2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * $DragonFly: src/sys/kern/lwkt_ipiq.c,v 1.2 2004/02/15 05:15:25 dillon Exp $
30 * This module implements IPI message queueing and the MI portion of IPI
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
40 #include <sys/rtprio.h>
41 #include <sys/queue.h>
42 #include <sys/thread2.h>
43 #include <sys/sysctl.h>
44 #include <sys/kthread.h>
45 #include <machine/cpu.h>
50 #include <vm/vm_param.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_page.h>
54 #include <vm/vm_map.h>
55 #include <vm/vm_pager.h>
56 #include <vm/vm_extern.h>
57 #include <vm/vm_zone.h>
59 #include <machine/stdarg.h>
60 #include <machine/ipl.h>
61 #include <machine/smp.h>
62 #include <machine/atomic.h>
64 #define THREAD_STACK (UPAGES * PAGE_SIZE)
68 #include <sys/stdint.h>
69 #include <libcaps/thread.h>
70 #include <sys/thread.h>
71 #include <sys/msgport.h>
72 #include <sys/errno.h>
73 #include <libcaps/globaldata.h>
74 #include <sys/thread2.h>
75 #include <sys/msgport2.h>
79 #include <machine/cpufunc.h>
80 #include <machine/lock.h>
85 static __int64_t ipiq_count = 0;
86 static __int64_t ipiq_fifofull = 0;
92 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
93 SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
100 static int lwkt_process_ipiq1(lwkt_ipiq_t ip, struct intrframe *frame);
101 static void lwkt_cpusync_remote1(lwkt_cpusync_t poll);
102 static void lwkt_cpusync_remote2(lwkt_cpusync_t poll);
105 * Send a function execution request to another cpu. The request is queued
106 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
107 * possible target cpu. The FIFO can be written.
109 * YYY If the FIFO fills up we have to enable interrupts and process the
110 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
111 * Create a CPU_*() function to do this!
113 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
114 * end will take care of any pending interrupts.
116 * Must be called from a critical section.
119 lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg)
123 struct globaldata *gd = mycpu;
130 ++gd->gd_intr_nesting_level;
132 if (gd->gd_intr_nesting_level > 20)
133 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
135 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
137 ip = &gd->gd_ipiq[target->gd_cpuid];
140 * We always drain before the FIFO becomes full so it should never
141 * become full. We need to leave enough entries to deal with
144 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
145 windex = ip->ip_windex & MAXCPUFIFO_MASK;
146 ip->ip_func[windex] = (ipifunc2_t)func;
147 ip->ip_arg[windex] = arg;
148 /* YYY memory barrier */
150 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
151 unsigned int eflags = read_eflags();
154 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
155 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
158 write_eflags(eflags);
160 --gd->gd_intr_nesting_level;
161 cpu_send_ipiq(target->gd_cpuid); /* issues mem barrier if appropriate */
163 return(ip->ip_windex);
167 * deprecated, used only by fast int forwarding.
170 lwkt_send_ipiq_bycpu(int dcpu, ipifunc_t func, void *arg)
172 return(lwkt_send_ipiq(globaldata_find(dcpu), func, arg));
176 * Send a message to several target cpus. Typically used for scheduling.
177 * The message will not be sent to stopped cpus.
180 lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
185 mask &= ~stopped_cpus;
188 lwkt_send_ipiq(globaldata_find(cpuid), func, arg);
189 mask &= ~(1 << cpuid);
196 * Wait for the remote cpu to finish processing a function.
198 * YYY we have to enable interrupts and process the IPIQ while waiting
199 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
200 * function to do this! YYY we really should 'block' here.
202 * MUST be called from a critical section. This routine may be called
203 * from an interrupt (for example, if an interrupt wakes a foreign thread
207 lwkt_wait_ipiq(globaldata_t target, int seq)
210 int maxc = 100000000;
212 if (target != mycpu) {
213 ip = &mycpu->gd_ipiq[target->gd_cpuid];
214 if ((int)(ip->ip_xindex - seq) < 0) {
215 unsigned int eflags = read_eflags();
217 while ((int)(ip->ip_xindex - seq) < 0) {
220 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
222 panic("LWKT_WAIT_IPIQ");
224 write_eflags(eflags);
230 * Called from IPI interrupt (like a fast interrupt), which has placed
231 * us in a critical section. The MP lock may or may not be held.
232 * May also be called from doreti or splz, or be reentrantly called
233 * indirectly through the ip_func[] we run.
235 * There are two versions, one where no interrupt frame is available (when
236 * called from the send code and from splz, and one where an interrupt
237 * frame is available.
240 lwkt_process_ipiq(void)
242 globaldata_t gd = mycpu;
247 for (n = 0; n < ncpus; ++n) {
248 if (n != gd->gd_cpuid) {
249 ip = globaldata_find(n)->gd_ipiq;
251 while (lwkt_process_ipiq1(&ip[gd->gd_cpuid], NULL))
256 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
257 if (lwkt_process_ipiq1(&gd->gd_cpusyncq, NULL))
264 lwkt_process_ipiq_frame(struct intrframe frame)
266 globaldata_t gd = mycpu;
271 for (n = 0; n < ncpus; ++n) {
272 if (n != gd->gd_cpuid) {
273 ip = globaldata_find(n)->gd_ipiq;
275 while (lwkt_process_ipiq1(&ip[gd->gd_cpuid], &frame))
280 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
281 if (lwkt_process_ipiq1(&gd->gd_cpusyncq, &frame))
288 lwkt_process_ipiq1(lwkt_ipiq_t ip, struct intrframe *frame)
291 int wi = ip->ip_windex;
293 * Note: xindex is only updated after we are sure the function has
294 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
295 * function may send an IPI which may block/drain.
297 while ((ri = ip->ip_rindex) != wi) {
298 ip->ip_rindex = ri + 1;
299 ri &= MAXCPUFIFO_MASK;
300 ip->ip_func[ri](ip->ip_arg[ri], frame);
301 /* YYY memory barrier */
302 ip->ip_xindex = ip->ip_rindex;
304 return(wi != ip->ip_windex);
308 * CPU Synchronization Support
310 * lwkt_cpusync_simple()
312 * The function is executed synchronously before return on remote cpus.
313 * A lwkt_cpusync_t pointer is passed as an argument. The data can
314 * be accessed via arg->cs_data.
316 * XXX should I just pass the data as an argument to be consistent?
320 lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data)
322 struct lwkt_cpusync cmd;
324 cmd.cs_run_func = NULL;
325 cmd.cs_fin1_func = func;
326 cmd.cs_fin2_func = NULL;
328 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
329 if (mask & (1 << mycpu->gd_cpuid))
331 lwkt_cpusync_finish(&cmd);
335 * lwkt_cpusync_fastdata()
337 * The function is executed in tandem with return on remote cpus.
338 * The data is directly passed as an argument. Do not pass pointers to
339 * temporary storage as the storage might have
340 * gone poof by the time the target cpu executes
343 * At the moment lwkt_cpusync is declared on the stack and we must wait
344 * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future
345 * optimization we should be able to put a counter in the globaldata
346 * structure (if it is not otherwise being used) and just poke it and
347 * return without waiting. XXX
350 lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data)
352 struct lwkt_cpusync cmd;
354 cmd.cs_run_func = NULL;
355 cmd.cs_fin1_func = NULL;
356 cmd.cs_fin2_func = func;
358 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
359 if (mask & (1 << mycpu->gd_cpuid))
361 lwkt_cpusync_finish(&cmd);
365 * lwkt_cpusync_start()
367 * Start synchronization with a set of target cpus, return once they are
368 * known to be in a synchronization loop. The target cpus will execute
369 * poll->cs_run_func() IN TANDEM WITH THE RETURN.
371 * XXX future: add lwkt_cpusync_start_quick() and require a call to
372 * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to
373 * potentially absorb the IPI latency doing something useful.
376 lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll)
379 poll->cs_mask = mask;
380 poll->cs_maxcount = lwkt_send_ipiq_mask(mask & mycpu->gd_other_cpus,
381 (ipifunc_t)lwkt_cpusync_remote1, poll);
382 if (mask & (1 << mycpu->gd_cpuid)) {
383 if (poll->cs_run_func)
384 poll->cs_run_func(poll);
386 while (poll->cs_count != poll->cs_maxcount) {
394 lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll)
396 mask &= ~poll->cs_mask;
397 poll->cs_mask |= mask;
398 poll->cs_maxcount += lwkt_send_ipiq_mask(mask & mycpu->gd_other_cpus,
399 (ipifunc_t)lwkt_cpusync_remote1, poll);
400 if (mask & (1 << mycpu->gd_cpuid)) {
401 if (poll->cs_run_func)
402 poll->cs_run_func(poll);
404 while (poll->cs_count != poll->cs_maxcount) {
412 * Finish synchronization with a set of target cpus. The target cpus will
413 * execute cs_fin1_func(poll) prior to this function returning, and will
414 * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN.
417 lwkt_cpusync_finish(lwkt_cpusync_t poll)
421 count = -(poll->cs_maxcount + 1);
423 if (poll->cs_mask & (1 << mycpu->gd_cpuid)) {
424 if (poll->cs_fin1_func)
425 poll->cs_fin1_func(poll);
426 if (poll->cs_fin2_func)
427 poll->cs_fin2_func(poll->cs_data);
429 while (poll->cs_count != count) {
437 * helper IPI remote messaging function.
439 * Called on remote cpu when a new cpu synchronization request has been
440 * sent to us. Execute the run function and adjust cs_count, then requeue
441 * the request so we spin on it.
444 lwkt_cpusync_remote1(lwkt_cpusync_t poll)
446 atomic_add_int(&poll->cs_count, 1);
447 if (poll->cs_run_func)
448 poll->cs_run_func(poll);
449 lwkt_cpusync_remote2(poll);
453 * helper IPI remote messaging function.
455 * Poll for the originator telling us to finish. If it hasn't, requeue
456 * our request so we spin on it. When the originator requests that we
457 * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data)
458 * in tandem with the release.
461 lwkt_cpusync_remote2(lwkt_cpusync_t poll)
463 if (poll->cs_count < 0) {
464 cpusync_func2_t savef;
467 if (poll->cs_fin1_func)
468 poll->cs_fin1_func(poll);
469 if (poll->cs_fin2_func) {
470 savef = poll->cs_fin2_func;
471 saved = poll->cs_data;
472 atomic_add_int(&poll->cs_count, -1);
475 atomic_add_int(&poll->cs_count, -1);
478 globaldata_t gd = mycpu;
482 ip = &gd->gd_cpusyncq;
483 wi = ip->ip_windex & MAXCPUFIFO_MASK;
484 ip->ip_func[wi] = (ipifunc2_t)lwkt_cpusync_remote2;
485 ip->ip_arg[wi] = poll;
493 * !SMP dummy routines
497 lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg)
499 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", target->gd_cpuid, func, arg);
500 return(0); /* NOT REACHED */
504 lwkt_wait_ipiq(globaldata_t target, int seq)
506 panic("lwkt_wait_ipiq: UP box! (%d,%d)", target->gd_cpuid, seq);