xref: /dragonfly/sys/kern/lwkt_ipiq.c (revision cae2835b)
1 /*
2  * Copyright (c) 2003-2016 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 
55 #include <vm/vm.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>
64 
65 #include <machine/stdarg.h>
66 #include <machine/smp.h>
67 #include <machine/clock.h>
68 #include <machine/atomic.h>
69 
70 #ifdef _KERNEL_VIRTUAL
71 #include <pthread.h>
72 #endif
73 
74 struct ipiq_stats {
75     int64_t ipiq_count;		/* total calls to lwkt_send_ipiq*() */
76     int64_t ipiq_fifofull;	/* number of fifo full conditions detected */
77     int64_t ipiq_avoided;	/* interlock with target avoids cpu ipi */
78     int64_t ipiq_passive;	/* passive IPI messages */
79     int64_t ipiq_cscount;	/* number of cpu synchronizations */
80 } __cachealign;
81 
82 static struct ipiq_stats ipiq_stats_percpu[MAXCPU];
83 #define ipiq_stat(gd)	ipiq_stats_percpu[(gd)->gd_cpuid]
84 
85 static int ipiq_debug;		/* set to 1 for debug */
86 #ifdef PANIC_DEBUG
87 static int	panic_ipiq_cpu = -1;
88 static int	panic_ipiq_count = 100;
89 #endif
90 
91 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_debug, CTLFLAG_RW, &ipiq_debug, 0,
92     "");
93 #ifdef PANIC_DEBUG
94 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, "");
95 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, "");
96 #endif
97 
98 #define IPIQ_STRING	"func=%p arg1=%p arg2=%d scpu=%d dcpu=%d"
99 #define IPIQ_ARGS	void *func, void *arg1, int arg2, int scpu, int dcpu
100 
101 #if !defined(KTR_IPIQ)
102 #define KTR_IPIQ	KTR_ALL
103 #endif
104 KTR_INFO_MASTER(ipiq);
105 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARGS);
106 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARGS);
107 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARGS);
108 KTR_INFO(KTR_IPIQ, ipiq, sync_start, 5, "cpumask=%08lx", unsigned long mask);
109 KTR_INFO(KTR_IPIQ, ipiq, sync_end, 6, "cpumask=%08lx", unsigned long mask);
110 KTR_INFO(KTR_IPIQ, ipiq, cpu_send, 7, IPIQ_STRING, IPIQ_ARGS);
111 KTR_INFO(KTR_IPIQ, ipiq, send_end, 8, IPIQ_STRING, IPIQ_ARGS);
112 KTR_INFO(KTR_IPIQ, ipiq, sync_quick, 9, "cpumask=%08lx", unsigned long mask);
113 
114 #define logipiq(name, func, arg1, arg2, sgd, dgd)	\
115 	KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid)
116 #define logipiq2(name, arg)	\
117 	KTR_LOG(ipiq_ ## name, arg)
118 
119 static void lwkt_process_ipiq_nested(void);
120 static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
121 				  struct intrframe *frame, int limit);
122 static void lwkt_cpusync_remote1(lwkt_cpusync_t cs);
123 static void lwkt_cpusync_remote2(lwkt_cpusync_t cs);
124 
125 #define IPIQ_SYSCTL(name)				\
126 static int						\
127 sysctl_##name(SYSCTL_HANDLER_ARGS)			\
128 {							\
129     int64_t val = 0;					\
130     int cpu, error;					\
131 							\
132     for (cpu = 0; cpu < ncpus; ++cpu)			\
133 	val += ipiq_stats_percpu[cpu].name;		\
134 							\
135     error = sysctl_handle_quad(oidp, &val, 0, req);	\
136     if (error || req->newptr == NULL)			\
137 	return error;					\
138 							\
139     for (cpu = 0; cpu < ncpus; ++cpu)			\
140     	ipiq_stats_percpu[cpu].name = val;		\
141 							\
142     return 0;						\
143 }
144 
145 IPIQ_SYSCTL(ipiq_count);
146 IPIQ_SYSCTL(ipiq_fifofull);
147 IPIQ_SYSCTL(ipiq_avoided);
148 IPIQ_SYSCTL(ipiq_passive);
149 IPIQ_SYSCTL(ipiq_cscount);
150 
151 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_count, (CTLTYPE_QUAD | CTLFLAG_RW),
152     0, 0, sysctl_ipiq_count, "Q", "Number of IPI's sent");
153 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_fifofull, (CTLTYPE_QUAD | CTLFLAG_RW),
154     0, 0, sysctl_ipiq_fifofull, "Q",
155     "Number of fifo full conditions detected");
156 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_avoided, (CTLTYPE_QUAD | CTLFLAG_RW),
157     0, 0, sysctl_ipiq_avoided, "Q",
158     "Number of IPI's avoided by interlock with target cpu");
159 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_passive, (CTLTYPE_QUAD | CTLFLAG_RW),
160     0, 0, sysctl_ipiq_passive, "Q",
161     "Number of passive IPI messages sent");
162 SYSCTL_PROC(_lwkt, OID_AUTO, ipiq_cscount, (CTLTYPE_QUAD | CTLFLAG_RW),
163     0, 0, sysctl_ipiq_cscount, "Q",
164     "Number of cpu synchronizations");
165 
166 /*
167  * Send a function execution request to another cpu.  The request is queued
168  * on the cpu<->cpu ipiq matrix.  Each cpu owns a unique ipiq FIFO for every
169  * possible target cpu.  The FIFO can be written.
170  *
171  * If the FIFO fills up we have to enable interrupts to avoid an APIC
172  * deadlock and process pending IPIQs while waiting for it to empty.
173  * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
174  *
175  * We can safely bump gd_intr_nesting_level because our crit_exit() at the
176  * end will take care of any pending interrupts.
177  *
178  * The actual hardware IPI is avoided if the target cpu is already processing
179  * the queue from a prior IPI.  It is possible to pipeline IPI messages
180  * very quickly between cpus due to the FIFO hysteresis.
181  *
182  * Need not be called from a critical section.
183  */
184 int
185 lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2)
186 {
187     lwkt_ipiq_t ip;
188     int windex;
189     int level1;
190     int level2;
191     long rflags;
192     struct globaldata *gd = mycpu;
193 
194     logipiq(send_norm, func, arg1, arg2, gd, target);
195 
196     if (target == gd) {
197 	func(arg1, arg2, NULL);
198 	logipiq(send_end, func, arg1, arg2, gd, target);
199 	return(0);
200     }
201     crit_enter();
202     ++gd->gd_intr_nesting_level;
203 #ifdef INVARIANTS
204     if (gd->gd_intr_nesting_level > 20)
205 	panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
206 #endif
207     KKASSERT(curthread->td_critcount);
208     ++ipiq_stat(gd).ipiq_count;
209     ip = &gd->gd_ipiq[target->gd_cpuid];
210 
211     /*
212      * Do not allow the FIFO to become full.  Interrupts must be physically
213      * enabled while we liveloop to avoid deadlocking the APIC.
214      *
215      * When we are not nested inside a processing loop we allow the FIFO
216      * to get 1/2 full.  Once it exceeds 1/2 full we must wait for it to
217      * drain, executing any incoming IPIs while we wait.
218      *
219      * When we are nested we allow the FIFO to get almost completely full.
220      * This allows us to queue IPIs sent from IPI callbacks.  The processing
221      * code will only process incoming FIFOs that are trying to drain while
222      * we wait, and only to the only-slightly-less-full point, to avoid a
223      * deadlock.
224      *
225      * We are guaranteed
226      */
227 
228     if (gd->gd_processing_ipiq == 0) {
229 	level1 = MAXCPUFIFO / 2;
230 	level2 = MAXCPUFIFO / 4;
231     } else {
232 	level1 = MAXCPUFIFO - 3;
233 	level2 = MAXCPUFIFO - 5;
234     }
235 
236     if (ip->ip_windex - ip->ip_rindex > level1) {
237 #ifndef _KERNEL_VIRTUAL
238 	uint64_t tsc_base = rdtsc();
239 #endif
240 	int repeating = 0;
241 	int olimit;
242 
243 	rflags = read_rflags();
244 	cpu_enable_intr();
245 	++ipiq_stat(gd).ipiq_fifofull;
246 	DEBUG_PUSH_INFO("send_ipiq3");
247 	olimit = atomic_swap_int(&ip->ip_drain, level2);
248 	while (ip->ip_windex - ip->ip_rindex > level2) {
249 	    KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
250 	    lwkt_process_ipiq_nested();
251 	    cpu_pause();
252 
253 	    /*
254 	     * Check for target not draining issue.  This should be fixed but
255 	     * leave the code in-place anyway as it can recover an otherwise
256 	     * dead system.
257 	     */
258 #ifdef _KERNEL_VIRTUAL
259 	    if (repeating++ > 10)
260 		    pthread_yield();
261 #else
262 	    if (rdtsc() - tsc_base > tsc_frequency) {
263 		++repeating;
264 		if (repeating > 10) {
265 			kprintf("send_ipiq %d->%d tgt not draining (%d) sniff=%p,%p\n",
266 				gd->gd_cpuid, target->gd_cpuid, repeating,
267 				target->gd_sample_pc, target->gd_sample_sp);
268 			smp_sniff();
269 			ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid);
270 			cpu_send_ipiq(target->gd_cpuid);
271 		} else {
272 			kprintf("send_ipiq %d->%d tgt not draining (%d)\n",
273 				gd->gd_cpuid, target->gd_cpuid, repeating);
274 			smp_sniff();
275 		}
276 		tsc_base = rdtsc();
277 	    }
278 #endif
279 	}
280 	atomic_swap_int(&ip->ip_drain, olimit);
281 	DEBUG_POP_INFO();
282 #if defined(__x86_64__)
283 	write_rflags(rflags);
284 #else
285 #error "no write_*flags"
286 #endif
287     }
288 
289     /*
290      * Queue the new message and signal the target cpu.  For now we need to
291      * physically disable interrupts because the target will not get signalled
292      * by other cpus once we set target->gd_npoll and we don't want to get
293      * interrupted.
294      *
295      * XXX not sure why this is a problem, the critical section should prevent
296      *     any stalls (incoming interrupts except Xinvltlb and Xsnoop will
297      *	   just be made pending).
298      */
299     rflags = read_rflags();
300     cpu_disable_intr();
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_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid);
309 
310     /*
311      * signal the target cpu that there is work pending.
312      */
313     if (atomic_swap_int(&target->gd_npoll, 1) == 0) {
314 	logipiq(cpu_send, func, arg1, arg2, gd, target);
315 	cpu_send_ipiq(target->gd_cpuid);
316     } else {
317 	++ipiq_stat(gd).ipiq_avoided;
318     }
319     write_rflags(rflags);
320 
321     --gd->gd_intr_nesting_level;
322     crit_exit();
323     logipiq(send_end, func, arg1, arg2, gd, target);
324 
325     return(ip->ip_windex);
326 }
327 
328 /*
329  * Similar to lwkt_send_ipiq() but this function does not actually initiate
330  * the IPI to the target cpu unless the FIFO is greater than 1/4 full.
331  * This function is usually very fast.
332  *
333  * This function is used for non-critical IPI messages, such as memory
334  * deallocations.  The queue will typically be flushed by the target cpu at
335  * the next clock interrupt.
336  *
337  * Need not be called from a critical section.
338  */
339 int
340 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
341 			void *arg1, int arg2)
342 {
343     lwkt_ipiq_t ip;
344     int windex;
345     struct globaldata *gd = mycpu;
346 
347     KKASSERT(target != gd);
348     crit_enter_gd(gd);
349     ++gd->gd_intr_nesting_level;
350     ip = &gd->gd_ipiq[target->gd_cpuid];
351 
352     /*
353      * If the FIFO is too full send the IPI actively.
354      *
355      * WARNING! This level must be low enough not to trigger a wait loop
356      *		in the active sending code since we are not signalling the
357      *		target cpu.
358      */
359     if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO / 4) {
360 	--gd->gd_intr_nesting_level;
361 	crit_exit_gd(gd);
362 	return lwkt_send_ipiq3(target, func, arg1, arg2);
363     }
364 
365     /*
366      * Else we can do it passively.
367      */
368     logipiq(send_pasv, func, arg1, arg2, gd, target);
369     ++ipiq_stat(gd).ipiq_count;
370     ++ipiq_stat(gd).ipiq_passive;
371 
372     /*
373      * Queue the new message
374      */
375     windex = ip->ip_windex & MAXCPUFIFO_MASK;
376     ip->ip_info[windex].func = func;
377     ip->ip_info[windex].arg1 = arg1;
378     ip->ip_info[windex].arg2 = arg2;
379     cpu_sfence();
380     ++ip->ip_windex;
381     ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid);
382     --gd->gd_intr_nesting_level;
383 
384     /*
385      * Do not signal the target cpu, it will pick up the IPI when it next
386      * polls (typically on the next tick).
387      */
388     crit_exit();
389     logipiq(send_end, func, arg1, arg2, gd, target);
390 
391     return(ip->ip_windex);
392 }
393 
394 /*
395  * deprecated, used only by fast int forwarding.
396  */
397 int
398 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
399 {
400     return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
401 }
402 
403 /*
404  * Send a message to several target cpus.  Typically used for scheduling.
405  * The message will not be sent to stopped cpus.
406  *
407  * To prevent treating low-numbered cpus as favored sons, the IPIs are
408  * issued in order starting at mycpu upward, then from 0 through mycpu.
409  * This is particularly important to prevent random scheduler pickups
410  * from favoring cpu 0.
411  */
412 int
413 lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2)
414 {
415     int cpuid;
416     int count = 0;
417     cpumask_t amask;
418 
419     CPUMASK_NANDMASK(mask, stopped_cpus);
420 
421     /*
422      * All cpus in mask which are >= mycpu
423      */
424     CPUMASK_ASSBMASK(amask, mycpu->gd_cpuid);
425     CPUMASK_INVMASK(amask);
426     CPUMASK_ANDMASK(amask, mask);
427     while (CPUMASK_TESTNZERO(amask)) {
428 	cpuid = BSFCPUMASK(amask);
429 	lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
430 	CPUMASK_NANDBIT(amask, cpuid);
431 	++count;
432     }
433 
434     /*
435      * All cpus in mask which are < mycpu
436      */
437     CPUMASK_ASSBMASK(amask, mycpu->gd_cpuid);
438     CPUMASK_ANDMASK(amask, mask);
439     while (CPUMASK_TESTNZERO(amask)) {
440 	cpuid = BSFCPUMASK(amask);
441 	lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
442 	CPUMASK_NANDBIT(amask, cpuid);
443 	++count;
444     }
445     return(count);
446 }
447 
448 /*
449  * Wait for the remote cpu to finish processing a function.
450  *
451  * YYY we have to enable interrupts and process the IPIQ while waiting
452  * for it to empty or we may deadlock with another cpu.  Create a CPU_*()
453  * function to do this!  YYY we really should 'block' here.
454  *
455  * MUST be called from a critical section.  This routine may be called
456  * from an interrupt (for example, if an interrupt wakes a foreign thread
457  * up).
458  */
459 void
460 lwkt_wait_ipiq(globaldata_t target, int seq)
461 {
462     lwkt_ipiq_t ip;
463 
464     if (target != mycpu) {
465 	ip = &mycpu->gd_ipiq[target->gd_cpuid];
466 	if ((int)(ip->ip_xindex - seq) < 0) {
467 #if defined(__x86_64__)
468 	    unsigned long rflags = read_rflags();
469 #else
470 #error "no read_*flags"
471 #endif
472 	    int64_t time_tgt = tsc_get_target(1000000000LL);
473 	    int time_loops = 10;
474 	    int benice = 0;
475 #ifdef _KERNEL_VIRTUAL
476 	    int repeating = 0;
477 #endif
478 
479 	    cpu_enable_intr();
480 	    DEBUG_PUSH_INFO("wait_ipiq");
481 	    while ((int)(ip->ip_xindex - seq) < 0) {
482 		crit_enter();
483 		lwkt_process_ipiq();
484 		crit_exit();
485 #ifdef _KERNEL_VIRTUAL
486 		if (repeating++ > 10)
487 			pthread_yield();
488 #endif
489 
490 		/*
491 		 * IPIQs must be handled within 10 seconds and this code
492 		 * will warn after one second.
493 		 */
494 		if ((benice & 255) == 0 && tsc_test_target(time_tgt) > 0) {
495 			kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n",
496 				mycpu->gd_cpuid, target->gd_cpuid,
497 				ip->ip_xindex - seq);
498 			if (--time_loops == 0)
499 				panic("LWKT_WAIT_IPIQ");
500 			time_tgt = tsc_get_target(1000000000LL);
501 		}
502 		++benice;
503 
504 		/*
505 		 * xindex may be modified by another cpu, use a load fence
506 		 * to ensure that the loop does not use a speculative value
507 		 * (which may improve performance).
508 		 */
509 		cpu_pause();
510 		cpu_lfence();
511 	    }
512 	    DEBUG_POP_INFO();
513 #if defined(__x86_64__)
514 	    write_rflags(rflags);
515 #else
516 #error "no write_*flags"
517 #endif
518 	}
519     }
520 }
521 
522 /*
523  * Called from IPI interrupt (like a fast interrupt), which has placed
524  * us in a critical section.  The MP lock may or may not be held.
525  * May also be called from doreti or splz, or be reentrantly called
526  * indirectly through the ip_info[].func we run.
527  *
528  * There are two versions, one where no interrupt frame is available (when
529  * called from the send code and from splz, and one where an interrupt
530  * frame is available.
531  *
532  * When the current cpu is mastering a cpusync we do NOT internally loop
533  * on the cpusyncq poll.  We also do not re-flag a pending ipi due to
534  * the cpusyncq poll because this can cause doreti/splz to loop internally.
535  * The cpusync master's own loop must be allowed to run to avoid a deadlock.
536  */
537 void
538 lwkt_process_ipiq(void)
539 {
540     globaldata_t gd = mycpu;
541     globaldata_t sgd;
542     lwkt_ipiq_t ip;
543     cpumask_t mask;
544     int n;
545 
546     ++gd->gd_processing_ipiq;
547 again:
548     mask = gd->gd_ipimask;
549     cpu_ccfence();
550     while (CPUMASK_TESTNZERO(mask)) {
551 	n = BSFCPUMASK(mask);
552 	if (n != gd->gd_cpuid) {
553 	    sgd = globaldata_find(n);
554 	    ip = sgd->gd_ipiq;
555 	    if (ip != NULL) {
556 		ip += gd->gd_cpuid;
557 		while (lwkt_process_ipiq_core(sgd, ip, NULL, 0))
558 		    ;
559 		ATOMIC_CPUMASK_NANDBIT(gd->gd_ipimask, n);
560 		if (ip->ip_rindex != ip->ip_windex)
561 			ATOMIC_CPUMASK_ORBIT(gd->gd_ipimask, n);
562 	    }
563 	}
564 	CPUMASK_NANDBIT(mask, n);
565     }
566 
567     /*
568      * Process pending cpusyncs.  If the current thread has a cpusync
569      * active cpusync we only run the list once and do not re-flag
570      * as the thread itself is processing its interlock.
571      */
572     if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL, 0)) {
573 	if (gd->gd_curthread->td_cscount == 0)
574 	    goto again;
575 	/* need_ipiq(); do not reflag */
576     }
577 
578     /*
579      * Interlock to allow more IPI interrupts.
580      */
581     --gd->gd_processing_ipiq;
582 }
583 
584 void
585 lwkt_process_ipiq_frame(struct intrframe *frame)
586 {
587     globaldata_t gd = mycpu;
588     globaldata_t sgd;
589     lwkt_ipiq_t ip;
590     cpumask_t mask;
591     int n;
592 
593     ++gd->gd_processing_ipiq;
594 again:
595     mask = gd->gd_ipimask;
596     cpu_ccfence();
597     while (CPUMASK_TESTNZERO(mask)) {
598 	n = BSFCPUMASK(mask);
599 	if (n != gd->gd_cpuid) {
600 	    sgd = globaldata_find(n);
601 	    ip = sgd->gd_ipiq;
602 	    if (ip != NULL) {
603 		ip += gd->gd_cpuid;
604 		while (lwkt_process_ipiq_core(sgd, ip, frame, 0))
605 		    ;
606 		ATOMIC_CPUMASK_NANDBIT(gd->gd_ipimask, n);
607 		if (ip->ip_rindex != ip->ip_windex)
608 			ATOMIC_CPUMASK_ORBIT(gd->gd_ipimask, n);
609 	    }
610 	}
611 	CPUMASK_NANDBIT(mask, n);
612     }
613     if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
614 	if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame, 0)) {
615 	    if (gd->gd_curthread->td_cscount == 0)
616 		goto again;
617 	    /* need_ipiq(); do not reflag */
618 	}
619     }
620     --gd->gd_processing_ipiq;
621 }
622 
623 /*
624  * Only process incoming IPIQs from draining senders and only process them
625  * to the point where the draining sender is able to continue.  This is
626  * necessary to avoid deadlocking the IPI subsystem because we are acting on
627  * incoming messages and the callback may queue additional messages.
628  *
629  * We only want to have to act on senders that are blocked to limit the
630  * number of additional messages sent.  At the same time, recipients are
631  * trying to drain our own queue.  Theoretically this create a pipeline that
632  * cannot deadlock.
633  */
634 static void
635 lwkt_process_ipiq_nested(void)
636 {
637     globaldata_t gd = mycpu;
638     globaldata_t sgd;
639     lwkt_ipiq_t ip;
640     cpumask_t mask;
641     int n;
642     int limit;
643 
644     ++gd->gd_processing_ipiq;
645 again:
646     mask = gd->gd_ipimask;
647     cpu_ccfence();
648     while (CPUMASK_TESTNZERO(mask)) {
649 	n = BSFCPUMASK(mask);
650 	if (n != gd->gd_cpuid) {
651 	    sgd = globaldata_find(n);
652 	    ip = sgd->gd_ipiq;
653 
654 	    /*
655 	     * NOTE: We do not mess with the cpumask at all, instead we allow
656 	     *	     the top-level ipiq processor deal with it.
657 	     */
658 	    if (ip != NULL) {
659 		ip += gd->gd_cpuid;
660 		if ((limit = ip->ip_drain) != 0) {
661 		    lwkt_process_ipiq_core(sgd, ip, NULL, limit);
662 		    /* no gd_ipimask when doing limited processing */
663 		}
664 	    }
665 	}
666 	CPUMASK_NANDBIT(mask, n);
667     }
668 
669     /*
670      * Process pending cpusyncs.  If the current thread has a cpusync
671      * active cpusync we only run the list once and do not re-flag
672      * as the thread itself is processing its interlock.
673      */
674     if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL, 0)) {
675 	if (gd->gd_curthread->td_cscount == 0)
676 	    goto again;
677 	/* need_ipiq(); do not reflag */
678     }
679     --gd->gd_processing_ipiq;
680 }
681 
682 /*
683  * Process incoming IPI requests until only <limit> are left (0 to exhaust
684  * all incoming IPI requests).
685  */
686 static int
687 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
688 		       struct intrframe *frame, int limit)
689 {
690     globaldata_t mygd = mycpu;
691     int ri;
692     int wi;
693     ipifunc3_t copy_func;
694     void *copy_arg1;
695     int copy_arg2;
696 
697     /*
698      * Clear the originating core from our ipimask, we will process all
699      * incoming messages.
700      *
701      * Obtain the current write index, which is modified by a remote cpu.
702      * Issue a load fence to prevent speculative reads of e.g. data written
703      * by the other cpu prior to it updating the index.
704      */
705     KKASSERT(curthread->td_critcount);
706     wi = ip->ip_windex;
707     cpu_lfence();
708     ++mygd->gd_intr_nesting_level;
709 
710     /*
711      * NOTE: xindex is only updated after we are sure the function has
712      *	     finished execution.  Beware lwkt_process_ipiq() reentrancy!
713      *	     The function may send an IPI which may block/drain.
714      *
715      * NOTE: Due to additional IPI operations that the callback function
716      *	     may make, it is possible for both rindex and windex to advance and
717      *	     thus for rindex to advance passed our cached windex.
718      *
719      * NOTE: A load fence is required to prevent speculative loads prior
720      *	     to the loading of ip_rindex.  Even though stores might be
721      *	     ordered, loads are probably not.  A memory fence is required
722      *	     to prevent reordering of the loads after the ip_rindex update.
723      *
724      * NOTE: Single pass only.  Returns non-zero if the queue is not empty
725      *	     on return.
726      */
727     while (wi - (ri = ip->ip_rindex) > limit) {
728 	ri &= MAXCPUFIFO_MASK;
729 	cpu_lfence();
730 	copy_func = ip->ip_info[ri].func;
731 	copy_arg1 = ip->ip_info[ri].arg1;
732 	copy_arg2 = ip->ip_info[ri].arg2;
733 	cpu_mfence();
734 	++ip->ip_rindex;
735 	KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) ==
736 		 ((ri + 1) & MAXCPUFIFO_MASK));
737 	logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
738 #ifdef INVARIANTS
739 	if (ipiq_debug && (ip->ip_rindex & 0xFFFFFF) == 0) {
740 		kprintf("cpu %d ipifunc %p %p %d (frame %p)\n",
741 			mycpu->gd_cpuid,
742 			copy_func, copy_arg1, copy_arg2,
743 #if defined(__x86_64__)
744 			(frame ? (void *)frame->if_rip : NULL));
745 #else
746 			NULL);
747 #endif
748 	}
749 #endif
750 	copy_func(copy_arg1, copy_arg2, frame);
751 	cpu_sfence();
752 	ip->ip_xindex = ip->ip_rindex;
753 
754 #ifdef PANIC_DEBUG
755 	/*
756 	 * Simulate panics during the processing of an IPI
757 	 */
758 	if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
759 		if (--panic_ipiq_count == 0) {
760 #ifdef DDB
761 			Debugger("PANIC_DEBUG");
762 #else
763 			panic("PANIC_DEBUG");
764 #endif
765 		}
766 	}
767 #endif
768     }
769     --mygd->gd_intr_nesting_level;
770 
771     /*
772      * Return non-zero if there is still more in the queue.  Don't worry
773      * about fencing, we will get another interrupt if necessary.
774      */
775     return (ip->ip_rindex != ip->ip_windex);
776 }
777 
778 static void
779 lwkt_sync_ipiq(void *arg)
780 {
781     volatile cpumask_t *cpumask = arg;
782 
783     ATOMIC_CPUMASK_NANDBIT(*cpumask, mycpu->gd_cpuid);
784     if (CPUMASK_TESTZERO(*cpumask))
785 	wakeup(cpumask);
786 }
787 
788 void
789 lwkt_synchronize_ipiqs(const char *wmesg)
790 {
791     volatile cpumask_t other_cpumask;
792 
793     other_cpumask = smp_active_mask;
794     CPUMASK_ANDMASK(other_cpumask, mycpu->gd_other_cpus);
795     lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq,
796 			__DEVOLATILE(void *, &other_cpumask));
797 
798     while (CPUMASK_TESTNZERO(other_cpumask)) {
799 	tsleep_interlock(&other_cpumask, 0);
800 	if (CPUMASK_TESTNZERO(other_cpumask))
801 	    tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0);
802     }
803 }
804 
805 /*
806  * CPU Synchronization Support
807  *
808  * lwkt_cpusync_interlock()	- Place specified cpus in a quiescent state.
809  *				  The current cpu is placed in a hard critical
810  *				  section.
811  *
812  * lwkt_cpusync_deinterlock()	- Execute cs_func on specified cpus, including
813  *				  current cpu if specified, then return.
814  */
815 void
816 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *arg)
817 {
818     struct lwkt_cpusync cs;
819 
820     lwkt_cpusync_init(&cs, mask, func, arg);
821     lwkt_cpusync_interlock(&cs);
822     lwkt_cpusync_deinterlock(&cs);
823 }
824 
825 
826 void
827 lwkt_cpusync_interlock(lwkt_cpusync_t cs)
828 {
829     globaldata_t gd = mycpu;
830     cpumask_t mask;
831 
832     /*
833      * mask acknowledge (cs_mack):  0->mask for stage 1
834      *
835      * mack does not include the current cpu.
836      */
837     mask = cs->cs_mask;
838     CPUMASK_ANDMASK(mask, gd->gd_other_cpus);
839     CPUMASK_ANDMASK(mask, smp_active_mask);
840     CPUMASK_ASSZERO(cs->cs_mack);
841 
842     crit_enter_id("cpusync");
843     if (CPUMASK_TESTNZERO(mask)) {
844 	DEBUG_PUSH_INFO("cpusync_interlock");
845 	++ipiq_stat(gd).ipiq_cscount;
846 	++gd->gd_curthread->td_cscount;
847 	lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote1, cs);
848 	logipiq2(sync_start, (long)CPUMASK_LOWMASK(mask));
849 	while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) {
850 	    lwkt_process_ipiq();
851 	    cpu_pause();
852 #ifdef _KERNEL_VIRTUAL
853 	    pthread_yield();
854 #endif
855 	}
856 	DEBUG_POP_INFO();
857     }
858 }
859 
860 /*
861  * Interlocked cpus have executed remote1 and are polling in remote2.
862  * To deinterlock we clear cs_mack and wait for the cpus to execute
863  * the func and set their bit in cs_mack again.
864  *
865  */
866 void
867 lwkt_cpusync_deinterlock(lwkt_cpusync_t cs)
868 {
869     globaldata_t gd = mycpu;
870     cpumask_t mask;
871 
872     /*
873      * mask acknowledge (cs_mack):  mack->0->mack for stage 2
874      *
875      * Clearing cpu bits for polling cpus in cs_mack will cause them to
876      * execute stage 2, which executes the cs_func(cs_data) and then sets
877      * their bit in cs_mack again.
878      *
879      * mack does not include the current cpu.
880      */
881     mask = cs->cs_mack;
882     cpu_ccfence();
883     CPUMASK_ASSZERO(cs->cs_mack);
884     cpu_ccfence();
885     if (cs->cs_func && CPUMASK_TESTBIT(cs->cs_mask, gd->gd_cpuid))
886 	    cs->cs_func(cs->cs_data);
887     if (CPUMASK_TESTNZERO(mask)) {
888 	DEBUG_PUSH_INFO("cpusync_deinterlock");
889 	while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) {
890 	    lwkt_process_ipiq();
891 	    cpu_pause();
892 #ifdef _KERNEL_VIRTUAL
893 	    pthread_yield();
894 #endif
895 	}
896 	DEBUG_POP_INFO();
897 	/*
898 	 * cpusyncq ipis may be left queued without the RQF flag set due to
899 	 * a non-zero td_cscount, so be sure to process any laggards after
900 	 * decrementing td_cscount.
901 	 */
902 	--gd->gd_curthread->td_cscount;
903 	lwkt_process_ipiq();
904 	logipiq2(sync_end, (long)CPUMASK_LOWMASK(mask));
905     }
906     crit_exit_id("cpusync");
907 }
908 
909 /*
910  * The quick version does not quiesce the target cpu(s) but instead executes
911  * the function on the target cpu(s) and waits for all to acknowledge.  This
912  * avoids spinning on the target cpus.
913  *
914  * This function is typically only used for kernel_pmap updates.  User pmaps
915  * have to be quiesced.
916  */
917 void
918 lwkt_cpusync_quick(lwkt_cpusync_t cs)
919 {
920     globaldata_t gd = mycpu;
921     cpumask_t mask;
922 
923     /*
924      * stage-2 cs_mack only.
925      */
926     mask = cs->cs_mask;
927     CPUMASK_ANDMASK(mask, gd->gd_other_cpus);
928     CPUMASK_ANDMASK(mask, smp_active_mask);
929     CPUMASK_ASSZERO(cs->cs_mack);
930 
931     crit_enter_id("cpusync");
932     if (CPUMASK_TESTNZERO(mask)) {
933 	DEBUG_PUSH_INFO("cpusync_interlock");
934 	++ipiq_stat(gd).ipiq_cscount;
935 	++gd->gd_curthread->td_cscount;
936 	lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote2, cs);
937 	logipiq2(sync_quick, (long)CPUMASK_LOWMASK(mask));
938 	while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) {
939 	    lwkt_process_ipiq();
940 	    cpu_pause();
941 #ifdef _KERNEL_VIRTUAL
942 	    pthread_yield();
943 #endif
944 	}
945 
946 	/*
947 	 * cpusyncq ipis may be left queued without the RQF flag set due to
948 	 * a non-zero td_cscount, so be sure to process any laggards after
949 	 * decrementing td_cscount.
950 	 */
951 	DEBUG_POP_INFO();
952 	--gd->gd_curthread->td_cscount;
953 	lwkt_process_ipiq();
954     }
955     if (cs->cs_func && CPUMASK_TESTBIT(cs->cs_mask, gd->gd_cpuid))
956 	    cs->cs_func(cs->cs_data);
957     crit_exit_id("cpusync");
958 }
959 
960 /*
961  * helper IPI remote messaging function.
962  *
963  * Called on remote cpu when a new cpu synchronization request has been
964  * sent to us.  Execute the run function and adjust cs_count, then requeue
965  * the request so we spin on it.
966  */
967 static void
968 lwkt_cpusync_remote1(lwkt_cpusync_t cs)
969 {
970     globaldata_t gd = mycpu;
971 
972     ATOMIC_CPUMASK_ORBIT(cs->cs_mack, gd->gd_cpuid);
973     lwkt_cpusync_remote2(cs);
974 }
975 
976 /*
977  * helper IPI remote messaging function.
978  *
979  * Poll for the originator telling us to finish.  If it hasn't, requeue
980  * our request so we spin on it.
981  */
982 static void
983 lwkt_cpusync_remote2(lwkt_cpusync_t cs)
984 {
985     globaldata_t gd = mycpu;
986 
987     if (CPUMASK_TESTMASK(cs->cs_mack, gd->gd_cpumask) == 0) {
988 	if (cs->cs_func)
989 		cs->cs_func(cs->cs_data);
990 	ATOMIC_CPUMASK_ORBIT(cs->cs_mack, gd->gd_cpuid);
991 	/* cs can be ripped out at this point */
992     } else {
993 	lwkt_ipiq_t ip;
994 	int wi;
995 
996 	cpu_pause();
997 #ifdef _KERNEL_VIRTUAL
998 	pthread_yield();
999 #endif
1000 	cpu_lfence();
1001 
1002 	/*
1003 	 * Requeue our IPI to avoid a deep stack recursion.  If no other
1004 	 * IPIs are pending we can just loop up, which should help VMs
1005 	 * better-detect spin loops.
1006 	 */
1007 	ip = &gd->gd_cpusyncq;
1008 
1009 	wi = ip->ip_windex & MAXCPUFIFO_MASK;
1010 	ip->ip_info[wi].func = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
1011 	ip->ip_info[wi].arg1 = cs;
1012 	ip->ip_info[wi].arg2 = 0;
1013 	cpu_sfence();
1014 	KKASSERT(ip->ip_windex - ip->ip_rindex < MAXCPUFIFO);
1015 	++ip->ip_windex;
1016 	if (ipiq_debug && (ip->ip_windex & 0xFFFFFF) == 0) {
1017 		kprintf("cpu %d cm=%016jx %016jx f=%p\n",
1018 			gd->gd_cpuid,
1019 			(intmax_t)CPUMASK_LOWMASK(cs->cs_mask),
1020 			(intmax_t)CPUMASK_LOWMASK(cs->cs_mack),
1021 			cs->cs_func);
1022 	}
1023     }
1024 }
1025 
1026 #define LWKT_IPIQ_NLATENCY	8
1027 #define LWKT_IPIQ_NLATENCY_MASK	(LWKT_IPIQ_NLATENCY - 1)
1028 
1029 struct lwkt_ipiq_latency_log {
1030 	int		idx;	/* unmasked index */
1031 	int		pad;
1032 	uint64_t	latency[LWKT_IPIQ_NLATENCY];
1033 };
1034 
1035 static struct lwkt_ipiq_latency_log	lwkt_ipiq_latency_logs[MAXCPU];
1036 static uint64_t save_tsc;
1037 
1038 /*
1039  * IPI callback (already in a critical section)
1040  */
1041 static void
1042 lwkt_ipiq_latency_testfunc(void *arg __unused)
1043 {
1044 	uint64_t delta_tsc;
1045 	struct globaldata *gd;
1046 	struct lwkt_ipiq_latency_log *lat;
1047 
1048 	/*
1049 	 * Get delta TSC (assume TSCs are synchronized) as quickly as
1050 	 * possible and then convert to nanoseconds.
1051 	 */
1052 	delta_tsc = rdtsc_ordered() - save_tsc;
1053 	delta_tsc = delta_tsc * 1000000000LU / tsc_frequency;
1054 
1055 	/*
1056 	 * Record in our save array.
1057 	 */
1058 	gd = mycpu;
1059 	lat = &lwkt_ipiq_latency_logs[gd->gd_cpuid];
1060 	lat->latency[lat->idx & LWKT_IPIQ_NLATENCY_MASK] = delta_tsc;
1061 	++lat->idx;
1062 }
1063 
1064 /*
1065  * Send IPI from cpu0 to other cpus
1066  *
1067  * NOTE: Machine must be idle for test to run dependably, and also probably
1068  *	 a good idea not to be running powerd.
1069  *
1070  * NOTE: Caller should use 'usched :1 <command>' to lock itself to cpu 0.
1071  *	 See 'ipitest' script in /usr/src/test/sysperf/ipitest
1072  */
1073 static int
1074 lwkt_ipiq_latency_test(SYSCTL_HANDLER_ARGS)
1075 {
1076 	struct globaldata *gd;
1077 	int cpu = 0, orig_cpu, error;
1078 
1079 	error = sysctl_handle_int(oidp, &cpu, arg2, req);
1080 	if (error || req->newptr == NULL)
1081 		return error;
1082 
1083 	if (cpu == 0)
1084 		return 0;
1085 	else if (cpu >= ncpus || cpu < 0)
1086 		return EINVAL;
1087 
1088 	orig_cpu = mycpuid;
1089 	lwkt_migratecpu(0);
1090 
1091 	gd = globaldata_find(cpu);
1092 
1093 	save_tsc = rdtsc_ordered();
1094 	lwkt_send_ipiq(gd, lwkt_ipiq_latency_testfunc, NULL);
1095 
1096 	lwkt_migratecpu(orig_cpu);
1097 	return 0;
1098 }
1099 
1100 SYSCTL_NODE(_debug, OID_AUTO, ipiq, CTLFLAG_RW, 0, "");
1101 SYSCTL_PROC(_debug_ipiq, OID_AUTO, latency_test, CTLTYPE_INT | CTLFLAG_RW,
1102     NULL, 0, lwkt_ipiq_latency_test, "I",
1103     "ipi latency test, arg: remote cpuid");
1104 
1105 static int
1106 lwkt_ipiq_latency(SYSCTL_HANDLER_ARGS)
1107 {
1108 	struct lwkt_ipiq_latency_log *latency = arg1;
1109 	uint64_t lat[LWKT_IPIQ_NLATENCY];
1110 	int i;
1111 
1112 	for (i = 0; i < LWKT_IPIQ_NLATENCY; ++i)
1113 		lat[i] = latency->latency[i];
1114 
1115 	return sysctl_handle_opaque(oidp, lat, sizeof(lat), req);
1116 }
1117 
1118 static void
1119 lwkt_ipiq_latency_init(void *dummy __unused)
1120 {
1121 	int cpu;
1122 
1123 	for (cpu = 0; cpu < ncpus; ++cpu) {
1124 		char name[32];
1125 
1126 		ksnprintf(name, sizeof(name), "latency%d", cpu);
1127 		SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_debug_ipiq),
1128 		    OID_AUTO, name, CTLTYPE_OPAQUE | CTLFLAG_RD,
1129 		    &lwkt_ipiq_latency_logs[cpu], 0, lwkt_ipiq_latency,
1130 		    "LU", "7 latest ipi latency measurement results");
1131 	}
1132 }
1133 SYSINIT(lwkt_ipiq_latency, SI_SUB_CONFIGURE, SI_ORDER_ANY,
1134     lwkt_ipiq_latency_init, NULL);
1135