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 1.19 2006/12/23 00:35:04 swildner 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>
74 #define THREAD_STACK (UPAGES * PAGE_SIZE)
78 #include <sys/stdint.h>
79 #include <libcaps/thread.h>
80 #include <sys/thread.h>
81 #include <sys/msgport.h>
82 #include <sys/errno.h>
83 #include <libcaps/globaldata.h>
84 #include <machine/cpufunc.h>
85 #include <sys/thread2.h>
86 #include <sys/msgport2.h>
90 #include <machine/lock.h>
91 #include <machine/cpu.h>
92 #include <machine/atomic.h>
97 static __int64_t ipiq_count; /* total calls to lwkt_send_ipiq*() */
98 static __int64_t ipiq_fifofull; /* number of fifo full conditions detected */
99 static __int64_t ipiq_avoided; /* interlock with target avoids cpu ipi */
100 static __int64_t ipiq_passive; /* passive IPI messages */
101 static __int64_t ipiq_cscount; /* number of cpu synchronizations */
102 static int ipiq_optimized = 1; /* XXX temporary sysctl */
104 static int panic_ipiq_cpu = -1;
105 static int panic_ipiq_count = 100;
112 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
113 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
114 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_avoided, CTLFLAG_RW, &ipiq_avoided, 0, "");
115 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_passive, CTLFLAG_RW, &ipiq_passive, 0, "");
116 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0, "");
117 SYSCTL_INT(_lwkt, OID_AUTO, ipiq_optimized, CTLFLAG_RW, &ipiq_optimized, 0, "");
119 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_cpu, CTLFLAG_RW, &panic_ipiq_cpu, 0, "");
120 SYSCTL_INT(_lwkt, OID_AUTO, panic_ipiq_count, CTLFLAG_RW, &panic_ipiq_count, 0, "");
123 #define IPIQ_STRING "func=%p arg1=%p arg2=%d scpu=%d dcpu=%d"
124 #define IPIQ_ARG_SIZE (sizeof(void *) * 2 + sizeof(int) * 2)
126 #if !defined(KTR_IPIQ)
127 #define KTR_IPIQ KTR_ALL
129 KTR_INFO_MASTER(ipiq);
130 KTR_INFO(KTR_IPIQ, ipiq, send_norm, 0, IPIQ_STRING, IPIQ_ARG_SIZE);
131 KTR_INFO(KTR_IPIQ, ipiq, send_pasv, 1, IPIQ_STRING, IPIQ_ARG_SIZE);
132 KTR_INFO(KTR_IPIQ, ipiq, send_nbio, 2, IPIQ_STRING, IPIQ_ARG_SIZE);
133 KTR_INFO(KTR_IPIQ, ipiq, send_fail, 3, IPIQ_STRING, IPIQ_ARG_SIZE);
134 KTR_INFO(KTR_IPIQ, ipiq, receive, 4, IPIQ_STRING, IPIQ_ARG_SIZE);
136 #define logipiq(name, func, arg1, arg2, sgd, dgd) \
137 KTR_LOG(ipiq_ ## name, func, arg1, arg2, sgd->gd_cpuid, dgd->gd_cpuid)
144 static int lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
145 struct intrframe *frame);
146 static void lwkt_cpusync_remote1(lwkt_cpusync_t poll);
147 static void lwkt_cpusync_remote2(lwkt_cpusync_t poll);
150 * Send a function execution request to another cpu. The request is queued
151 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
152 * possible target cpu. The FIFO can be written.
154 * If the FIFO fills up we have to enable interrupts to avoid an APIC
155 * deadlock and process pending IPIQs while waiting for it to empty.
156 * Otherwise we may soft-deadlock with another cpu whos FIFO is also full.
158 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
159 * end will take care of any pending interrupts.
161 * The actual hardware IPI is avoided if the target cpu is already processing
162 * the queue from a prior IPI. It is possible to pipeline IPI messages
163 * very quickly between cpus due to the FIFO hysteresis.
165 * Need not be called from a critical section.
168 lwkt_send_ipiq3(globaldata_t target, ipifunc3_t func, void *arg1, int arg2)
172 struct globaldata *gd = mycpu;
174 logipiq(send_norm, func, arg1, arg2, gd, target);
177 func(arg1, arg2, NULL);
181 ++gd->gd_intr_nesting_level;
183 if (gd->gd_intr_nesting_level > 20)
184 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
186 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
188 ip = &gd->gd_ipiq[target->gd_cpuid];
191 * Do not allow the FIFO to become full. Interrupts must be physically
192 * enabled while we liveloop to avoid deadlocking the APIC.
194 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
195 unsigned int eflags = read_eflags();
197 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
198 cpu_send_ipiq(target->gd_cpuid);
201 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
202 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
205 write_eflags(eflags);
209 * Queue the new message
211 windex = ip->ip_windex & MAXCPUFIFO_MASK;
212 ip->ip_func[windex] = func;
213 ip->ip_arg1[windex] = arg1;
214 ip->ip_arg2[windex] = arg2;
217 --gd->gd_intr_nesting_level;
220 * signal the target cpu that there is work pending.
222 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
223 cpu_send_ipiq(target->gd_cpuid);
225 if (ipiq_optimized == 0)
226 cpu_send_ipiq(target->gd_cpuid);
230 return(ip->ip_windex);
234 * Similar to lwkt_send_ipiq() but this function does not actually initiate
235 * the IPI to the target cpu unless the FIFO has become too full, so it is
238 * This function is used for non-critical IPI messages, such as memory
239 * deallocations. The queue will typically be flushed by the target cpu at
240 * the next clock interrupt.
242 * Need not be called from a critical section.
245 lwkt_send_ipiq3_passive(globaldata_t target, ipifunc3_t func,
246 void *arg1, int arg2)
250 struct globaldata *gd = mycpu;
252 KKASSERT(target != gd);
254 logipiq(send_pasv, func, arg1, arg2, gd, target);
255 ++gd->gd_intr_nesting_level;
257 if (gd->gd_intr_nesting_level > 20)
258 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
260 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
263 ip = &gd->gd_ipiq[target->gd_cpuid];
266 * Do not allow the FIFO to become full. Interrupts must be physically
267 * enabled while we liveloop to avoid deadlocking the APIC.
269 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
270 unsigned int eflags = read_eflags();
272 if (atomic_poll_acquire_int(&ip->ip_npoll) || ipiq_optimized == 0)
273 cpu_send_ipiq(target->gd_cpuid);
276 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
277 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
280 write_eflags(eflags);
284 * Queue the new message
286 windex = ip->ip_windex & MAXCPUFIFO_MASK;
287 ip->ip_func[windex] = func;
288 ip->ip_arg1[windex] = arg1;
289 ip->ip_arg2[windex] = arg2;
292 --gd->gd_intr_nesting_level;
295 * Do not signal the target cpu, it will pick up the IPI when it next
296 * polls (typically on the next tick).
299 return(ip->ip_windex);
303 * Send an IPI request without blocking, return 0 on success, ENOENT on
304 * failure. The actual queueing of the hardware IPI may still force us
305 * to spin and process incoming IPIs but that will eventually go away
306 * when we've gotten rid of the other general IPIs.
309 lwkt_send_ipiq3_nowait(globaldata_t target, ipifunc3_t func,
310 void *arg1, int arg2)
314 struct globaldata *gd = mycpu;
316 logipiq(send_nbio, func, arg1, arg2, gd, target);
317 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
319 func(arg1, arg2, NULL);
323 ip = &gd->gd_ipiq[target->gd_cpuid];
325 if (ip->ip_windex - ip->ip_rindex >= MAXCPUFIFO * 2 / 3) {
326 logipiq(send_fail, func, arg1, arg2, gd, target);
329 windex = ip->ip_windex & MAXCPUFIFO_MASK;
330 ip->ip_func[windex] = func;
331 ip->ip_arg1[windex] = arg1;
332 ip->ip_arg2[windex] = arg2;
337 * This isn't a passive IPI, we still have to signal the target cpu.
339 if (atomic_poll_acquire_int(&ip->ip_npoll)) {
340 cpu_send_ipiq(target->gd_cpuid);
342 if (ipiq_optimized == 0)
343 cpu_send_ipiq(target->gd_cpuid);
351 * deprecated, used only by fast int forwarding.
354 lwkt_send_ipiq3_bycpu(int dcpu, ipifunc3_t func, void *arg1, int arg2)
356 return(lwkt_send_ipiq3(globaldata_find(dcpu), func, arg1, arg2));
360 * Send a message to several target cpus. Typically used for scheduling.
361 * The message will not be sent to stopped cpus.
364 lwkt_send_ipiq3_mask(u_int32_t mask, ipifunc3_t func, void *arg1, int arg2)
369 mask &= ~stopped_cpus;
372 lwkt_send_ipiq3(globaldata_find(cpuid), func, arg1, arg2);
373 mask &= ~(1 << cpuid);
380 * Wait for the remote cpu to finish processing a function.
382 * YYY we have to enable interrupts and process the IPIQ while waiting
383 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
384 * function to do this! YYY we really should 'block' here.
386 * MUST be called from a critical section. This routine may be called
387 * from an interrupt (for example, if an interrupt wakes a foreign thread
391 lwkt_wait_ipiq(globaldata_t target, int seq)
394 int maxc = 100000000;
396 if (target != mycpu) {
397 ip = &mycpu->gd_ipiq[target->gd_cpuid];
398 if ((int)(ip->ip_xindex - seq) < 0) {
399 unsigned int eflags = read_eflags();
401 while ((int)(ip->ip_xindex - seq) < 0) {
406 kprintf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
408 panic("LWKT_WAIT_IPIQ");
410 * xindex may be modified by another cpu, use a load fence
411 * to ensure that the loop does not use a speculative value
412 * (which may improve performance).
416 write_eflags(eflags);
422 lwkt_seq_ipiq(globaldata_t target)
426 ip = &mycpu->gd_ipiq[target->gd_cpuid];
427 return(ip->ip_windex);
431 * Called from IPI interrupt (like a fast interrupt), which has placed
432 * us in a critical section. The MP lock may or may not be held.
433 * May also be called from doreti or splz, or be reentrantly called
434 * indirectly through the ip_func[] we run.
436 * There are two versions, one where no interrupt frame is available (when
437 * called from the send code and from splz, and one where an interrupt
438 * frame is available.
441 lwkt_process_ipiq(void)
443 globaldata_t gd = mycpu;
449 for (n = 0; n < ncpus; ++n) {
450 if (n != gd->gd_cpuid) {
451 sgd = globaldata_find(n);
454 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], NULL))
459 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
460 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, NULL)) {
461 if (gd->gd_curthread->td_cscount == 0)
470 lwkt_process_ipiq_frame(struct intrframe frame)
472 globaldata_t gd = mycpu;
478 for (n = 0; n < ncpus; ++n) {
479 if (n != gd->gd_cpuid) {
480 sgd = globaldata_find(n);
483 while (lwkt_process_ipiq_core(sgd, &ip[gd->gd_cpuid], &frame))
488 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
489 if (lwkt_process_ipiq_core(gd, &gd->gd_cpusyncq, &frame)) {
490 if (gd->gd_curthread->td_cscount == 0)
499 lwkt_process_ipiq_core(globaldata_t sgd, lwkt_ipiq_t ip,
500 struct intrframe *frame)
504 ipifunc3_t copy_func;
509 * Obtain the current write index, which is modified by a remote cpu.
510 * Issue a load fence to prevent speculative reads of e.g. data written
511 * by the other cpu prior to it updating the index.
513 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
518 * Note: xindex is only updated after we are sure the function has
519 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
520 * function may send an IPI which may block/drain.
522 * Note: due to additional IPI operations that the callback function
523 * may make, it is possible for both rindex and windex to advance and
524 * thus for rindex to advance passed our cached windex.
526 while (wi - (ri = ip->ip_rindex) > 0) {
527 ri &= MAXCPUFIFO_MASK;
528 copy_func = ip->ip_func[ri];
529 copy_arg1 = ip->ip_arg1[ri];
530 copy_arg2 = ip->ip_arg2[ri];
533 KKASSERT((ip->ip_rindex & MAXCPUFIFO_MASK) == ((ri + 1) & MAXCPUFIFO_MASK));
534 logipiq(receive, copy_func, copy_arg1, copy_arg2, sgd, mycpu);
535 copy_func(copy_arg1, copy_arg2, frame);
537 ip->ip_xindex = ip->ip_rindex;
541 * Simulate panics during the processing of an IPI
543 if (mycpu->gd_cpuid == panic_ipiq_cpu && panic_ipiq_count) {
544 if (--panic_ipiq_count == 0) {
546 Debugger("PANIC_DEBUG");
548 panic("PANIC_DEBUG");
556 * Return non-zero if there are more IPI messages pending on this
557 * ipiq. ip_npoll is left set as long as possible to reduce the
558 * number of IPIs queued by the originating cpu, but must be cleared
559 * *BEFORE* checking windex.
561 atomic_poll_release_int(&ip->ip_npoll);
562 return(wi != ip->ip_windex);
568 * CPU Synchronization Support
570 * lwkt_cpusync_simple()
572 * The function is executed synchronously before return on remote cpus.
573 * A lwkt_cpusync_t pointer is passed as an argument. The data can
574 * be accessed via arg->cs_data.
576 * XXX should I just pass the data as an argument to be consistent?
580 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data)
582 struct lwkt_cpusync cmd;
584 cmd.cs_run_func = NULL;
585 cmd.cs_fin1_func = func;
586 cmd.cs_fin2_func = NULL;
588 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
589 if (mask & (1 << mycpu->gd_cpuid))
591 lwkt_cpusync_finish(&cmd);
595 * lwkt_cpusync_fastdata()
597 * The function is executed in tandem with return on remote cpus.
598 * The data is directly passed as an argument. Do not pass pointers to
599 * temporary storage as the storage might have
600 * gone poof by the time the target cpu executes
603 * At the moment lwkt_cpusync is declared on the stack and we must wait
604 * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future
605 * optimization we should be able to put a counter in the globaldata
606 * structure (if it is not otherwise being used) and just poke it and
607 * return without waiting. XXX
610 lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data)
612 struct lwkt_cpusync cmd;
614 cmd.cs_run_func = NULL;
615 cmd.cs_fin1_func = NULL;
616 cmd.cs_fin2_func = func;
618 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
619 if (mask & (1 << mycpu->gd_cpuid))
621 lwkt_cpusync_finish(&cmd);
625 * lwkt_cpusync_start()
627 * Start synchronization with a set of target cpus, return once they are
628 * known to be in a synchronization loop. The target cpus will execute
629 * poll->cs_run_func() IN TANDEM WITH THE RETURN.
631 * XXX future: add lwkt_cpusync_start_quick() and require a call to
632 * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to
633 * potentially absorb the IPI latency doing something useful.
636 lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll)
638 globaldata_t gd = mycpu;
641 poll->cs_mask = mask;
643 poll->cs_maxcount = lwkt_send_ipiq_mask(
644 mask & gd->gd_other_cpus & smp_active_mask,
645 (ipifunc1_t)lwkt_cpusync_remote1, poll);
647 if (mask & gd->gd_cpumask) {
648 if (poll->cs_run_func)
649 poll->cs_run_func(poll);
652 if (poll->cs_maxcount) {
654 ++gd->gd_curthread->td_cscount;
655 while (poll->cs_count != poll->cs_maxcount) {
665 lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll)
667 globaldata_t gd = mycpu;
672 mask &= ~poll->cs_mask;
673 poll->cs_mask |= mask;
675 count = lwkt_send_ipiq_mask(
676 mask & gd->gd_other_cpus & smp_active_mask,
677 (ipifunc1_t)lwkt_cpusync_remote1, poll);
679 if (mask & gd->gd_cpumask) {
680 if (poll->cs_run_func)
681 poll->cs_run_func(poll);
684 poll->cs_maxcount += count;
685 if (poll->cs_maxcount) {
686 if (poll->cs_maxcount == count)
687 ++gd->gd_curthread->td_cscount;
688 while (poll->cs_count != poll->cs_maxcount) {
698 * Finish synchronization with a set of target cpus. The target cpus will
699 * execute cs_fin1_func(poll) prior to this function returning, and will
700 * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN.
702 * If cs_maxcount is non-zero then we are mastering a cpusync with one or
703 * more remote cpus and must account for it in our thread structure.
706 lwkt_cpusync_finish(lwkt_cpusync_t poll)
708 globaldata_t gd = mycpu;
711 if (poll->cs_mask & gd->gd_cpumask) {
712 if (poll->cs_fin1_func)
713 poll->cs_fin1_func(poll);
714 if (poll->cs_fin2_func)
715 poll->cs_fin2_func(poll->cs_data);
718 if (poll->cs_maxcount) {
719 while (poll->cs_count != -(poll->cs_maxcount + 1)) {
724 --gd->gd_curthread->td_cscount;
732 * helper IPI remote messaging function.
734 * Called on remote cpu when a new cpu synchronization request has been
735 * sent to us. Execute the run function and adjust cs_count, then requeue
736 * the request so we spin on it.
739 lwkt_cpusync_remote1(lwkt_cpusync_t poll)
741 atomic_add_int(&poll->cs_count, 1);
742 if (poll->cs_run_func)
743 poll->cs_run_func(poll);
744 lwkt_cpusync_remote2(poll);
748 * helper IPI remote messaging function.
750 * Poll for the originator telling us to finish. If it hasn't, requeue
751 * our request so we spin on it. When the originator requests that we
752 * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data)
753 * in tandem with the release.
756 lwkt_cpusync_remote2(lwkt_cpusync_t poll)
758 if (poll->cs_count < 0) {
759 cpusync_func2_t savef;
762 if (poll->cs_fin1_func)
763 poll->cs_fin1_func(poll);
764 if (poll->cs_fin2_func) {
765 savef = poll->cs_fin2_func;
766 saved = poll->cs_data;
767 atomic_add_int(&poll->cs_count, -1);
770 atomic_add_int(&poll->cs_count, -1);
773 globaldata_t gd = mycpu;
777 ip = &gd->gd_cpusyncq;
778 wi = ip->ip_windex & MAXCPUFIFO_MASK;
779 ip->ip_func[wi] = (ipifunc3_t)(ipifunc1_t)lwkt_cpusync_remote2;
780 ip->ip_arg1[wi] = poll;