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 1.13 2005/06/21 05:25:17 dillon Exp $
38 * This module implements IPI message queueing and the MI portion of IPI
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/kernel.h>
48 #include <sys/rtprio.h>
49 #include <sys/queue.h>
50 #include <sys/thread2.h>
51 #include <sys/sysctl.h>
53 #include <sys/kthread.h>
54 #include <machine/cpu.h>
59 #include <vm/vm_param.h>
60 #include <vm/vm_kern.h>
61 #include <vm/vm_object.h>
62 #include <vm/vm_page.h>
63 #include <vm/vm_map.h>
64 #include <vm/vm_pager.h>
65 #include <vm/vm_extern.h>
66 #include <vm/vm_zone.h>
68 #include <machine/stdarg.h>
69 #include <machine/ipl.h>
70 #include <machine/smp.h>
71 #include <machine/atomic.h>
73 #define THREAD_STACK (UPAGES * PAGE_SIZE)
77 #include <sys/stdint.h>
78 #include <libcaps/thread.h>
79 #include <sys/thread.h>
80 #include <sys/msgport.h>
81 #include <sys/errno.h>
82 #include <libcaps/globaldata.h>
83 #include <machine/cpufunc.h>
84 #include <sys/thread2.h>
85 #include <sys/msgport2.h>
89 #include <machine/lock.h>
90 #include <machine/cpu.h>
91 #include <machine/atomic.h>
96 static __int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */
97 static __int64_t ipiq_fifofull; /* number of fifo full conditions detected */
98 static __int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */
99 static __int64_t ipiq_passive; /* passive IPI messages */
100 static __int64_t ipiq_cscount; /* number of cpu synchronizations */
101 static int ipiq_optimized = 1; /* XXX temporary sysctl */
107 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
108 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
109 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0, "");
110 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0, "");
111 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0, "");
112 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_optimized, CTLFLAG_RW, &ipiq_optimized, 0, "");
114 #define IPIQ_STRING "func=%p arg=%p scpu=%d dcpu=%d"
115 #define IPIQ_ARG_SIZE (sizeof(void *) * 2 + sizeof(int) * 2)
117 #if !defined(KTR_IPIQ)
118 #define KTR_IPIQ KTR_ALL
120 KTR_INFO_MASTER(ipiq);
121 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARG_SIZE);
122 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARG_SIZE);
123 KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARG_SIZE);
124 KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARG_SIZE);
125 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARG_SIZE);
127 #define logipiq(name, func, arg, sgd, dgd) \
128 KTR_LOG(ipiq_ ## name, func, arg, sgd->gd_cpuid, dgd->gd_cpuid)
135 static int lwkt_process_ipiq1(globaldata_t sgd, lwkt_ipiq_t ip, struct intrframe *frame);
136 static void lwkt_cpusync_remote1(lwkt_cpusync_t poll);
137 static void lwkt_cpusync_remote2(lwkt_cpusync_t poll);
140 * Send a function execution request to another cpu. The request is queued
141 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
142 * possible target cpu. The FIFO can be written.
144 * If the FIFO fills up we have to enable interrupts to avoid an APIC
145 * deadlock and process pending IPIQs while waiting for it to empty.
146 * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
148 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
149 * end will take care of any pending interrupts.
151 * The actual hardware IPI is avoided if the target cpu is already processing
152 * the queue from a prior IPI. It is possible to pipeline IPI messages
153 * very quickly between cpus due to the FIFO hysteresis.
155 * Need not be called from a critical section.
158 lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg)
162 struct globaldata *gd = mycpu;
164 logipiq(send_norm, func, arg, gd, target);
171 ++gd->gd_intr_nesting_level;
173 if (gd->gd_intr_nesting_level > 20)
174 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
176 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
178 ip = &gd->gd_ipiq[target->gd_cpuid];
181 * Do not allow the FIFO to become full. Interrupts must be physically
182 * enabled while we liveloop to avoid deadlocking the APIC.
184 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
185 unsigned int eflags = read_eflags();
187 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
188 cpu_send_ipiq(target->gd_cpuid);
191 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
192 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
195 write_eflags(eflags);
199 * Queue the new message
201 windex = ip->ip_windex & MAXCPUFIFO_MASK;
202 ip->ip_func[windex] = (ipifunc2_t)func;
203 ip->ip_arg[windex] = arg;
206 --gd->gd_intr_nesting_level;
209 * signal the target cpu that there is work pending.
211 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
212 cpu_send_ipiq(target->gd_cpuid);
214 if (ipiq_optimized == 0)
215 cpu_send_ipiq(target->gd_cpuid);
219 return(ip->ip_windex);
223 * Similar to lwkt_send_ipiq() but this function does not actually initiate
224 * the IPI to the target cpu unless the FIFO has become too full, so it is
227 * This function is used for non-critical IPI messages, such as memory
228 * deallocations. The queue will typically be flushed by the target cpu at
229 * the next clock interrupt.
231 * Need not be called from a critical section.
234 lwkt_send_ipiq_passive(globaldata_t target, ipifunc_t func, void *arg)
238 struct globaldata *gd = mycpu;
240 KKASSERT(target != gd);
242 logipiq(send_pasv, func, arg, gd, target);
243 ++gd->gd_intr_nesting_level;
245 if (gd->gd_intr_nesting_level > 20)
246 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
248 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
251 ip = &gd->gd_ipiq[target->gd_cpuid];
254 * Do not allow the FIFO to become full. Interrupts must be physically
255 * enabled while we liveloop to avoid deadlocking the APIC.
257 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
258 unsigned int eflags = read_eflags();
260 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
261 cpu_send_ipiq(target->gd_cpuid);
264 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
265 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
268 write_eflags(eflags);
272 * Queue the new message
274 windex = ip->ip_windex & MAXCPUFIFO_MASK;
275 ip->ip_func[windex] = (ipifunc2_t)func;
276 ip->ip_arg[windex] = arg;
279 --gd->gd_intr_nesting_level;
282 * Do not signal the target cpu, it will pick up the IPI when it next
283 * polls (typically on the next tick).
286 return(ip->ip_windex);
290 * Send an IPI request without blocking, return 0 on success, ENOENT on
291 * failure. The actual queueing of the hardware IPI may still force us
292 * to spin and process incoming IPIs but that will eventually go away
293 * when we've gotten rid of the other general IPIs.
296 lwkt_send_ipiq_nowait(globaldata_t target, ipifunc_t func, void *arg)
300 struct globaldata *gd = mycpu;
302 logipiq(send_nbio, func, arg, gd, target);
303 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
309 ip = &gd->gd_ipiq[target->gd_cpuid];
311 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
312 logipiq(send_fail, func, arg, gd, target);
315 windex = ip->ip_windex & MAXCPUFIFO_MASK;
316 ip->ip_func[windex] = (ipifunc2_t)func;
317 ip->ip_arg[windex] = arg;
322 * This isn't a passive IPI, we still have to signal the target cpu.
324 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
325 cpu_send_ipiq(target->gd_cpuid);
327 if (ipiq_optimized == 0)
328 cpu_send_ipiq(target->gd_cpuid);
336 * deprecated, used only by fast int forwarding.
339 lwkt_send_ipiq_bycpu(int dcpu, ipifunc_t func, void *arg)
341 return(lwkt_send_ipiq(globaldata_find(dcpu), func, arg));
345 * Send a message to several target cpus. Typically used for scheduling.
346 * The message will not be sent to stopped cpus.
349 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
354 mask &= ~stopped_cpus;
357 lwkt_send_ipiq(globaldata_find(cpuid), func, arg);
358 mask &= ~(1 << cpuid);
365 * Wait for the remote cpu to finish processing a function.
367 * YYY we have to enable interrupts and process the IPIQ while waiting
368 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
369 * function to do this! YYY we really should 'block' here.
371 * MUST be called from a critical section. This routine may be called
372 * from an interrupt (for example, if an interrupt wakes a foreign thread
376 lwkt_wait_ipiq(globaldata_t target, int seq)
379 int maxc = 100000000;
381 if (target != mycpu) {
382 ip = &mycpu->gd_ipiq[target->gd_cpuid];
383 if ((int)(ip->ip_xindex - seq) < 0) {
384 unsigned int eflags = read_eflags();
386 while ((int)(ip->ip_xindex - seq) < 0) {
391 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
393 panic("LWKT_WAIT_IPIQ");
395 * xindex may be modified by another cpu, use a load fence
396 * to ensure that the loop does not use a speculative value
397 * (which may improve performance).
401 write_eflags(eflags);
407 lwkt_seq_ipiq(globaldata_t target)
411 ip = &mycpu->gd_ipiq[target->gd_cpuid];
412 return(ip->ip_windex);
416 * Called from IPI interrupt (like a fast interrupt), which has placed
417 * us in a critical section. The MP lock may or may not be held.
418 * May also be called from doreti or splz, or be reentrantly called
419 * indirectly through the ip_func[] we run.
421 * There are two versions, one where no interrupt frame is available (when
422 * called from the send code and from splz, and one where an interrupt
423 * frame is available.
426 lwkt_process_ipiq(void)
428 globaldata_t gd = mycpu;
434 for (n = 0; n < ncpus; ++n) {
435 if (n != gd->gd_cpuid) {
436 sgd = globaldata_find(n);
439 while (lwkt_process_ipiq1(sgd, &ip[gd->gd_cpuid], NULL))
444 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
445 if (lwkt_process_ipiq1(gd, &gd->gd_cpusyncq, NULL)) {
446 if (gd->gd_curthread->td_cscount == 0)
455 lwkt_process_ipiq_frame(struct intrframe frame)
457 globaldata_t gd = mycpu;
463 for (n = 0; n < ncpus; ++n) {
464 if (n != gd->gd_cpuid) {
465 sgd = globaldata_find(n);
468 while (lwkt_process_ipiq1(sgd, &ip[gd->gd_cpuid], &frame))
473 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
474 if (lwkt_process_ipiq1(gd, &gd->gd_cpusyncq, &frame)) {
475 if (gd->gd_curthread->td_cscount == 0)
484 lwkt_process_ipiq1(globaldata_t sgd, lwkt_ipiq_t ip, struct intrframe *frame)
488 void (*copy_func)(void *data, struct intrframe *frame);
492 * Obtain the current write index, which is modified by a remote cpu.
493 * Issue a load fence to prevent speculative reads of e.g. data written
494 * by the other cpu prior to it updating the index.
496 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
501 * Note: xindex is only updated after we are sure the function has
502 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
503 * function may send an IPI which may block/drain.
505 while ((ri = ip->ip_rindex) != wi) {
506 ri &= MAXCPUFIFO_MASK;
507 copy_func = ip->ip_func[ri];
508 copy_arg = ip->ip_arg[ri];
511 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) == ((ri + 1) & MAXCPUFIFO_MASK));
512 logipiq(receive, copy_func, copy_arg, sgd, mycpu);
513 copy_func(copy_arg, frame);
515 ip->ip_xindex = ip->ip_rindex;
519 * Return non-zero if there are more IPI messages pending on this
520 * ipiq. ip_npoll is left set as long as possible to reduce the
521 * number of IPIs queued by the originating cpu, but must be cleared
522 * *BEFORE* checking windex.
524 atomic_poll_release_int(&ip->ip_npoll);
525 return(wi != ip->ip_windex);
531 * !SMP dummy routines
535 lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg)
537 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", target->gd_cpuid, func, arg);
538 return(0); /* NOT REACHED */
542 lwkt_wait_ipiq(globaldata_t target, int seq)
544 panic("lwkt_wait_ipiq: UP box! (%d,%d)", target->gd_cpuid, seq);
550 * CPU Synchronization Support
552 * lwkt_cpusync_simple()
554 * The function is executed synchronously before return on remote cpus.
555 * A lwkt_cpusync_t pointer is passed as an argument. The data can
556 * be accessed via arg->cs_data.
558 * XXX should I just pass the data as an argument to be consistent?
562 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data)
564 struct lwkt_cpusync cmd;
566 cmd.cs_run_func = NULL;
567 cmd.cs_fin1_func = func;
568 cmd.cs_fin2_func = NULL;
570 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
571 if (mask & (1 << mycpu->gd_cpuid))
573 lwkt_cpusync_finish(&cmd);
577 * lwkt_cpusync_fastdata()
579 * The function is executed in tandem with return on remote cpus.
580 * The data is directly passed as an argument. Do not pass pointers to
581 * temporary storage as the storage might have
582 * gone poof by the time the target cpu executes
585 * At the moment lwkt_cpusync is declared on the stack and we must wait
586 * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future
587 * optimization we should be able to put a counter in the globaldata
588 * structure (if it is not otherwise being used) and just poke it and
589 * return without waiting. XXX
592 lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data)
594 struct lwkt_cpusync cmd;
596 cmd.cs_run_func = NULL;
597 cmd.cs_fin1_func = NULL;
598 cmd.cs_fin2_func = func;
600 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
601 if (mask & (1 << mycpu->gd_cpuid))
603 lwkt_cpusync_finish(&cmd);
607 * lwkt_cpusync_start()
609 * Start synchronization with a set of target cpus, return once they are
610 * known to be in a synchronization loop. The target cpus will execute
611 * poll->cs_run_func() IN TANDEM WITH THE RETURN.
613 * XXX future: add lwkt_cpusync_start_quick() and require a call to
614 * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to
615 * potentially absorb the IPI latency doing something useful.
618 lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll)
620 globaldata_t gd = mycpu;
623 poll->cs_mask = mask;
625 poll->cs_maxcount = lwkt_send_ipiq_mask(
626 mask & gd->gd_other_cpus & smp_active_mask,
627 (ipifunc_t)lwkt_cpusync_remote1, poll);
629 if (mask & gd->gd_cpumask) {
630 if (poll->cs_run_func)
631 poll->cs_run_func(poll);
634 if (poll->cs_maxcount) {
636 ++gd->gd_curthread->td_cscount;
637 while (poll->cs_count != poll->cs_maxcount) {
647 lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll)
649 globaldata_t gd = mycpu;
654 mask &= ~poll->cs_mask;
655 poll->cs_mask |= mask;
657 count = lwkt_send_ipiq_mask(
658 mask & gd->gd_other_cpus & smp_active_mask,
659 (ipifunc_t)lwkt_cpusync_remote1, poll);
661 if (mask & gd->gd_cpumask) {
662 if (poll->cs_run_func)
663 poll->cs_run_func(poll);
666 poll->cs_maxcount += count;
667 if (poll->cs_maxcount) {
668 if (poll->cs_maxcount == count)
669 ++gd->gd_curthread->td_cscount;
670 while (poll->cs_count != poll->cs_maxcount) {
680 * Finish synchronization with a set of target cpus. The target cpus will
681 * execute cs_fin1_func(poll) prior to this function returning, and will
682 * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN.
684 * If cs_maxcount is non-zero then we are mastering a cpusync with one or
685 * more remote cpus and must account for it in our thread structure.
688 lwkt_cpusync_finish(lwkt_cpusync_t poll)
690 globaldata_t gd = mycpu;
693 if (poll->cs_mask & gd->gd_cpumask) {
694 if (poll->cs_fin1_func)
695 poll->cs_fin1_func(poll);
696 if (poll->cs_fin2_func)
697 poll->cs_fin2_func(poll->cs_data);
700 if (poll->cs_maxcount) {
701 while (poll->cs_count != -(poll->cs_maxcount + 1)) {
706 --gd->gd_curthread->td_cscount;
714 * helper IPI remote messaging function.
716 * Called on remote cpu when a new cpu synchronization request has been
717 * sent to us. Execute the run function and adjust cs_count, then requeue
718 * the request so we spin on it.
721 lwkt_cpusync_remote1(lwkt_cpusync_t poll)
723 atomic_add_int(&poll->cs_count, 1);
724 if (poll->cs_run_func)
725 poll->cs_run_func(poll);
726 lwkt_cpusync_remote2(poll);
730 * helper IPI remote messaging function.
732 * Poll for the originator telling us to finish. If it hasn't, requeue
733 * our request so we spin on it. When the originator requests that we
734 * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data)
735 * in tandem with the release.
738 lwkt_cpusync_remote2(lwkt_cpusync_t poll)
740 if (poll->cs_count < 0) {
741 cpusync_func2_t savef;
744 if (poll->cs_fin1_func)
745 poll->cs_fin1_func(poll);
746 if (poll->cs_fin2_func) {
747 savef = poll->cs_fin2_func;
748 saved = poll->cs_data;
749 atomic_add_int(&poll->cs_count, -1);
752 atomic_add_int(&poll->cs_count, -1);
755 globaldata_t gd = mycpu;
759 ip = &gd->gd_cpusyncq;
760 wi = ip->ip_windex & MAXCPUFIFO_MASK;
761 ip->ip_func[wi] = (ipifunc2_t)lwkt_cpusync_remote2;
762 ip->ip_arg[wi] = poll;