xref: /dflybsd-src/sys/kern/lwkt_ipiq.c (revision e586f31ca9899b49a4fc156613d9ecd853defcec)
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 			cpu_disable_intr();
270 			ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid);
271 			cpu_send_ipiq(target->gd_cpuid);
272 			cpu_enable_intr();
273 		} else {
274 			kprintf("send_ipiq %d->%d tgt not draining (%d)\n",
275 				gd->gd_cpuid, target->gd_cpuid, repeating);
276 			smp_sniff();
277 		}
278 		tsc_base = rdtsc();
279 	    }
280 #endif
281 	}
282 	atomic_swap_int(&ip->ip_drain, olimit);
283 	DEBUG_POP_INFO();
284 #if defined(__x86_64__)
285 	write_rflags(rflags);
286 #else
287 #error "no write_*flags"
288 #endif
289     }
290 
291     /*
292      * Queue the new message and signal the target cpu.  For now we need to
293      * physically disable interrupts because the target will not get signalled
294      * by other cpus once we set target->gd_npoll and we don't want to get
295      * interrupted.
296      *
297      * XXX not sure why this is a problem, the critical section should prevent
298      *     any stalls (incoming interrupts except Xinvltlb and Xsnoop will
299      *	   just be made pending).
300      */
301     rflags = read_rflags();
302 #ifndef _KERNEL_VIRTUAL
303     cpu_disable_intr();
304 #endif
305 
306     windex = ip->ip_windex & MAXCPUFIFO_MASK;
307     ip->ip_info[windex].func = func;
308     ip->ip_info[windex].arg1 = arg1;
309     ip->ip_info[windex].arg2 = arg2;
310     cpu_sfence();
311     ++ip->ip_windex;
312     ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid);
313 
314     /*
315      * signal the target cpu that there is work pending.
316      */
317     if (atomic_swap_int(&target->gd_npoll, 1) == 0) {
318 	logipiq(cpu_send, func, arg1, arg2, gd, target);
319 	cpu_send_ipiq(target->gd_cpuid);
320     } else {
321 	++ipiq_stat(gd).ipiq_avoided;
322     }
323     write_rflags(rflags);
324 
325     --gd->gd_intr_nesting_level;
326     crit_exit();
327     logipiq(send_end, func, arg1, arg2, gd, target);
328 
329     return(ip->ip_windex);
330 }
331 
332 /*
333  * Similar to lwkt_send_ipiq() but this function does not actually initiate
334  * the IPI to the target cpu unless the FIFO is greater than 1/4 full.
335  * This function is usually very fast.
336  *
337  * This function is used for non-critical IPI messages, such as memory
338  * deallocations.  The queue will typically be flushed by the target cpu at
339  * the next clock interrupt.
340  *
341  * Need not be called from a critical section.
342  */
343 int
344 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
345 			void *arg1, int arg2)
346 {
347     lwkt_ipiq_t ip;
348     int windex;
349     struct globaldata *gd = mycpu;
350 
351     KKASSERT(target != gd);
352     crit_enter_gd(gd);
353     ++gd->gd_intr_nesting_level;
354     ip = &gd->gd_ipiq[target->gd_cpuid];
355 
356     /*
357      * If the FIFO is too full send the IPI actively.
358      *
359      * WARNING! This level must be low enough not to trigger a wait loop
360      *		in the active sending code since we are not signalling the
361      *		target cpu.
362      */
363     if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO / 4) {
364 	--gd->gd_intr_nesting_level;
365 	crit_exit_gd(gd);
366 	return lwkt_send_ipiq3(target, func, arg1, arg2);
367     }
368 
369     /*
370      * Else we can do it passively.
371      */
372     logipiq(send_pasv, func, arg1, arg2, gd, target);
373     ++ipiq_stat(gd).ipiq_count;
374     ++ipiq_stat(gd).ipiq_passive;
375 
376     /*
377      * Queue the new message
378      */
379     windex = ip->ip_windex & MAXCPUFIFO_MASK;
380     ip->ip_info[windex].func = func;
381     ip->ip_info[windex].arg1 = arg1;
382     ip->ip_info[windex].arg2 = arg2;
383     cpu_sfence();
384     ++ip->ip_windex;
385     ATOMIC_CPUMASK_ORBIT(target->gd_ipimask, gd->gd_cpuid);
386     --gd->gd_intr_nesting_level;
387 
388     /*
389      * Do not signal the target cpu, it will pick up the IPI when it next
390      * polls (typically on the next tick).
391      */
392     crit_exit();
393     logipiq(send_end, func, arg1, arg2, gd, target);
394 
395     return(ip->ip_windex);
396 }
397 
398 /*
399  * deprecated, used only by fast int forwarding.
400  */
401 int
402 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
403 {
404     return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
405 }
406 
407 /*
408  * Send a message to several target cpus.  Typically used for scheduling.
409  * The message will not be sent to stopped cpus.
410  *
411  * To prevent treating low-numbered cpus as favored sons, the IPIs are
412  * issued in order starting at mycpu upward, then from 0 through mycpu.
413  * This is particularly important to prevent random scheduler pickups
414  * from favoring cpu 0.
415  */
416 int
417 lwkt_send_ipiq3_mask(cpumask_t mask, ipifunc3_t func, void *arg1, int arg2)
418 {
419     int cpuid;
420     int count = 0;
421     cpumask_t amask;
422 
423     CPUMASK_NANDMASK(mask, stopped_cpus);
424 
425     /*
426      * All cpus in mask which are >= mycpu
427      */
428     CPUMASK_ASSBMASK(amask, mycpu->gd_cpuid);
429     CPUMASK_INVMASK(amask);
430     CPUMASK_ANDMASK(amask, mask);
431     while (CPUMASK_TESTNZERO(amask)) {
432 	cpuid = BSFCPUMASK(amask);
433 	lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
434 	CPUMASK_NANDBIT(amask, cpuid);
435 	++count;
436     }
437 
438     /*
439      * All cpus in mask which are < mycpu
440      */
441     CPUMASK_ASSBMASK(amask, mycpu->gd_cpuid);
442     CPUMASK_ANDMASK(amask, mask);
443     while (CPUMASK_TESTNZERO(amask)) {
444 	cpuid = BSFCPUMASK(amask);
445 	lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
446 	CPUMASK_NANDBIT(amask, cpuid);
447 	++count;
448     }
449     return(count);
450 }
451 
452 /*
453  * Wait for the remote cpu to finish processing a function.
454  *
455  * YYY we have to enable interrupts and process the IPIQ while waiting
456  * for it to empty or we may deadlock with another cpu.  Create a CPU_*()
457  * function to do this!  YYY we really should 'block' here.
458  *
459  * MUST be called from a critical section.  This routine may be called
460  * from an interrupt (for example, if an interrupt wakes a foreign thread
461  * up).
462  */
463 void
464 lwkt_wait_ipiq(globaldata_t target, int seq)
465 {
466     lwkt_ipiq_t ip;
467 
468     if (target != mycpu) {
469 	ip = &mycpu->gd_ipiq[target->gd_cpuid];
470 	if ((int)(ip->ip_xindex - seq) < 0) {
471 #if defined(__x86_64__)
472 	    unsigned long rflags = read_rflags();
473 #else
474 #error "no read_*flags"
475 #endif
476 	    int64_t time_tgt = tsc_get_target(1000000000LL);
477 	    int time_loops = 10;
478 	    int benice = 0;
479 #ifdef _KERNEL_VIRTUAL
480 	    int repeating = 0;
481 #endif
482 
483 	    cpu_enable_intr();
484 	    DEBUG_PUSH_INFO("wait_ipiq");
485 	    while ((int)(ip->ip_xindex - seq) < 0) {
486 		crit_enter();
487 		lwkt_process_ipiq();
488 		crit_exit();
489 #ifdef _KERNEL_VIRTUAL
490 		if (repeating++ > 10)
491 			pthread_yield();
492 #endif
493 
494 		/*
495 		 * IPIQs must be handled within 10 seconds and this code
496 		 * will warn after one second.
497 		 */
498 		if ((benice & 255) == 0 && tsc_test_target(time_tgt) > 0) {
499 			kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n",
500 				mycpu->gd_cpuid, target->gd_cpuid,
501 				ip->ip_xindex - seq);
502 			if (--time_loops == 0)
503 				panic("LWKT_WAIT_IPIQ");
504 			time_tgt = tsc_get_target(1000000000LL);
505 		}
506 		++benice;
507 
508 		/*
509 		 * xindex may be modified by another cpu, use a load fence
510 		 * to ensure that the loop does not use a speculative value
511 		 * (which may improve performance).
512 		 */
513 		cpu_pause();
514 		cpu_lfence();
515 	    }
516 	    DEBUG_POP_INFO();
517 #if defined(__x86_64__)
518 	    write_rflags(rflags);
519 #else
520 #error "no write_*flags"
521 #endif
522 	}
523     }
524 }
525 
526 /*
527  * Called from IPI interrupt (like a fast interrupt), which has placed
528  * us in a critical section.  The MP lock may or may not be held.
529  * May also be called from doreti or splz, or be reentrantly called
530  * indirectly through the ip_info[].func we run.
531  *
532  * There are two versions, one where no interrupt frame is available (when
533  * called from the send code and from splz, and one where an interrupt
534  * frame is available.
535  *
536  * When the current cpu is mastering a cpusync we do NOT internally loop
537  * on the cpusyncq poll.  We also do not re-flag a pending ipi due to
538  * the cpusyncq poll because this can cause doreti/splz to loop internally.
539  * The cpusync master's own loop must be allowed to run to avoid a deadlock.
540  */
541 void
542 lwkt_process_ipiq(void)
543 {
544     globaldata_t gd = mycpu;
545     globaldata_t sgd;
546     lwkt_ipiq_t ip;
547     cpumask_t mask;
548     int n;
549 
550     ++gd->gd_processing_ipiq;
551 again:
552     mask = gd->gd_ipimask;
553     cpu_ccfence();
554     while (CPUMASK_TESTNZERO(mask)) {
555 	n = BSFCPUMASK(mask);
556 	if (n != gd->gd_cpuid) {
557 	    sgd = globaldata_find(n);
558 	    ip = sgd->gd_ipiq;
559 	    if (ip != NULL) {
560 		ip += gd->gd_cpuid;
561 		while (lwkt_process_ipiq_core(sgd, ip, NULL, 0))
562 		    ;
563 		ATOMIC_CPUMASK_NANDBIT(gd->gd_ipimask, n);
564 		if (ip->ip_rindex != ip->ip_windex)
565 			ATOMIC_CPUMASK_ORBIT(gd->gd_ipimask, n);
566 	    }
567 	}
568 	CPUMASK_NANDBIT(mask, n);
569     }
570 
571     /*
572      * Process pending cpusyncs.  If the current thread has a cpusync
573      * active cpusync we only run the list once and do not re-flag
574      * as the thread itself is processing its interlock.
575      */
576     if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL, 0)) {
577 	if (gd->gd_curthread->td_cscount == 0)
578 	    goto again;
579 	/* need_ipiq(); do not reflag */
580     }
581 
582     /*
583      * Interlock to allow more IPI interrupts.
584      */
585     --gd->gd_processing_ipiq;
586 }
587 
588 void
589 lwkt_process_ipiq_frame(struct intrframe *frame)
590 {
591     globaldata_t gd = mycpu;
592     globaldata_t sgd;
593     lwkt_ipiq_t ip;
594     cpumask_t mask;
595     int n;
596 
597     ++gd->gd_processing_ipiq;
598 again:
599     mask = gd->gd_ipimask;
600     cpu_ccfence();
601     while (CPUMASK_TESTNZERO(mask)) {
602 	n = BSFCPUMASK(mask);
603 	if (n != gd->gd_cpuid) {
604 	    sgd = globaldata_find(n);
605 	    ip = sgd->gd_ipiq;
606 	    if (ip != NULL) {
607 		ip += gd->gd_cpuid;
608 		while (lwkt_process_ipiq_core(sgd, ip, frame, 0))
609 		    ;
610 		ATOMIC_CPUMASK_NANDBIT(gd->gd_ipimask, n);
611 		if (ip->ip_rindex != ip->ip_windex)
612 			ATOMIC_CPUMASK_ORBIT(gd->gd_ipimask, n);
613 	    }
614 	}
615 	CPUMASK_NANDBIT(mask, n);
616     }
617     if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
618 	if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame, 0)) {
619 	    if (gd->gd_curthread->td_cscount == 0)
620 		goto again;
621 	    /* need_ipiq(); do not reflag */
622 	}
623     }
624     --gd->gd_processing_ipiq;
625 }
626 
627 /*
628  * Only process incoming IPIQs from draining senders and only process them
629  * to the point where the draining sender is able to continue.  This is
630  * necessary to avoid deadlocking the IPI subsystem because we are acting on
631  * incoming messages and the callback may queue additional messages.
632  *
633  * We only want to have to act on senders that are blocked to limit the
634  * number of additional messages sent.  At the same time, recipients are
635  * trying to drain our own queue.  Theoretically this create a pipeline that
636  * cannot deadlock.
637  */
638 static void
639 lwkt_process_ipiq_nested(void)
640 {
641     globaldata_t gd = mycpu;
642     globaldata_t sgd;
643     lwkt_ipiq_t ip;
644     cpumask_t mask;
645     int n;
646     int limit;
647 
648     ++gd->gd_processing_ipiq;
649 again:
650     mask = gd->gd_ipimask;
651     cpu_ccfence();
652     while (CPUMASK_TESTNZERO(mask)) {
653 	n = BSFCPUMASK(mask);
654 	if (n != gd->gd_cpuid) {
655 	    sgd = globaldata_find(n);
656 	    ip = sgd->gd_ipiq;
657 
658 	    /*
659 	     * NOTE: We do not mess with the cpumask at all, instead we allow
660 	     *	     the top-level ipiq processor deal with it.
661 	     */
662 	    if (ip != NULL) {
663 		ip += gd->gd_cpuid;
664 		if ((limit = ip->ip_drain) != 0) {
665 		    lwkt_process_ipiq_core(sgd, ip, NULL, limit);
666 		    /* no gd_ipimask when doing limited processing */
667 		}
668 	    }
669 	}
670 	CPUMASK_NANDBIT(mask, n);
671     }
672 
673     /*
674      * Process pending cpusyncs.  If the current thread has a cpusync
675      * active cpusync we only run the list once and do not re-flag
676      * as the thread itself is processing its interlock.
677      */
678     if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL, 0)) {
679 	if (gd->gd_curthread->td_cscount == 0)
680 	    goto again;
681 	/* need_ipiq(); do not reflag */
682     }
683     --gd->gd_processing_ipiq;
684 }
685 
686 /*
687  * Process incoming IPI requests until only <limit> are left (0 to exhaust
688  * all incoming IPI requests).
689  */
690 static int
691 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
692 		       struct intrframe *frame, int limit)
693 {
694     globaldata_t mygd = mycpu;
695     int ri;
696     int wi;
697     ipifunc3_t copy_func;
698     void *copy_arg1;
699     int copy_arg2;
700 
701     /*
702      * Clear the originating core from our ipimask, we will process all
703      * incoming messages.
704      *
705      * Obtain the current write index, which is modified by a remote cpu.
706      * Issue a load fence to prevent speculative reads of e.g. data written
707      * by the other cpu prior to it updating the index.
708      */
709     KKASSERT(curthread->td_critcount);
710     wi = ip->ip_windex;
711     cpu_lfence();
712     ++mygd->gd_intr_nesting_level;
713 
714     /*
715      * NOTE: xindex is only updated after we are sure the function has
716      *	     finished execution.  Beware lwkt_process_ipiq() reentrancy!
717      *	     The function may send an IPI which may block/drain.
718      *
719      * NOTE: Due to additional IPI operations that the callback function
720      *	     may make, it is possible for both rindex and windex to advance and
721      *	     thus for rindex to advance passed our cached windex.
722      *
723      * NOTE: A load fence is required to prevent speculative loads prior
724      *	     to the loading of ip_rindex.  Even though stores might be
725      *	     ordered, loads are probably not.  A memory fence is required
726      *	     to prevent reordering of the loads after the ip_rindex update.
727      *
728      * NOTE: Single pass only.  Returns non-zero if the queue is not empty
729      *	     on return.
730      */
731     while (wi - (ri = ip->ip_rindex) > limit) {
732 	ri &= MAXCPUFIFO_MASK;
733 	cpu_lfence();
734 	copy_func = ip->ip_info[ri].func;
735 	copy_arg1 = ip->ip_info[ri].arg1;
736 	copy_arg2 = ip->ip_info[ri].arg2;
737 	cpu_mfence();
738 	++ip->ip_rindex;
739 	KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) ==
740 		 ((ri + 1) & MAXCPUFIFO_MASK));
741 	logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
742 #ifdef INVARIANTS
743 	if (ipiq_debug && (ip->ip_rindex & 0xFFFFFF) == 0) {
744 		kprintf("cpu %d ipifunc %p %p %d (frame %p)\n",
745 			mycpu->gd_cpuid,
746 			copy_func, copy_arg1, copy_arg2,
747 #if defined(__x86_64__)
748 			(frame ? (void *)frame->if_rip : NULL));
749 #else
750 			NULL);
751 #endif
752 	}
753 #endif
754 	copy_func(copy_arg1, copy_arg2, frame);
755 	cpu_sfence();
756 	ip->ip_xindex = ip->ip_rindex;
757 
758 #ifdef PANIC_DEBUG
759 	/*
760 	 * Simulate panics during the processing of an IPI
761 	 */
762 	if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
763 		if (--panic_ipiq_count == 0) {
764 #ifdef DDB
765 			Debugger("PANIC_DEBUG");
766 #else
767 			panic("PANIC_DEBUG");
768 #endif
769 		}
770 	}
771 #endif
772     }
773     --mygd->gd_intr_nesting_level;
774 
775     /*
776      * Return non-zero if there is still more in the queue.  Don't worry
777      * about fencing, we will get another interrupt if necessary.
778      */
779     return (ip->ip_rindex != ip->ip_windex);
780 }
781 
782 static void
783 lwkt_sync_ipiq(void *arg)
784 {
785     volatile cpumask_t *cpumask = arg;
786 
787     ATOMIC_CPUMASK_NANDBIT(*cpumask, mycpu->gd_cpuid);
788     if (CPUMASK_TESTZERO(*cpumask))
789 	wakeup(cpumask);
790 }
791 
792 void
793 lwkt_synchronize_ipiqs(const char *wmesg)
794 {
795     volatile cpumask_t other_cpumask;
796 
797     other_cpumask = smp_active_mask;
798     CPUMASK_ANDMASK(other_cpumask, mycpu->gd_other_cpus);
799     lwkt_send_ipiq_mask(other_cpumask, lwkt_sync_ipiq,
800 			__DEVOLATILE(void *, &other_cpumask));
801 
802     while (CPUMASK_TESTNZERO(other_cpumask)) {
803 	tsleep_interlock(&other_cpumask, 0);
804 	if (CPUMASK_TESTNZERO(other_cpumask))
805 	    tsleep(&other_cpumask, PINTERLOCKED, wmesg, 0);
806     }
807 }
808 
809 /*
810  * CPU Synchronization Support
811  *
812  * lwkt_cpusync_interlock()	- Place specified cpus in a quiescent state.
813  *				  The current cpu is placed in a hard critical
814  *				  section.
815  *
816  * lwkt_cpusync_deinterlock()	- Execute cs_func on specified cpus, including
817  *				  current cpu if specified, then return.
818  */
819 void
820 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *arg)
821 {
822     struct lwkt_cpusync cs;
823 
824     lwkt_cpusync_init(&cs, mask, func, arg);
825     lwkt_cpusync_interlock(&cs);
826     lwkt_cpusync_deinterlock(&cs);
827 }
828 
829 
830 void
831 lwkt_cpusync_interlock(lwkt_cpusync_t cs)
832 {
833     globaldata_t gd = mycpu;
834     cpumask_t mask;
835 
836     /*
837      * mask acknowledge (cs_mack):  0->mask for stage 1
838      *
839      * mack does not include the current cpu.
840      */
841     mask = cs->cs_mask;
842     CPUMASK_ANDMASK(mask, gd->gd_other_cpus);
843     CPUMASK_ANDMASK(mask, smp_active_mask);
844     CPUMASK_ASSZERO(cs->cs_mack);
845 
846     crit_enter_id("cpusync");
847     if (CPUMASK_TESTNZERO(mask)) {
848 	DEBUG_PUSH_INFO("cpusync_interlock");
849 	++ipiq_stat(gd).ipiq_cscount;
850 	++gd->gd_curthread->td_cscount;
851 	lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote1, cs);
852 	logipiq2(sync_start, (long)CPUMASK_LOWMASK(mask));
853 	while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) {
854 	    lwkt_process_ipiq();
855 	    cpu_pause();
856 #ifdef _KERNEL_VIRTUAL
857 	    pthread_yield();
858 #endif
859 	}
860 	DEBUG_POP_INFO();
861     }
862 }
863 
864 /*
865  * Interlocked cpus have executed remote1 and are polling in remote2.
866  * To deinterlock we clear cs_mack and wait for the cpus to execute
867  * the func and set their bit in cs_mack again.
868  *
869  */
870 void
871 lwkt_cpusync_deinterlock(lwkt_cpusync_t cs)
872 {
873     globaldata_t gd = mycpu;
874     cpumask_t mask;
875 
876     /*
877      * mask acknowledge (cs_mack):  mack->0->mack for stage 2
878      *
879      * Clearing cpu bits for polling cpus in cs_mack will cause them to
880      * execute stage 2, which executes the cs_func(cs_data) and then sets
881      * their bit in cs_mack again.
882      *
883      * mack does not include the current cpu.
884      */
885     mask = cs->cs_mack;
886     cpu_ccfence();
887     CPUMASK_ASSZERO(cs->cs_mack);
888     cpu_ccfence();
889     if (cs->cs_func && CPUMASK_TESTBIT(cs->cs_mask, gd->gd_cpuid))
890 	    cs->cs_func(cs->cs_data);
891     if (CPUMASK_TESTNZERO(mask)) {
892 	DEBUG_PUSH_INFO("cpusync_deinterlock");
893 	while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) {
894 	    lwkt_process_ipiq();
895 	    cpu_pause();
896 #ifdef _KERNEL_VIRTUAL
897 	    pthread_yield();
898 #endif
899 	}
900 	DEBUG_POP_INFO();
901 	/*
902 	 * cpusyncq ipis may be left queued without the RQF flag set due to
903 	 * a non-zero td_cscount, so be sure to process any laggards after
904 	 * decrementing td_cscount.
905 	 */
906 	--gd->gd_curthread->td_cscount;
907 	lwkt_process_ipiq();
908 	logipiq2(sync_end, (long)CPUMASK_LOWMASK(mask));
909     }
910     crit_exit_id("cpusync");
911 }
912 
913 /*
914  * The quick version does not quiesce the target cpu(s) but instead executes
915  * the function on the target cpu(s) and waits for all to acknowledge.  This
916  * avoids spinning on the target cpus.
917  *
918  * This function is typically only used for kernel_pmap updates.  User pmaps
919  * have to be quiesced.
920  */
921 void
922 lwkt_cpusync_quick(lwkt_cpusync_t cs)
923 {
924     globaldata_t gd = mycpu;
925     cpumask_t mask;
926 
927     /*
928      * stage-2 cs_mack only.
929      */
930     mask = cs->cs_mask;
931     CPUMASK_ANDMASK(mask, gd->gd_other_cpus);
932     CPUMASK_ANDMASK(mask, smp_active_mask);
933     CPUMASK_ASSZERO(cs->cs_mack);
934 
935     crit_enter_id("cpusync");
936     if (CPUMASK_TESTNZERO(mask)) {
937 	DEBUG_PUSH_INFO("cpusync_interlock");
938 	++ipiq_stat(gd).ipiq_cscount;
939 	++gd->gd_curthread->td_cscount;
940 	lwkt_send_ipiq_mask(mask, (ipifunc1_t)lwkt_cpusync_remote2, cs);
941 	logipiq2(sync_quick, (long)CPUMASK_LOWMASK(mask));
942 	while (CPUMASK_CMPMASKNEQ(cs->cs_mack, mask)) {
943 	    lwkt_process_ipiq();
944 	    cpu_pause();
945 #ifdef _KERNEL_VIRTUAL
946 	    pthread_yield();
947 #endif
948 	}
949 
950 	/*
951 	 * cpusyncq ipis may be left queued without the RQF flag set due to
952 	 * a non-zero td_cscount, so be sure to process any laggards after
953 	 * decrementing td_cscount.
954 	 */
955 	DEBUG_POP_INFO();
956 	--gd->gd_curthread->td_cscount;
957 	lwkt_process_ipiq();
958     }
959     if (cs->cs_func && CPUMASK_TESTBIT(cs->cs_mask, gd->gd_cpuid))
960 	    cs->cs_func(cs->cs_data);
961     crit_exit_id("cpusync");
962 }
963 
964 /*
965  * helper IPI remote messaging function.
966  *
967  * Called on remote cpu when a new cpu synchronization request has been
968  * sent to us.  Execute the run function and adjust cs_count, then requeue
969  * the request so we spin on it.
970  */
971 static void
972 lwkt_cpusync_remote1(lwkt_cpusync_t cs)
973 {
974     globaldata_t gd = mycpu;
975 
976     ATOMIC_CPUMASK_ORBIT(cs->cs_mack, gd->gd_cpuid);
977     lwkt_cpusync_remote2(cs);
978 }
979 
980 /*
981  * helper IPI remote messaging function.
982  *
983  * Poll for the originator telling us to finish.  If it hasn't, requeue
984  * our request so we spin on it.
985  */
986 static void
987 lwkt_cpusync_remote2(lwkt_cpusync_t cs)
988 {
989     globaldata_t gd = mycpu;
990 
991     if (CPUMASK_TESTMASK(cs->cs_mack, gd->gd_cpumask) == 0) {
992 	if (cs->cs_func)
993 		cs->cs_func(cs->cs_data);
994 	ATOMIC_CPUMASK_ORBIT(cs->cs_mack, gd->gd_cpuid);
995 	/* cs can be ripped out at this point */
996     } else {
997 	lwkt_ipiq_t ip;
998 	int wi;
999 
1000 	cpu_pause();
1001 #ifdef _KERNEL_VIRTUAL
1002 	pthread_yield();
1003 #endif
1004 	cpu_lfence();
1005 
1006 	/*
1007 	 * Requeue our IPI to avoid a deep stack recursion.  If no other
1008 	 * IPIs are pending we can just loop up, which should help VMs
1009 	 * better-detect spin loops.
1010 	 */
1011 	ip = &gd->gd_cpusyncq;
1012 
1013 	wi = ip->ip_windex & MAXCPUFIFO_MASK;
1014 	ip->ip_info[wi].func = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
1015 	ip->ip_info[wi].arg1 = cs;
1016 	ip->ip_info[wi].arg2 = 0;
1017 	cpu_sfence();
1018 	KKASSERT(ip->ip_windex - ip->ip_rindex < MAXCPUFIFO);
1019 	++ip->ip_windex;
1020 	if (ipiq_debug && (ip->ip_windex & 0xFFFFFF) == 0) {
1021 		kprintf("cpu %d cm=%016jx %016jx f=%p\n",
1022 			gd->gd_cpuid,
1023 			(intmax_t)CPUMASK_LOWMASK(cs->cs_mask),
1024 			(intmax_t)CPUMASK_LOWMASK(cs->cs_mack),
1025 			cs->cs_func);
1026 	}
1027     }
1028 }
1029 
1030 #define LWKT_IPIQ_NLATENCY	8
1031 #define LWKT_IPIQ_NLATENCY_MASK	(LWKT_IPIQ_NLATENCY - 1)
1032 
1033 struct lwkt_ipiq_latency_log {
1034 	int		idx;	/* unmasked index */
1035 	int		pad;
1036 	uint64_t	latency[LWKT_IPIQ_NLATENCY];
1037 };
1038 
1039 static struct lwkt_ipiq_latency_log	lwkt_ipiq_latency_logs[MAXCPU];
1040 static uint64_t save_tsc;
1041 
1042 /*
1043  * IPI callback (already in a critical section)
1044  */
1045 static void
1046 lwkt_ipiq_latency_testfunc(void *arg __unused)
1047 {
1048 	uint64_t delta_tsc;
1049 	struct globaldata *gd;
1050 	struct lwkt_ipiq_latency_log *lat;
1051 
1052 	/*
1053 	 * Get delta TSC (assume TSCs are synchronized) as quickly as
1054 	 * possible and then convert to nanoseconds.
1055 	 */
1056 	delta_tsc = rdtsc_ordered() - save_tsc;
1057 	delta_tsc = delta_tsc * 1000000000LU / tsc_frequency;
1058 
1059 	/*
1060 	 * Record in our save array.
1061 	 */
1062 	gd = mycpu;
1063 	lat = &lwkt_ipiq_latency_logs[gd->gd_cpuid];
1064 	lat->latency[lat->idx & LWKT_IPIQ_NLATENCY_MASK] = delta_tsc;
1065 	++lat->idx;
1066 }
1067 
1068 /*
1069  * Send IPI from cpu0 to other cpus
1070  *
1071  * NOTE: Machine must be idle for test to run dependably, and also probably
1072  *	 a good idea not to be running powerd.
1073  *
1074  * NOTE: Caller should use 'usched :1 <command>' to lock itself to cpu 0.
1075  *	 See 'ipitest' script in /usr/src/test/sysperf/ipitest
1076  */
1077 static int
1078 lwkt_ipiq_latency_test(SYSCTL_HANDLER_ARGS)
1079 {
1080 	struct globaldata *gd;
1081 	int cpu = 0, orig_cpu, error;
1082 
1083 	error = sysctl_handle_int(oidp, &cpu, arg2, req);
1084 	if (error || req->newptr == NULL)
1085 		return error;
1086 
1087 	if (cpu == 0)
1088 		return 0;
1089 	else if (cpu >= ncpus || cpu < 0)
1090 		return EINVAL;
1091 
1092 	orig_cpu = mycpuid;
1093 	lwkt_migratecpu(0);
1094 
1095 	gd = globaldata_find(cpu);
1096 
1097 	save_tsc = rdtsc_ordered();
1098 	lwkt_send_ipiq(gd, lwkt_ipiq_latency_testfunc, NULL);
1099 
1100 	lwkt_migratecpu(orig_cpu);
1101 	return 0;
1102 }
1103 
1104 SYSCTL_NODE(_debug, OID_AUTO, ipiq, CTLFLAG_RW, 0, "");
1105 SYSCTL_PROC(_debug_ipiq, OID_AUTO, latency_test, CTLTYPE_INT | CTLFLAG_RW,
1106     NULL, 0, lwkt_ipiq_latency_test, "I",
1107     "ipi latency test, arg: remote cpuid");
1108 
1109 static int
1110 lwkt_ipiq_latency(SYSCTL_HANDLER_ARGS)
1111 {
1112 	struct lwkt_ipiq_latency_log *latency = arg1;
1113 	uint64_t lat[LWKT_IPIQ_NLATENCY];
1114 	int i;
1115 
1116 	for (i = 0; i < LWKT_IPIQ_NLATENCY; ++i)
1117 		lat[i] = latency->latency[i];
1118 
1119 	return sysctl_handle_opaque(oidp, lat, sizeof(lat), req);
1120 }
1121 
1122 static void
1123 lwkt_ipiq_latency_init(void *dummy __unused)
1124 {
1125 	int cpu;
1126 
1127 	for (cpu = 0; cpu < ncpus; ++cpu) {
1128 		char name[32];
1129 
1130 		ksnprintf(name, sizeof(name), "latency%d", cpu);
1131 		SYSCTL_ADD_PROC(NULL, SYSCTL_STATIC_CHILDREN(_debug_ipiq),
1132 		    OID_AUTO, name, CTLTYPE_OPAQUE | CTLFLAG_RD,
1133 		    &lwkt_ipiq_latency_logs[cpu], 0, lwkt_ipiq_latency,
1134 		    "LU", "7 latest ipi latency measurement results");
1135 	}
1136 }
1137 SYSINIT(lwkt_ipiq_latency, SI_SUB_CONFIGURE, SI_ORDER_ANY,
1138     lwkt_ipiq_latency_init, NULL);
1139