2 * Copyright (c) 2003,2004 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * $DragonFly: src/sys/kern/lwkt_ipiq.c,v 1.21 2007/01/22 19:37:04 corecode Exp $
38 * This module implements IPI message queueing and the MI portion of IPI
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
50 #include <sys/rtprio.h>
51 #include <sys/queue.h>
52 #include <sys/thread2.h>
53 #include <sys/sysctl.h>
55 #include <sys/kthread.h>
56 #include <machine/cpu.h>
61 #include <vm/vm_param.h>
62 #include <vm/vm_kern.h>
63 #include <vm/vm_object.h>
64 #include <vm/vm_page.h>
65 #include <vm/vm_map.h>
66 #include <vm/vm_pager.h>
67 #include <vm/vm_extern.h>
68 #include <vm/vm_zone.h>
70 #include <machine/stdarg.h>
71 #include <machine/smp.h>
72 #include <machine/atomic.h>
76 #include <sys/stdint.h>
77 #include <libcaps/thread.h>
78 #include <sys/thread.h>
79 #include <sys/msgport.h>
80 #include <sys/errno.h>
81 #include <libcaps/globaldata.h>
82 #include <machine/cpufunc.h>
83 #include <sys/thread2.h>
84 #include <sys/msgport2.h>
88 #include <machine/lock.h>
89 #include <machine/cpu.h>
90 #include <machine/atomic.h>
95 static __int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */
96 static __int64_t ipiq_fifofull; /* number of fifo full conditions detected */
97 static __int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */
98 static __int64_t ipiq_passive; /* passive IPI messages */
99 static __int64_t ipiq_cscount; /* number of cpu synchronizations */
100 static int ipiq_optimized = 1; /* XXX temporary sysctl */
102 static int panic_ipiq_cpu = -1;
103 static int panic_ipiq_count = 100;
110 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
112 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0, "");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0, "");
115 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_optimized, CTLFLAG_RW, &ipiq_optimized, 0, "");
117 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, "");
118 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, "");
121 #define IPIQ_STRING "func=%p arg1=%p arg2=%d scpu=%d dcpu=%d"
122 #define IPIQ_ARG_SIZE (sizeof(void *) * 2 + sizeof(int) * 2)
124 #if !defined(KTR_IPIQ)
125 #define KTR_IPIQ KTR_ALL
127 KTR_INFO_MASTER(ipiq);
128 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARG_SIZE);
129 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARG_SIZE);
130 KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARG_SIZE);
131 KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARG_SIZE);
132 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARG_SIZE);
134 #define logipiq(name, func, arg1, arg2, sgd, dgd) \
135 KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid)
142 static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
143 struct intrframe *frame);
144 static void lwkt_cpusync_remote1(lwkt_cpusync_t poll);
145 static void lwkt_cpusync_remote2(lwkt_cpusync_t poll);
148 * Send a function execution request to another cpu. The request is queued
149 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
150 * possible target cpu. The FIFO can be written.
152 * If the FIFO fills up we have to enable interrupts to avoid an APIC
153 * deadlock and process pending IPIQs while waiting for it to empty.
154 * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
156 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
157 * end will take care of any pending interrupts.
159 * The actual hardware IPI is avoided if the target cpu is already processing
160 * the queue from a prior IPI. It is possible to pipeline IPI messages
161 * very quickly between cpus due to the FIFO hysteresis.
163 * Need not be called from a critical section.
166 lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2)
170 struct globaldata *gd = mycpu;
172 logipiq(send_norm, func, arg1, arg2, gd, target);
175 func(arg1, arg2, NULL);
179 ++gd->gd_intr_nesting_level;
181 if (gd->gd_intr_nesting_level > 20)
182 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
184 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
186 ip = &gd->gd_ipiq[target->gd_cpuid];
189 * Do not allow the FIFO to become full. Interrupts must be physically
190 * enabled while we liveloop to avoid deadlocking the APIC.
192 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
193 unsigned int eflags = read_eflags();
195 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
196 cpu_send_ipiq(target->gd_cpuid);
199 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
200 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
203 write_eflags(eflags);
207 * Queue the new message
209 windex = ip->ip_windex & MAXCPUFIFO_MASK;
210 ip->ip_func[windex] = func;
211 ip->ip_arg1[windex] = arg1;
212 ip->ip_arg2[windex] = arg2;
215 --gd->gd_intr_nesting_level;
218 * signal the target cpu that there is work pending.
220 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
221 cpu_send_ipiq(target->gd_cpuid);
223 if (ipiq_optimized == 0)
224 cpu_send_ipiq(target->gd_cpuid);
228 return(ip->ip_windex);
232 * Similar to lwkt_send_ipiq() but this function does not actually initiate
233 * the IPI to the target cpu unless the FIFO has become too full, so it is
236 * This function is used for non-critical IPI messages, such as memory
237 * deallocations. The queue will typically be flushed by the target cpu at
238 * the next clock interrupt.
240 * Need not be called from a critical section.
243 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
244 void *arg1, int arg2)
248 struct globaldata *gd = mycpu;
250 KKASSERT(target != gd);
252 logipiq(send_pasv, func, arg1, arg2, gd, target);
253 ++gd->gd_intr_nesting_level;
255 if (gd->gd_intr_nesting_level > 20)
256 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
258 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
261 ip = &gd->gd_ipiq[target->gd_cpuid];
264 * Do not allow the FIFO to become full. Interrupts must be physically
265 * enabled while we liveloop to avoid deadlocking the APIC.
267 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
268 unsigned int eflags = read_eflags();
270 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
271 cpu_send_ipiq(target->gd_cpuid);
274 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
275 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
278 write_eflags(eflags);
282 * Queue the new message
284 windex = ip->ip_windex & MAXCPUFIFO_MASK;
285 ip->ip_func[windex] = func;
286 ip->ip_arg1[windex] = arg1;
287 ip->ip_arg2[windex] = arg2;
290 --gd->gd_intr_nesting_level;
293 * Do not signal the target cpu, it will pick up the IPI when it next
294 * polls (typically on the next tick).
297 return(ip->ip_windex);
301 * Send an IPI request without blocking, return 0 on success, ENOENT on
302 * failure. The actual queueing of the hardware IPI may still force us
303 * to spin and process incoming IPIs but that will eventually go away
304 * when we've gotten rid of the other general IPIs.
307 lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func,
308 void *arg1, int arg2)
312 struct globaldata *gd = mycpu;
314 logipiq(send_nbio, func, arg1, arg2, gd, target);
315 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
317 func(arg1, arg2, NULL);
321 ip = &gd->gd_ipiq[target->gd_cpuid];
323 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
324 logipiq(send_fail, func, arg1, arg2, gd, target);
327 windex = ip->ip_windex & MAXCPUFIFO_MASK;
328 ip->ip_func[windex] = func;
329 ip->ip_arg1[windex] = arg1;
330 ip->ip_arg2[windex] = arg2;
335 * This isn't a passive IPI, we still have to signal the target cpu.
337 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
338 cpu_send_ipiq(target->gd_cpuid);
340 if (ipiq_optimized == 0)
341 cpu_send_ipiq(target->gd_cpuid);
349 * deprecated, used only by fast int forwarding.
352 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
354 return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
358 * Send a message to several target cpus. Typically used for scheduling.
359 * The message will not be sent to stopped cpus.
362 lwkt_send_ipiq3_mask(u_int32_t mask, ipifunc3_t func, void *arg1, int arg2)
367 mask &= ~stopped_cpus;
370 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
371 mask &= ~(1 << cpuid);
378 * Wait for the remote cpu to finish processing a function.
380 * YYY we have to enable interrupts and process the IPIQ while waiting
381 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
382 * function to do this! YYY we really should 'block' here.
384 * MUST be called from a critical section. This routine may be called
385 * from an interrupt (for example, if an interrupt wakes a foreign thread
389 lwkt_wait_ipiq(globaldata_t target, int seq)
392 int maxc = 100000000;
394 if (target != mycpu) {
395 ip = &mycpu->gd_ipiq[target->gd_cpuid];
396 if ((int)(ip->ip_xindex - seq) < 0) {
397 unsigned int eflags = read_eflags();
399 while ((int)(ip->ip_xindex - seq) < 0) {
404 kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
406 panic("LWKT_WAIT_IPIQ");
408 * xindex may be modified by another cpu, use a load fence
409 * to ensure that the loop does not use a speculative value
410 * (which may improve performance).
414 write_eflags(eflags);
420 lwkt_seq_ipiq(globaldata_t target)
424 ip = &mycpu->gd_ipiq[target->gd_cpuid];
425 return(ip->ip_windex);
429 * Called from IPI interrupt (like a fast interrupt), which has placed
430 * us in a critical section. The MP lock may or may not be held.
431 * May also be called from doreti or splz, or be reentrantly called
432 * indirectly through the ip_func[] we run.
434 * There are two versions, one where no interrupt frame is available (when
435 * called from the send code and from splz, and one where an interrupt
436 * frame is available.
439 lwkt_process_ipiq(void)
441 globaldata_t gd = mycpu;
447 for (n = 0; n < ncpus; ++n) {
448 if (n != gd->gd_cpuid) {
449 sgd = globaldata_find(n);
452 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL))
457 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
458 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) {
459 if (gd->gd_curthread->td_cscount == 0)
468 lwkt_process_ipiq_frame(struct intrframe *frame)
470 globaldata_t gd = mycpu;
476 for (n = 0; n < ncpus; ++n) {
477 if (n != gd->gd_cpuid) {
478 sgd = globaldata_find(n);
481 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], frame))
486 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
487 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, frame)) {
488 if (gd->gd_curthread->td_cscount == 0)
497 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
498 struct intrframe *frame)
502 ipifunc3_t copy_func;
507 * Obtain the current write index, which is modified by a remote cpu.
508 * Issue a load fence to prevent speculative reads of e.g. data written
509 * by the other cpu prior to it updating the index.
511 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
516 * Note: xindex is only updated after we are sure the function has
517 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
518 * function may send an IPI which may block/drain.
520 * Note: due to additional IPI operations that the callback function
521 * may make, it is possible for both rindex and windex to advance and
522 * thus for rindex to advance passed our cached windex.
524 while (wi - (ri = ip->ip_rindex) > 0) {
525 ri &= MAXCPUFIFO_MASK;
526 copy_func = ip->ip_func[ri];
527 copy_arg1 = ip->ip_arg1[ri];
528 copy_arg2 = ip->ip_arg2[ri];
531 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) == ((ri + 1) & MAXCPUFIFO_MASK));
532 logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
533 copy_func(copy_arg1, copy_arg2, frame);
535 ip->ip_xindex = ip->ip_rindex;
539 * Simulate panics during the processing of an IPI
541 if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
542 if (--panic_ipiq_count == 0) {
544 Debugger("PANIC_DEBUG");
546 panic("PANIC_DEBUG");
554 * Return non-zero if there are more IPI messages pending on this
555 * ipiq. ip_npoll is left set as long as possible to reduce the
556 * number of IPIs queued by the originating cpu, but must be cleared
557 * *BEFORE* checking windex.
559 atomic_poll_release_int(&ip->ip_npoll);
560 return(wi != ip->ip_windex);
566 * CPU Synchronization Support
568 * lwkt_cpusync_simple()
570 * The function is executed synchronously before return on remote cpus.
571 * A lwkt_cpusync_t pointer is passed as an argument. The data can
572 * be accessed via arg->cs_data.
574 * XXX should I just pass the data as an argument to be consistent?
578 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data)
580 struct lwkt_cpusync cmd;
582 cmd.cs_run_func = NULL;
583 cmd.cs_fin1_func = func;
584 cmd.cs_fin2_func = NULL;
586 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
587 if (mask & (1 << mycpu->gd_cpuid))
589 lwkt_cpusync_finish(&cmd);
593 * lwkt_cpusync_fastdata()
595 * The function is executed in tandem with return on remote cpus.
596 * The data is directly passed as an argument. Do not pass pointers to
597 * temporary storage as the storage might have
598 * gone poof by the time the target cpu executes
601 * At the moment lwkt_cpusync is declared on the stack and we must wait
602 * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future
603 * optimization we should be able to put a counter in the globaldata
604 * structure (if it is not otherwise being used) and just poke it and
605 * return without waiting. XXX
608 lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data)
610 struct lwkt_cpusync cmd;
612 cmd.cs_run_func = NULL;
613 cmd.cs_fin1_func = NULL;
614 cmd.cs_fin2_func = func;
616 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
617 if (mask & (1 << mycpu->gd_cpuid))
619 lwkt_cpusync_finish(&cmd);
623 * lwkt_cpusync_start()
625 * Start synchronization with a set of target cpus, return once they are
626 * known to be in a synchronization loop. The target cpus will execute
627 * poll->cs_run_func() IN TANDEM WITH THE RETURN.
629 * XXX future: add lwkt_cpusync_start_quick() and require a call to
630 * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to
631 * potentially absorb the IPI latency doing something useful.
634 lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll)
636 globaldata_t gd = mycpu;
639 poll->cs_mask = mask;
641 poll->cs_maxcount = lwkt_send_ipiq_mask(
642 mask & gd->gd_other_cpus & smp_active_mask,
643 (ipifunc1_t)lwkt_cpusync_remote1, poll);
645 if (mask & gd->gd_cpumask) {
646 if (poll->cs_run_func)
647 poll->cs_run_func(poll);
650 if (poll->cs_maxcount) {
652 ++gd->gd_curthread->td_cscount;
653 while (poll->cs_count != poll->cs_maxcount) {
663 lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll)
665 globaldata_t gd = mycpu;
670 mask &= ~poll->cs_mask;
671 poll->cs_mask |= mask;
673 count = lwkt_send_ipiq_mask(
674 mask & gd->gd_other_cpus & smp_active_mask,
675 (ipifunc1_t)lwkt_cpusync_remote1, poll);
677 if (mask & gd->gd_cpumask) {
678 if (poll->cs_run_func)
679 poll->cs_run_func(poll);
682 poll->cs_maxcount += count;
683 if (poll->cs_maxcount) {
684 if (poll->cs_maxcount == count)
685 ++gd->gd_curthread->td_cscount;
686 while (poll->cs_count != poll->cs_maxcount) {
696 * Finish synchronization with a set of target cpus. The target cpus will
697 * execute cs_fin1_func(poll) prior to this function returning, and will
698 * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN.
700 * If cs_maxcount is non-zero then we are mastering a cpusync with one or
701 * more remote cpus and must account for it in our thread structure.
704 lwkt_cpusync_finish(lwkt_cpusync_t poll)
706 globaldata_t gd = mycpu;
709 if (poll->cs_mask & gd->gd_cpumask) {
710 if (poll->cs_fin1_func)
711 poll->cs_fin1_func(poll);
712 if (poll->cs_fin2_func)
713 poll->cs_fin2_func(poll->cs_data);
716 if (poll->cs_maxcount) {
717 while (poll->cs_count != -(poll->cs_maxcount + 1)) {
722 --gd->gd_curthread->td_cscount;
730 * helper IPI remote messaging function.
732 * Called on remote cpu when a new cpu synchronization request has been
733 * sent to us. Execute the run function and adjust cs_count, then requeue
734 * the request so we spin on it.
737 lwkt_cpusync_remote1(lwkt_cpusync_t poll)
739 atomic_add_int(&poll->cs_count, 1);
740 if (poll->cs_run_func)
741 poll->cs_run_func(poll);
742 lwkt_cpusync_remote2(poll);
746 * helper IPI remote messaging function.
748 * Poll for the originator telling us to finish. If it hasn't, requeue
749 * our request so we spin on it. When the originator requests that we
750 * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data)
751 * in tandem with the release.
754 lwkt_cpusync_remote2(lwkt_cpusync_t poll)
756 if (poll->cs_count < 0) {
757 cpusync_func2_t savef;
760 if (poll->cs_fin1_func)
761 poll->cs_fin1_func(poll);
762 if (poll->cs_fin2_func) {
763 savef = poll->cs_fin2_func;
764 saved = poll->cs_data;
765 atomic_add_int(&poll->cs_count, -1);
768 atomic_add_int(&poll->cs_count, -1);
771 globaldata_t gd = mycpu;
775 ip = &gd->gd_cpusyncq;
776 wi = ip->ip_windex & MAXCPUFIFO_MASK;
777 ip->ip_func[wi] = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
778 ip->ip_arg1[wi] = poll;