Interlock to allow more IPI interrupts.  Recheck ipimask after
516      * releasing gd_npoll.
517      */
518     if (gd->gd_ipimask)
519 	goto again;
520     atomic_poll_release_int(&gd->gd_npoll);
521     cpu_mfence();
522     if (gd->gd_ipimask)
523 	goto again;
524     --gd->gd_processing_ipiq;
525 }
526 
527 void
528 lwkt_process_ipiq_frame(struct intrframe *frame)
529 {
530     globaldata_t gd = mycpu;
531     globaldata_t sgd;
532     lwkt_ipiq_t ip;
533     cpumask_t mask;
534     int n;
535 
536 again:
537     cpu_lfence();
538     mask = gd->gd_ipimask;
539     atomic_clear_cpumask(&gd->gd_ipimask, mask);
540     while (mask) {
541 	n = BSFCPUMASK(mask);
542 	if (n != gd->gd_cpuid) {
543 	    sgd = globaldata_find(n);
544 	    ip = sgd->gd_ipiq;
545 	    if (ip != NULL) {
546 		while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], frame))
547 		    ;
548 	    }
549 	}
550 	mask &= ~CPUMASK(n);
551     }
552     if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
553 	if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame)) {
554 	    if (gd->gd_curthread->td_cscount == 0)
555 		goto again;
556 	    /* need_ipiq(); do not reflag */
557 	}
558     }
559 
560     /*
561      * Interlock to allow more IPI interrupts.  Recheck ipimask after
562      * releasing gd_npoll.
563      */
564     if (gd->gd_ipimask)
565 	goto again;
566     atomic_poll_release_int(&gd->gd_npoll);
567     cpu_mfence();
568     if (gd->gd_ipimask)
569 	goto again;
570 }
571 
572 #if 0
573 static int iqticks[SMP_MAXCPU];
574 static int iqcount[SMP_MAXCPU];
575 #endif
576 #if 0
577 static int iqterm[SMP_MAXCPU];
578 #endif
579 
580 static int
581 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
582 		       struct intrframe *frame)
583 {
584     globaldata_t mygd = mycpu;
585     int ri;
586     int wi;
587     ipifunc3_t copy_func;
588     void *copy_arg1;
589     int copy_arg2;
590 
591 #if 0
592     if (iqticks[mygd->gd_cpuid] != ticks) {
593 	    iqticks[mygd->gd_cpuid] = ticks;
594 	    iqcount[mygd->gd_cpuid] = 0;
595     }
596     if (++iqcount[mygd->gd_cpuid] > 3000000) {
597 	kprintf("cpu %d ipiq maxed cscount %d spin %d\n",
598 		mygd->gd_cpuid,
599 		mygd->gd_curthread->td_cscount,
600 		mygd->gd_spinlocks_wr);
601 	iqcount[mygd->gd_cpuid] = 0;
602 #if 0
603 	if (++iqterm[mygd->gd_cpuid] > 10)
604 		panic("cpu %d ipiq maxed", mygd->gd_cpuid);
605 #endif
606 	int i;
607 	for (i = 0; i < ncpus; ++i) {
608 		if (globaldata_find(i)->gd_infomsg)
609 			kprintf(" %s", globaldata_find(i)->gd_infomsg);
610 	}
611 	kprintf("\n");
612     }
613 #endif
614 
615     /*
616      * Clear the originating core from our ipimask, we will process all
617      * incoming messages.
618      *
619      * Obtain the current write index, which is modified by a remote cpu.
620      * Issue a load fence to prevent speculative reads of e.g. data written
621      * by the other cpu prior to it updating the index.
622      */
623     KKASSERT(curthread->td_critcount);
624     wi = ip->ip_windex;
625     cpu_lfence();
626     ++mygd->gd_intr_nesting_level;
627 
628     /*
629      * NOTE: xindex is only updated after we are sure the function has
630      *	     finished execution.  Beware lwkt_process_ipiq() reentrancy!
631      *	     The function may send an IPI which may block/drain.
632      *
633      * NOTE: Due to additional IPI operations that the callback function
634      *	     may make, it is possible for both rindex and windex to advance and
635      *	     thus for rindex to advance passed our cached windex.
636      *
637      * NOTE: A load fence is required to prevent speculative loads prior
638      *	     to the loading of ip_rindex.  Even though stores might be
639      *	     ordered, loads are probably not.  A memory fence is required
640      *	     to prevent reordering of the loads after the ip_rindex update.
641      */
642     while (wi - (ri = ip->ip_rindex) > 0) {
643 	ri &= MAXCPUFIFO_MASK;
644 	cpu_lfence();
645 	copy_func = ip->ip_info[ri].func;
646 	copy_arg1 = ip->ip_info[ri].arg1;
647 	copy_arg2 = ip->ip_info[ri].arg2;
648 	cpu_mfence();
649 	++ip->ip_rindex;
650 	KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) ==
651 		 ((ri + 1) & MAXCPUFIFO_MASK));
652 	logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
653 #ifdef INVARIANTS
654 	if (ipiq_debug && (ip->ip_rindex & 0xFFFFFF) == 0) {
655 		kprintf("cpu %d ipifunc %p %p %d (frame %p)\n",
656 			mycpu->gd_cpuid,
657 			copy_func, copy_arg1, copy_arg2,
658 #if defined(__i386__)
659 			(frame ? (void *)frame->if_eip : NULL));
660 #elif defined(__amd64__)
661 			(frame ? (void *)frame->if_rip : NULL));
662 #else
663 			NULL);
664 #endif
665 	}
666 #endif
667 	copy_func(copy_arg1, copy_arg2, frame);
668 	cpu_sfence();
669 	ip->ip_xindex = ip->ip_rindex;
670 
671 #ifdef PANIC_DEBUG
672 	/*
673 	 * Simulate panics during the processing of an IPI
674 	 */
675 	if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
676 		if (--panic_ipiq_count == 0) {
677 #ifdef DDB
678 			Debugger("PANIC_DEBUG");
679 #else
680 			panic("PANIC_DEBUG");
681 #endif
682 		}
683 	}
684 #endif
685     }
686     --mygd->gd_intr_nesting_level;
687 
688     /*
689      * Return non-zero if there is still more in the queue.
690      */
691     cpu_lfence();
692     return (ip->ip_rindex != ip->ip_windex);
693 }
694 
695 static void
696 lwkt_sync_ipiq(void *arg)
697 {
698     volatile cpumask_t *cpumask = arg;
699 
700     atomic_clear_cpumask(cpumask, mycpu->gd_cpumask);
701     if (*cpumask == 0)
702 	wakeup(cpumask);
703 }
704 
705 void
706 lwkt_synchronize_ipiqs(const char *wmesg)
707 {
708     volatile cpumask_t other_cpumask;
709 
710     other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
711     lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq,
712     	__DEVOLATILE(void *, &other_cpumask));
713 
714     while (other_cpumask != 0) {
715 	tsleep_interlock(&other_cpumask, 0);
716 	if (other_cpumask != 0)
717 	    tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0);
718     }
719 }
720 
721 #endif
722 
723 /*
724  * CPU Synchronization Support
725  *
726  * lwkt_cpusync_interlock()	- Place specified cpus in a quiescent state.
727  *				  The current cpu is placed in a hard critical
728  *				  section.
729  *
730  * lwkt_cpusync_deinterlock()	- Execute cs_func on specified cpus, including
731  *				  current cpu if specified, then return.
732  */
733 void
734 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *arg)
735 {
736     struct lwkt_cpusync cs;
737 
738     lwkt_cpusync_init(&cs, mask, func, arg);
739     lwkt_cpusync_interlock(&cs);
740     lwkt_cpusync_deinterlock(&cs);
741 }
742 
743 
744 void
745 lwkt_cpusync_interlock(lwkt_cpusync_t cs)
746 {
747 #ifdef SMP
748 #if 0
749     const char *smsg = "SMPSYNL";
750 #endif
751     globaldata_t gd = mycpu;
752     cpumask_t mask;
753 
754     /*
755      * mask acknowledge (cs_mack):  0->mask for stage 1
756      *
757      * mack does not include the current cpu.
758      */
759     mask = cs->cs_mask & gd->gd_other_cpus & smp_active_mask;
760     cs->cs_mack = 0;
761     crit_enter_id("cpusync");
762     if (mask) {
763 	DEBUG_PUSH_INFO("cpusync_interlock");
764 	++ipiq_cscount;
765 	++gd->gd_curthread->td_cscount;
766 	lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote1, cs);
767 	logipiq2(sync_start, mask);
768 #if 0
769 	if (gd->gd_curthread->td_wmesg == NULL)
770 		gd->gd_curthread->td_wmesg = smsg;
771 #endif
772 	while (cs->cs_mack != mask) {
773 	    lwkt_process_ipiq();
774 	    cpu_pause();
775 	}
776 #if 0
777 	if (gd->gd_curthread->td_wmesg == smsg)
778 		gd->gd_curthread->td_wmesg = NULL;
779 #endif
780 	DEBUG_POP_INFO();
781     }
782 #else
783     cs->cs_mack = 0;
784 #endif
785 }
786 
787 /*
788  * Interlocked cpus have executed remote1 and are polling in remote2.
789  * To deinterlock we clear cs_mack and wait for the cpus to execute
790  * the func and set their bit in cs_mack again.
791  *
792  */
793 void
794 lwkt_cpusync_deinterlock(lwkt_cpusync_t cs)
795 {
796     globaldata_t gd = mycpu;
797 #ifdef SMP
798 #if 0
799     const char *smsg = "SMPSYNU";
800 #endif
801     cpumask_t mask;
802 
803     /*
804      * mask acknowledge (cs_mack):  mack->0->mack for stage 2
805      *
806      * Clearing cpu bits for polling cpus in cs_mack will cause them to
807      * execute stage 2, which executes the cs_func(cs_data) and then sets
808      * their bit in cs_mack again.
809      *
810      * mack does not include the current cpu.
811      */
812     mask = cs->cs_mack;
813     cpu_ccfence();
814     cs->cs_mack = 0;
815     if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
816 	    cs->cs_func(cs->cs_data);
817     if (mask) {
818 	DEBUG_PUSH_INFO("cpusync_deinterlock");
819 #if 0
820 	if (gd->gd_curthread->td_wmesg == NULL)
821 		gd->gd_curthread->td_wmesg = smsg;
822 #endif
823 	while (cs->cs_mack != mask) {
824 	    lwkt_process_ipiq();
825 	    cpu_pause();
826 	}
827 #if 0
828 	if (gd->gd_curthread->td_wmesg == smsg)
829 		gd->gd_curthread->td_wmesg = NULL;
830 #endif
831 	DEBUG_POP_INFO();
832 	/*
833 	 * cpusyncq ipis may be left queued without the RQF flag set due to
834 	 * a non-zero td_cscount, so be sure to process any laggards after
835 	 * decrementing td_cscount.
836 	 */
837 	--gd->gd_curthread->td_cscount;
838 	lwkt_process_ipiq();
839 	logipiq2(sync_end, mask);
840     }
841     crit_exit_id("cpusync");
842 #else
843     if (cs->cs_func && (cs->cs_mask & gd->gd_cpumask))
844 	cs->cs_func(cs->cs_data);
845 #endif
846 }
847 
848 #ifdef SMP
849 
850 /*
851  * helper IPI remote messaging function.
852  *
853  * Called on remote cpu when a new cpu synchronization request has been
854  * sent to us.  Execute the run function and adjust cs_count, then requeue
855  * the request so we spin on it.
856  */
857 static void
858 lwkt_cpusync_remote1(lwkt_cpusync_t cs)
859 {
860     globaldata_t gd = mycpu;
861 
862     atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
863     lwkt_cpusync_remote2(cs);
864 }
865 
866 /*
867  * helper IPI remote messaging function.
868  *
869  * Poll for the originator telling us to finish.  If it hasn't, requeue
870  * our request so we spin on it.
871  */
872 static void
873 lwkt_cpusync_remote2(lwkt_cpusync_t cs)
874 {
875     globaldata_t gd = mycpu;
876 
877     if ((cs->cs_mack & gd->gd_cpumask) == 0) {
878 	if (cs->cs_func)
879 		cs->cs_func(cs->cs_data);
880 	atomic_set_cpumask(&cs->cs_mack, gd->gd_cpumask);
881     } else {
882 	lwkt_ipiq_t ip;
883 	int wi;
884 
885 	ip = &gd->gd_cpusyncq;
886 	wi = ip->ip_windex & MAXCPUFIFO_MASK;
887 	ip->ip_info[wi].func = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
888 	ip->ip_info[wi].arg1 = cs;
889 	ip->ip_info[wi].arg2 = 0;
890 	cpu_sfence();
891 	++ip->ip_windex;
892 	if (ipiq_debug && (ip->ip_windex & 0xFFFFFF) == 0) {
893 		kprintf("cpu %d cm=%016jx %016jx f=%p\n",
894 			gd->gd_cpuid,
895 			(intmax_t)cs->cs_mask, (intmax_t)cs->cs_mack,
896 			cs->cs_func);
897 	}
898     }
899 }
900 
901 #endif
902