Merge FreeBSD rev. 1.8:
[dragonfly.git] / sys / kern / lwkt_ipiq.c
CommitLineData
3b6b7bd1
MD
1/*
2 * Copyright (c) 2003 Matthew Dillon <dillon@backplane.com>
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
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.
13 *
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
24 * SUCH DAMAGE.
25 *
0f7a3396 26 * $DragonFly: src/sys/kern/lwkt_ipiq.c,v 1.3 2004/02/17 19:38:49 dillon Exp $
3b6b7bd1
MD
27 */
28
29/*
30 * This module implements IPI message queueing and the MI portion of IPI
31 * message processing.
32 */
33
34#ifdef _KERNEL
35
36#include <sys/param.h>
37#include <sys/systm.h>
38#include <sys/kernel.h>
39#include <sys/proc.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>
46#include <sys/lock.h>
47#include <sys/caps.h>
48
49#include <vm/vm.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>
58
59#include <machine/stdarg.h>
60#include <machine/ipl.h>
61#include <machine/smp.h>
62#include <machine/atomic.h>
63
64#define THREAD_STACK (UPAGES * PAGE_SIZE)
65
66#else
67
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>
76#include <stdio.h>
77#include <stdlib.h>
78#include <string.h>
79#include <machine/cpufunc.h>
80#include <machine/lock.h>
81
82#endif
83
84#ifdef SMP
0f7a3396
MD
85static __int64_t ipiq_count;
86static __int64_t ipiq_fifofull;
87static __int64_t ipiq_cscount;
3b6b7bd1
MD
88#endif
89
90#ifdef _KERNEL
91
92#ifdef SMP
93SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_count, CTLFLAG_RW, &ipiq_count, 0, "");
94SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_fifofull, CTLFLAG_RW, &ipiq_fifofull, 0, "");
0f7a3396 95SYSCTL_QUAD(_lwkt, OID_AUTO, ipiq_cscount, CTLFLAG_RW, &ipiq_cscount, 0, "");
3b6b7bd1
MD
96#endif
97
98#endif
99
100#ifdef SMP
101
102static int lwkt_process_ipiq1(lwkt_ipiq_t ip, struct intrframe *frame);
103static void lwkt_cpusync_remote1(lwkt_cpusync_t poll);
104static void lwkt_cpusync_remote2(lwkt_cpusync_t poll);
105
106/*
107 * Send a function execution request to another cpu. The request is queued
108 * on the cpu<->cpu ipiq matrix. Each cpu owns a unique ipiq FIFO for every
109 * possible target cpu. The FIFO can be written.
110 *
111 * YYY If the FIFO fills up we have to enable interrupts and process the
112 * IPIQ while waiting for it to empty or we may deadlock with another cpu.
113 * Create a CPU_*() function to do this!
114 *
115 * We can safely bump gd_intr_nesting_level because our crit_exit() at the
116 * end will take care of any pending interrupts.
117 *
118 * Must be called from a critical section.
119 */
120int
121lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg)
122{
123 lwkt_ipiq_t ip;
124 int windex;
125 struct globaldata *gd = mycpu;
126
127 if (target == gd) {
128 func(arg);
129 return(0);
130 }
131 crit_enter();
132 ++gd->gd_intr_nesting_level;
133#ifdef INVARIANTS
134 if (gd->gd_intr_nesting_level > 20)
135 panic("lwkt_send_ipiq: TOO HEAVILY NESTED!");
136#endif
137 KKASSERT(curthread->td_pri >= TDPRI_CRIT);
138 ++ipiq_count;
139 ip = &gd->gd_ipiq[target->gd_cpuid];
140
141 /*
142 * We always drain before the FIFO becomes full so it should never
143 * become full. We need to leave enough entries to deal with
144 * reentrancy.
145 */
146 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO);
147 windex = ip->ip_windex & MAXCPUFIFO_MASK;
148 ip->ip_func[windex] = (ipifunc2_t)func;
149 ip->ip_arg[windex] = arg;
150 /* YYY memory barrier */
151 ++ip->ip_windex;
152 if (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 2) {
153 unsigned int eflags = read_eflags();
154 cpu_enable_intr();
155 ++ipiq_fifofull;
156 while (ip->ip_windex - ip->ip_rindex > MAXCPUFIFO / 4) {
157 KKASSERT(ip->ip_windex - ip->ip_rindex != MAXCPUFIFO - 1);
158 lwkt_process_ipiq();
159 }
160 write_eflags(eflags);
161 }
162 --gd->gd_intr_nesting_level;
163 cpu_send_ipiq(target->gd_cpuid); /* issues mem barrier if appropriate */
164 crit_exit();
165 return(ip->ip_windex);
166}
167
168/*
169 * deprecated, used only by fast int forwarding.
170 */
171int
172lwkt_send_ipiq_bycpu(int dcpu, ipifunc_t func, void *arg)
173{
174 return(lwkt_send_ipiq(globaldata_find(dcpu), func, arg));
175}
176
177/*
178 * Send a message to several target cpus. Typically used for scheduling.
179 * The message will not be sent to stopped cpus.
180 */
181int
182lwkt_send_ipiq_mask(u_int32_t mask, ipifunc_t func, void *arg)
183{
184 int cpuid;
185 int count = 0;
186
187 mask &= ~stopped_cpus;
188 while (mask) {
189 cpuid = bsfl(mask);
190 lwkt_send_ipiq(globaldata_find(cpuid), func, arg);
191 mask &= ~(1 << cpuid);
192 ++count;
193 }
194 return(count);
195}
196
197/*
198 * Wait for the remote cpu to finish processing a function.
199 *
200 * YYY we have to enable interrupts and process the IPIQ while waiting
201 * for it to empty or we may deadlock with another cpu. Create a CPU_*()
202 * function to do this! YYY we really should 'block' here.
203 *
204 * MUST be called from a critical section. This routine may be called
205 * from an interrupt (for example, if an interrupt wakes a foreign thread
206 * up).
207 */
208void
209lwkt_wait_ipiq(globaldata_t target, int seq)
210{
211 lwkt_ipiq_t ip;
212 int maxc = 100000000;
213
214 if (target != mycpu) {
215 ip = &mycpu->gd_ipiq[target->gd_cpuid];
216 if ((int)(ip->ip_xindex - seq) < 0) {
217 unsigned int eflags = read_eflags();
218 cpu_enable_intr();
219 while ((int)(ip->ip_xindex - seq) < 0) {
220 lwkt_process_ipiq();
221 if (--maxc == 0)
222 printf("LWKT_WAIT_IPIQ WARNING! %d wait %d (%d)\n", mycpu->gd_cpuid, target->gd_cpuid, ip->ip_xindex - seq);
223 if (maxc < -1000000)
224 panic("LWKT_WAIT_IPIQ");
225 }
226 write_eflags(eflags);
227 }
228 }
229}
230
231/*
232 * Called from IPI interrupt (like a fast interrupt), which has placed
233 * us in a critical section. The MP lock may or may not be held.
234 * May also be called from doreti or splz, or be reentrantly called
235 * indirectly through the ip_func[] we run.
236 *
237 * There are two versions, one where no interrupt frame is available (when
238 * called from the send code and from splz, and one where an interrupt
239 * frame is available.
240 */
241void
242lwkt_process_ipiq(void)
243{
244 globaldata_t gd = mycpu;
245 lwkt_ipiq_t ip;
246 int n;
247
248again:
249 for (n = 0; n < ncpus; ++n) {
250 if (n != gd->gd_cpuid) {
251 ip = globaldata_find(n)->gd_ipiq;
252 if (ip != NULL) {
253 while (lwkt_process_ipiq1(&ip[gd->gd_cpuid], NULL))
254 ;
255 }
256 }
257 }
258 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
0f7a3396
MD
259 if (lwkt_process_ipiq1(&gd->gd_cpusyncq, NULL)) {
260 if (gd->gd_curthread->td_cscount == 0)
261 goto again;
262 need_ipiq();
263 }
3b6b7bd1
MD
264 }
265}
266
267#ifdef _KERNEL
268void
269lwkt_process_ipiq_frame(struct intrframe frame)
270{
271 globaldata_t gd = mycpu;
272 lwkt_ipiq_t ip;
273 int n;
274
275again:
276 for (n = 0; n < ncpus; ++n) {
277 if (n != gd->gd_cpuid) {
278 ip = globaldata_find(n)->gd_ipiq;
279 if (ip != NULL) {
280 while (lwkt_process_ipiq1(&ip[gd->gd_cpuid], &frame))
281 ;
282 }
283 }
284 }
285 if (gd->gd_cpusyncq.ip_rindex != gd->gd_cpusyncq.ip_windex) {
0f7a3396
MD
286 if (lwkt_process_ipiq1(&gd->gd_cpusyncq, &frame)) {
287 if (gd->gd_curthread->td_cscount == 0)
288 goto again;
289 need_ipiq();
290 }
3b6b7bd1
MD
291 }
292}
293#endif
294
295static int
296lwkt_process_ipiq1(lwkt_ipiq_t ip, struct intrframe *frame)
297{
298 int ri;
299 int wi = ip->ip_windex;
300 /*
301 * Note: xindex is only updated after we are sure the function has
302 * finished execution. Beware lwkt_process_ipiq() reentrancy! The
303 * function may send an IPI which may block/drain.
304 */
305 while ((ri = ip->ip_rindex) != wi) {
306 ip->ip_rindex = ri + 1;
307 ri &= MAXCPUFIFO_MASK;
308 ip->ip_func[ri](ip->ip_arg[ri], frame);
309 /* YYY memory barrier */
310 ip->ip_xindex = ip->ip_rindex;
311 }
312 return(wi != ip->ip_windex);
313}
314
0f7a3396
MD
315#else
316
317/*
318 * !SMP dummy routines
319 */
320
321int
322lwkt_send_ipiq(globaldata_t target, ipifunc_t func, void *arg)
323{
324 panic("lwkt_send_ipiq: UP box! (%d,%p,%p)", target->gd_cpuid, func, arg);
325 return(0); /* NOT REACHED */
326}
327
328void
329lwkt_wait_ipiq(globaldata_t target, int seq)
330{
331 panic("lwkt_wait_ipiq: UP box! (%d,%d)", target->gd_cpuid, seq);
332}
333
334#endif
335
3b6b7bd1
MD
336/*
337 * CPU Synchronization Support
5c71a36a
MD
338 *
339 * lwkt_cpusync_simple()
340 *
341 * The function is executed synchronously before return on remote cpus.
342 * A lwkt_cpusync_t pointer is passed as an argument. The data can
343 * be accessed via arg->cs_data.
344 *
345 * XXX should I just pass the data as an argument to be consistent?
3b6b7bd1
MD
346 */
347
348void
5c71a36a
MD
349lwkt_cpusync_simple(cpumask_t mask, cpusync_func_t func, void *data)
350{
351 struct lwkt_cpusync cmd;
352
353 cmd.cs_run_func = NULL;
354 cmd.cs_fin1_func = func;
355 cmd.cs_fin2_func = NULL;
356 cmd.cs_data = data;
357 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
358 if (mask & (1 << mycpu->gd_cpuid))
359 func(&cmd);
360 lwkt_cpusync_finish(&cmd);
361}
362
363/*
364 * lwkt_cpusync_fastdata()
365 *
366 * The function is executed in tandem with return on remote cpus.
367 * The data is directly passed as an argument. Do not pass pointers to
368 * temporary storage as the storage might have
369 * gone poof by the time the target cpu executes
370 * the function.
371 *
372 * At the moment lwkt_cpusync is declared on the stack and we must wait
373 * for all remote cpus to ack in lwkt_cpusync_finish(), but as a future
374 * optimization we should be able to put a counter in the globaldata
375 * structure (if it is not otherwise being used) and just poke it and
376 * return without waiting. XXX
377 */
378void
379lwkt_cpusync_fastdata(cpumask_t mask, cpusync_func2_t func, void *data)
3b6b7bd1
MD
380{
381 struct lwkt_cpusync cmd;
3b6b7bd1
MD
382
383 cmd.cs_run_func = NULL;
384 cmd.cs_fin1_func = NULL;
385 cmd.cs_fin2_func = func;
5c71a36a
MD
386 cmd.cs_data = NULL;
387 lwkt_cpusync_start(mask & mycpu->gd_other_cpus, &cmd);
3b6b7bd1
MD
388 if (mask & (1 << mycpu->gd_cpuid))
389 func(data);
5c71a36a 390 lwkt_cpusync_finish(&cmd);
3b6b7bd1
MD
391}
392
393/*
5c71a36a
MD
394 * lwkt_cpusync_start()
395 *
396 * Start synchronization with a set of target cpus, return once they are
397 * known to be in a synchronization loop. The target cpus will execute
398 * poll->cs_run_func() IN TANDEM WITH THE RETURN.
399 *
400 * XXX future: add lwkt_cpusync_start_quick() and require a call to
401 * lwkt_cpusync_add() or lwkt_cpusync_wait(), allowing the caller to
402 * potentially absorb the IPI latency doing something useful.
3b6b7bd1 403 */
5c71a36a 404void
3b6b7bd1
MD
405lwkt_cpusync_start(cpumask_t mask, lwkt_cpusync_t poll)
406{
0f7a3396
MD
407 globaldata_t gd = mycpu;
408
3b6b7bd1 409 poll->cs_count = 0;
5c71a36a 410 poll->cs_mask = mask;
0f7a3396
MD
411#ifdef SMP
412 poll->cs_maxcount = lwkt_send_ipiq_mask(
413 mask & gd->gd_other_cpus & smp_active_mask,
414 (ipifunc_t)lwkt_cpusync_remote1, poll);
415#endif
416 if (mask & (1 << gd->gd_cpuid)) {
5c71a36a
MD
417 if (poll->cs_run_func)
418 poll->cs_run_func(poll);
419 }
0f7a3396
MD
420#ifdef SMP
421 if (poll->cs_maxcount) {
422 ++ipiq_cscount;
423 ++gd->gd_curthread->td_cscount;
424 while (poll->cs_count != poll->cs_maxcount) {
425 crit_enter();
426 lwkt_process_ipiq();
427 crit_exit();
428 }
5c71a36a 429 }
0f7a3396 430#endif
5c71a36a
MD
431}
432
433void
434lwkt_cpusync_add(cpumask_t mask, lwkt_cpusync_t poll)
435{
0f7a3396
MD
436 globaldata_t gd = mycpu;
437 int count;
438
5c71a36a
MD
439 mask &= ~poll->cs_mask;
440 poll->cs_mask |= mask;
0f7a3396
MD
441#ifdef SMP
442 count = lwkt_send_ipiq_mask(
443 mask & gd->gd_other_cpus & smp_active_mask,
444 (ipifunc_t)lwkt_cpusync_remote1, poll);
445#endif
446 if (mask & (1 << gd->gd_cpuid)) {
5c71a36a
MD
447 if (poll->cs_run_func)
448 poll->cs_run_func(poll);
449 }
0f7a3396
MD
450#ifdef SMP
451 poll->cs_maxcount += count;
452 if (poll->cs_maxcount) {
453 if (poll->cs_maxcount == count)
454 ++gd->gd_curthread->td_cscount;
455 while (poll->cs_count != poll->cs_maxcount) {
456 crit_enter();
457 lwkt_process_ipiq();
458 crit_exit();
459 }
3b6b7bd1 460 }
0f7a3396 461#endif
3b6b7bd1
MD
462}
463
464/*
465 * Finish synchronization with a set of target cpus. The target cpus will
466 * execute cs_fin1_func(poll) prior to this function returning, and will
467 * execute cs_fin2_func(data) IN TANDEM WITH THIS FUNCTION'S RETURN.
0f7a3396
MD
468 *
469 * If cs_maxcount is non-zero then we are mastering a cpusync with one or
470 * more remote cpus and must account for it in our thread structure.
3b6b7bd1
MD
471 */
472void
5c71a36a 473lwkt_cpusync_finish(lwkt_cpusync_t poll)
3b6b7bd1 474{
0f7a3396 475 globaldata_t gd = mycpu;
5c71a36a 476
3b6b7bd1 477 poll->cs_count = -1;
0f7a3396 478 if (poll->cs_mask & (1 << gd->gd_cpuid)) {
5c71a36a
MD
479 if (poll->cs_fin1_func)
480 poll->cs_fin1_func(poll);
481 if (poll->cs_fin2_func)
482 poll->cs_fin2_func(poll->cs_data);
483 }
0f7a3396
MD
484#ifdef SMP
485 if (poll->cs_maxcount) {
486 while (poll->cs_count != -(poll->cs_maxcount + 1)) {
487 crit_enter();
488 lwkt_process_ipiq();
489 crit_exit();
490 }
491 --gd->gd_curthread->td_cscount;
3b6b7bd1 492 }
0f7a3396 493#endif
3b6b7bd1
MD
494}
495
0f7a3396
MD
496#ifdef SMP
497
3b6b7bd1
MD
498/*
499 * helper IPI remote messaging function.
500 *
501 * Called on remote cpu when a new cpu synchronization request has been
502 * sent to us. Execute the run function and adjust cs_count, then requeue
503 * the request so we spin on it.
504 */
505static void
506lwkt_cpusync_remote1(lwkt_cpusync_t poll)
507{
508 atomic_add_int(&poll->cs_count, 1);
509 if (poll->cs_run_func)
510 poll->cs_run_func(poll);
511 lwkt_cpusync_remote2(poll);
512}
513
514/*
515 * helper IPI remote messaging function.
516 *
517 * Poll for the originator telling us to finish. If it hasn't, requeue
518 * our request so we spin on it. When the originator requests that we
519 * finish we execute cs_fin1_func(poll) synchronously and cs_fin2_func(data)
520 * in tandem with the release.
521 */
522static void
523lwkt_cpusync_remote2(lwkt_cpusync_t poll)
524{
525 if (poll->cs_count < 0) {
526 cpusync_func2_t savef;
527 void *saved;
528
529 if (poll->cs_fin1_func)
530 poll->cs_fin1_func(poll);
531 if (poll->cs_fin2_func) {
532 savef = poll->cs_fin2_func;
533 saved = poll->cs_data;
534 atomic_add_int(&poll->cs_count, -1);
535 savef(saved);
536 } else {
537 atomic_add_int(&poll->cs_count, -1);
538 }
539 } else {
540 globaldata_t gd = mycpu;
541 lwkt_ipiq_t ip;
542 int wi;
543
544 ip = &gd->gd_cpusyncq;
545 wi = ip->ip_windex & MAXCPUFIFO_MASK;
546 ip->ip_func[wi] = (ipifunc2_t)lwkt_cpusync_remote2;
547 ip->ip_arg[wi] = poll;
548 ++ip->ip_windex;
549 }
550}
551
3b6b7bd1 552#endif