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 1.17 2005/10/26 10:46:45 sephe 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/ipl.h>
72 #include <machine/smp.h>
73 #include <machine/atomic.h>
75 #define THREAD_STACK (UPAGES * PAGE_SIZE)
79 #include <sys/stdint.h>
80 #include <libcaps/thread.h>
81 #include <sys/thread.h>
82 #include <sys/msgport.h>
83 #include <sys/errno.h>
84 #include <libcaps/globaldata.h>
85 #include <machine/cpufunc.h>
86 #include <sys/thread2.h>
87 #include <sys/msgport2.h>
91 #include <machine/lock.h>
92 #include <machine/cpu.h>
93 #include <machine/atomic.h>
98 static __int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */
99 static __int64_t ipiq_fifofull; /* number of fifo full conditions detected */
100 static __int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */
101 static __int64_t ipiq_passive; /* passive IPI messages */
102 static __int64_t ipiq_cscount; /* number of cpu synchronizations */
103 static int ipiq_optimized = 1; /* XXX temporary sysctl */
105 static int panic_ipiq_cpu = -1;
106 static int panic_ipiq_count = 100;
113 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0, "");
116 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0, "");
117 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0, "");
118 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_optimized, CTLFLAG_RW, &ipiq_optimized, 0, "");
120 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, "");
121 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, "");
124 #define IPIQ_STRING "func=%p arg1=%p arg2=%d scpu=%d dcpu=%d"
125 #define IPIQ_ARG_SIZE (sizeof(void *) * 2 + sizeof(int) * 2)
127 #if !defined(KTR_IPIQ)
128 #define KTR_IPIQ KTR_ALL
130 KTR_INFO_MASTER(ipiq);
131 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARG_SIZE);
132 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARG_SIZE);
133 KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARG_SIZE);
134 KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARG_SIZE);
135 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARG_SIZE);
137 #define logipiq(name, func, arg1, arg2, sgd, dgd) \
138 KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid)
145 static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
146 struct intrframe *frame);
147 static void lwkt_cpusync_remote1(lwkt_cpusync_t poll);
148 static void lwkt_cpusync_remote2(lwkt_cpusync_t poll);
151 * Send a function execution request to another cpu. The request is queued
152 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
153 * possible target cpu. The FIFO can be written.
155 * If the FIFO fills up we have to enable interrupts to avoid an APIC
156 * deadlock and process pending IPIQs while waiting for it to empty.
157 * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
159 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
160 * end will take care of any pending interrupts.
162 * The actual hardware IPI is avoided if the target cpu is already processing
163 * the queue from a prior IPI. It is possible to pipeline IPI messages
164 * very quickly between cpus due to the FIFO hysteresis.
166 * Need not be called from a critical section.
169 lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2)
173 struct globaldata *gd = mycpu;
175 logipiq(send_norm, func, arg1, arg2, gd, target);
178 func(arg1, arg2, NULL);
182 ++gd->gd_intr_nesting_level;
184 if (gd->gd_intr_nesting_level > 20)
185 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
187 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
189 ip = &gd->gd_ipiq[target->gd_cpuid];
192 * Do not allow the FIFO to become full. Interrupts must be physically
193 * enabled while we liveloop to avoid deadlocking the APIC.
195 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
196 unsigned int eflags = read_eflags();
198 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
199 cpu_send_ipiq(target->gd_cpuid);
202 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
203 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
206 write_eflags(eflags);
210 * Queue the new message
212 windex = ip->ip_windex & MAXCPUFIFO_MASK;
213 ip->ip_func[windex] = func;
214 ip->ip_arg1[windex] = arg1;
215 ip->ip_arg2[windex] = arg2;
218 --gd->gd_intr_nesting_level;
221 * signal the target cpu that there is work pending.
223 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
224 cpu_send_ipiq(target->gd_cpuid);
226 if (ipiq_optimized == 0)
227 cpu_send_ipiq(target->gd_cpuid);
231 return(ip->ip_windex);
235 * Similar to lwkt_send_ipiq() but this function does not actually initiate
236 * the IPI to the target cpu unless the FIFO has become too full, so it is
239 * This function is used for non-critical IPI messages, such as memory
240 * deallocations. The queue will typically be flushed by the target cpu at
241 * the next clock interrupt.
243 * Need not be called from a critical section.
246 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
247 void *arg1, int arg2)
251 struct globaldata *gd = mycpu;
253 KKASSERT(target != gd);
255 logipiq(send_pasv, func, arg1, arg2, gd, target);
256 ++gd->gd_intr_nesting_level;
258 if (gd->gd_intr_nesting_level > 20)
259 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
261 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
264 ip = &gd->gd_ipiq[target->gd_cpuid];
267 * Do not allow the FIFO to become full. Interrupts must be physically
268 * enabled while we liveloop to avoid deadlocking the APIC.
270 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
271 unsigned int eflags = read_eflags();
273 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
274 cpu_send_ipiq(target->gd_cpuid);
277 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
278 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
281 write_eflags(eflags);
285 * Queue the new message
287 windex = ip->ip_windex & MAXCPUFIFO_MASK;
288 ip->ip_func[windex] = func;
289 ip->ip_arg1[windex] = arg1;
290 ip->ip_arg2[windex] = arg2;
293 --gd->gd_intr_nesting_level;
296 * Do not signal the target cpu, it will pick up the IPI when it next
297 * polls (typically on the next tick).
300 return(ip->ip_windex);
304 * Send an IPI request without blocking, return 0 on success, ENOENT on
305 * failure. The actual queueing of the hardware IPI may still force us
306 * to spin and process incoming IPIs but that will eventually go away
307 * when we've gotten rid of the other general IPIs.
310 lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func,
311 void *arg1, int arg2)
315 struct globaldata *gd = mycpu;
317 logipiq(send_nbio, func, arg1, arg2, gd, target);
318 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
320 func(arg1, arg2, NULL);
324 ip = &gd->gd_ipiq[target->gd_cpuid];
326 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
327 logipiq(send_fail, func, arg1, arg2, gd, target);
330 windex = ip->ip_windex & MAXCPUFIFO_MASK;
331 ip->ip_func[windex] = func;
332 ip->ip_arg1[windex] = arg1;
333 ip->ip_arg2[windex] = arg2;
338 * This isn't a passive IPI, we still have to signal the target cpu.
340 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
341 cpu_send_ipiq(target->gd_cpuid);
343 if (ipiq_optimized == 0)
344 cpu_send_ipiq(target->gd_cpuid);
352 * deprecated, used only by fast int forwarding.
355 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
357 return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
361 * Send a message to several target cpus. Typically used for scheduling.
362 * The message will not be sent to stopped cpus.
365 lwkt_send_ipiq3_mask(u_int32_t mask, ipifunc3_t func, void *arg1, int arg2)
370 mask &= ~stopped_cpus;
373 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
374 mask &= ~(1 << cpuid);
381 * Wait for the remote cpu to finish processing a function.
383 * YYY we have to enable interrupts and process the IPIQ while waiting
384 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
385 * function to do this! YYY we really should 'block' here.
387 * MUST be called from a critical section. This routine may be called
388 * from an interrupt (for example, if an interrupt wakes a foreign thread
392 lwkt_wait_ipiq(globaldata_t target, int seq)
395 int maxc = 100000000;
397 if (target != mycpu) {
398 ip = &mycpu->gd_ipiq[target->gd_cpuid];
399 if ((int)(ip->ip_xindex - seq) < 0) {
400 unsigned int eflags = read_eflags();
402 while ((int)(ip->ip_xindex - seq) < 0) {
407 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
409 panic("LWKT_WAIT_IPIQ");
411 * xindex may be modified by another cpu, use a load fence
412 * to ensure that the loop does not use a speculative value
413 * (which may improve performance).
417 write_eflags(eflags);
423 lwkt_seq_ipiq(globaldata_t target)
427 ip = &mycpu->gd_ipiq[target->gd_cpuid];
428 return(ip->ip_windex);
432 * Called from IPI interrupt (like a fast interrupt), which has placed
433 * us in a critical section. The MP lock may or may not be held.
434 * May also be called from doreti or splz, or be reentrantly called
435 * indirectly through the ip_func[] we run.
437 * There are two versions, one where no interrupt frame is available (when
438 * called from the send code and from splz, and one where an interrupt
439 * frame is available.
442 lwkt_process_ipiq(void)
444 globaldata_t gd = mycpu;
450 for (n = 0; n < ncpus; ++n) {
451 if (n != gd->gd_cpuid) {
452 sgd = globaldata_find(n);
455 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL))
460 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
461 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) {
462 if (gd->gd_curthread->td_cscount == 0)
471 lwkt_process_ipiq_frame(struct intrframe frame)
473 globaldata_t gd = mycpu;
479 for (n = 0; n < ncpus; ++n) {
480 if (n != gd->gd_cpuid) {
481 sgd = globaldata_find(n);
484 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], &frame))
489 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
490 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, &frame)) {
491 if (gd->gd_curthread->td_cscount == 0)
500 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
501 struct intrframe *frame)
505 ipifunc3_t copy_func;
510 * Obtain the current write index, which is modified by a remote cpu.
511 * Issue a load fence to prevent speculative reads of e.g. data written
512 * by the other cpu prior to it updating the index.
514 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
519 * Note: xindex is only updated after we are sure the function has
520 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
521 * function may send an IPI which may block/drain.
523 * Note: due to additional IPI operations that the callback function
524 * may make, it is possible for both rindex and windex to advance and
525 * thus for rindex to advance passed our cached windex.
527 while (wi - (ri = ip->ip_rindex) > 0) {
528 ri &= MAXCPUFIFO_MASK;
529 copy_func = ip->ip_func[ri];
530 copy_arg1 = ip->ip_arg1[ri];
531 copy_arg2 = ip->ip_arg2[ri];
534 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) == ((ri + 1) & MAXCPUFIFO_MASK));
535 logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
536 copy_func(copy_arg1, copy_arg2, frame);
538 ip->ip_xindex = ip->ip_rindex;
542 * Simulate panics during the processing of an IPI
544 if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
545 if (--panic_ipiq_count == 0) {
547 Debugger("PANIC_DEBUG");
549 panic("PANIC_DEBUG");
557 * Return non-zero if there are more IPI messages pending on this
558 * ipiq. ip_npoll is left set as long as possible to reduce the
559 * number of IPIs queued by the originating cpu, but must be cleared
560 * *BEFORE* checking windex.
562 atomic_poll_release_int(&ip->ip_npoll);
563 return(wi != ip->ip_windex);
569 * CPU Synchronization Support
571 * lwkt_cpusync_simple()
573 * The function is executed synchronously before return on remote cpus.
574 * A lwkt_cpusync_t pointer is passed as an argument. The data can
575 * be accessed via arg->cs_data.
577 * XXX should I just pass the data as an argument to be consistent?
581 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data)
583 struct lwkt_cpusync cmd;
585 cmd.cs_run_func = NULL;
586 cmd.cs_fin1_func = func;
587 cmd.cs_fin2_func = NULL;
589 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
590 if (mask & (1 << mycpu->gd_cpuid))
592 lwkt_cpusync_finish(&cmd);
596 * lwkt_cpusync_fastdata()
598 * The function is executed in tandem with return on remote cpus.
599 * The data is directly passed as an argument. Do not pass pointers to
600 * temporary storage as the storage might have
601 * gone poof by the time the target cpu executes
604 * At the moment lwkt_cpusync is declared on the stack and we must wait
605 * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future
606 * optimization we should be able to put a counter in the globaldata
607 * structure (if it is not otherwise being used) and just poke it and
608 * return without waiting. XXX
611 lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data)
613 struct lwkt_cpusync cmd;
615 cmd.cs_run_func = NULL;
616 cmd.cs_fin1_func = NULL;
617 cmd.cs_fin2_func = func;
619 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
620 if (mask & (1 << mycpu->gd_cpuid))
622 lwkt_cpusync_finish(&cmd);
626 * lwkt_cpusync_start()
628 * Start synchronization with a set of target cpus, return once they are
629 * known to be in a synchronization loop. The target cpus will execute
630 * poll->cs_run_func() IN TANDEM WITH THE RETURN.
632 * XXX future: add lwkt_cpusync_start_quick() and require a call to
633 * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to
634 * potentially absorb the IPI latency doing something useful.
637 lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll)
639 globaldata_t gd = mycpu;
642 poll->cs_mask = mask;
644 poll->cs_maxcount = lwkt_send_ipiq_mask(
645 mask & gd->gd_other_cpus & smp_active_mask,
646 (ipifunc1_t)lwkt_cpusync_remote1, poll);
648 if (mask & gd->gd_cpumask) {
649 if (poll->cs_run_func)
650 poll->cs_run_func(poll);
653 if (poll->cs_maxcount) {
655 ++gd->gd_curthread->td_cscount;
656 while (poll->cs_count != poll->cs_maxcount) {
666 lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll)
668 globaldata_t gd = mycpu;
673 mask &= ~poll->cs_mask;
674 poll->cs_mask |= mask;
676 count = lwkt_send_ipiq_mask(
677 mask & gd->gd_other_cpus & smp_active_mask,
678 (ipifunc1_t)lwkt_cpusync_remote1, poll);
680 if (mask & gd->gd_cpumask) {
681 if (poll->cs_run_func)
682 poll->cs_run_func(poll);
685 poll->cs_maxcount += count;
686 if (poll->cs_maxcount) {
687 if (poll->cs_maxcount == count)
688 ++gd->gd_curthread->td_cscount;
689 while (poll->cs_count != poll->cs_maxcount) {
699 * Finish synchronization with a set of target cpus. The target cpus will
700 * execute cs_fin1_func(poll) prior to this function returning, and will
701 * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN.
703 * If cs_maxcount is non-zero then we are mastering a cpusync with one or
704 * more remote cpus and must account for it in our thread structure.
707 lwkt_cpusync_finish(lwkt_cpusync_t poll)
709 globaldata_t gd = mycpu;
712 if (poll->cs_mask & gd->gd_cpumask) {
713 if (poll->cs_fin1_func)
714 poll->cs_fin1_func(poll);
715 if (poll->cs_fin2_func)
716 poll->cs_fin2_func(poll->cs_data);
719 if (poll->cs_maxcount) {
720 while (poll->cs_count != -(poll->cs_maxcount + 1)) {
725 --gd->gd_curthread->td_cscount;
733 * helper IPI remote messaging function.
735 * Called on remote cpu when a new cpu synchronization request has been
736 * sent to us. Execute the run function and adjust cs_count, then requeue
737 * the request so we spin on it.
740 lwkt_cpusync_remote1(lwkt_cpusync_t poll)
742 atomic_add_int(&poll->cs_count, 1);
743 if (poll->cs_run_func)
744 poll->cs_run_func(poll);
745 lwkt_cpusync_remote2(poll);
749 * helper IPI remote messaging function.
751 * Poll for the originator telling us to finish. If it hasn't, requeue
752 * our request so we spin on it. When the originator requests that we
753 * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data)
754 * in tandem with the release.
757 lwkt_cpusync_remote2(lwkt_cpusync_t poll)
759 if (poll->cs_count < 0) {
760 cpusync_func2_t savef;
763 if (poll->cs_fin1_func)
764 poll->cs_fin1_func(poll);
765 if (poll->cs_fin2_func) {
766 savef = poll->cs_fin2_func;
767 saved = poll->cs_data;
768 atomic_add_int(&poll->cs_count, -1);
771 atomic_add_int(&poll->cs_count, -1);
774 globaldata_t gd = mycpu;
778 ip = &gd->gd_cpusyncq;
779 wi = ip->ip_windex & MAXCPUFIFO_MASK;
780 ip->ip_func[wi] = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
781 ip->ip_arg1[wi] = poll;