xref: /netbsd-src/sys/rump/librump/rumpkern/scheduler.c (revision 7f21db1c0118155e0dd40b75182e30c589d9f63e)
1 /*      $NetBSD: scheduler.c,v 1.9 2010/01/25 18:37:51 pooka Exp $	*/
2 
3 /*
4  * Copyright (c) 2009 Antti Kantee.  All Rights Reserved.
5  *
6  * Development of this software was supported by
7  * The Finnish Cultural Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
19  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
20  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 __KERNEL_RCSID(0, "$NetBSD: scheduler.c,v 1.9 2010/01/25 18:37:51 pooka Exp $");
33 
34 #include <sys/param.h>
35 #include <sys/cpu.h>
36 #include <sys/kmem.h>
37 #include <sys/mutex.h>
38 #include <sys/namei.h>
39 #include <sys/queue.h>
40 #include <sys/select.h>
41 
42 #include <rump/rumpuser.h>
43 
44 #include "rump_private.h"
45 
46 /* should go for MAXCPUS at some point */
47 static struct cpu_info rump_cpus[MAXCPUS];
48 static struct rumpcpu {
49 	struct cpu_info *rcpu_ci;
50 	int rcpu_flags;
51 	struct rumpuser_cv *rcpu_cv;
52 	LIST_ENTRY(rumpcpu) rcpu_entries;
53 } rcpu_storage[MAXCPUS];
54 struct cpu_info *rump_cpu = &rump_cpus[0];
55 int ncpu = 1;
56 
57 #define RCPU_WANTED	0x01	/* someone wants this specific CPU */
58 #define RCPU_BUSY	0x02	/* CPU is busy */
59 #define RCPU_FREELIST	0x04	/* CPU is on freelist */
60 
61 static LIST_HEAD(,rumpcpu) cpu_freelist = LIST_HEAD_INITIALIZER(cpu_freelist);
62 static struct rumpuser_mtx *schedmtx;
63 static struct rumpuser_cv *schedcv, *lwp0cv;
64 
65 static bool lwp0busy = false;
66 
67 struct cpu_info *
68 cpu_lookup(u_int index)
69 {
70 
71 	return &rump_cpus[index];
72 }
73 
74 void
75 rump_scheduler_init()
76 {
77 	struct rumpcpu *rcpu;
78 	struct cpu_info *ci;
79 	int i;
80 
81 	rumpuser_mutex_init(&schedmtx);
82 	rumpuser_cv_init(&schedcv);
83 	rumpuser_cv_init(&lwp0cv);
84 	for (i = 0; i < ncpu; i++) {
85 		rcpu = &rcpu_storage[i];
86 		ci = &rump_cpus[i];
87 		rump_cpu_bootstrap(ci);
88 		ci->ci_schedstate.spc_mutex =
89 		    mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE);
90 		ci->ci_schedstate.spc_flags = SPCF_RUNNING;
91 		rcpu->rcpu_ci = ci;
92 		LIST_INSERT_HEAD(&cpu_freelist, rcpu, rcpu_entries);
93 		rcpu->rcpu_flags = RCPU_FREELIST;
94 		rumpuser_cv_init(&rcpu->rcpu_cv);
95 	}
96 }
97 
98 void
99 rump_schedule()
100 {
101 	struct lwp *l;
102 
103 	/*
104 	 * If there is no dedicated lwp, allocate a temp one and
105 	 * set it to be free'd upon unschedule().  Use lwp0 context
106 	 * for reserving the necessary resources.
107 	 */
108 	l = rumpuser_get_curlwp();
109 	if (l == NULL) {
110 		/* busy lwp0 */
111 		rumpuser_mutex_enter_nowrap(schedmtx);
112 		while (lwp0busy)
113 			rumpuser_cv_wait_nowrap(lwp0cv, schedmtx);
114 		lwp0busy = true;
115 		rumpuser_mutex_exit(schedmtx);
116 
117 		/* schedule cpu and use lwp0 */
118 		rump_schedule_cpu(&lwp0);
119 		rumpuser_set_curlwp(&lwp0);
120 		l = rump_lwp_alloc(0, rump_nextlid());
121 
122 		/* release lwp0 */
123 		rump_lwp_switch(l);
124 		rumpuser_mutex_enter_nowrap(schedmtx);
125 		lwp0busy = false;
126 		rumpuser_cv_signal(lwp0cv);
127 		rumpuser_mutex_exit(schedmtx);
128 
129 		/* mark new lwp as dead-on-exit */
130 		rump_lwp_release(l);
131 	} else {
132 		rump_schedule_cpu(l);
133 	}
134 }
135 
136 void
137 rump_schedule_cpu(struct lwp *l)
138 {
139 	struct rumpcpu *rcpu;
140 
141 	rumpuser_mutex_enter_nowrap(schedmtx);
142 	if (l->l_pflag & LP_BOUND) {
143 		KASSERT(l->l_cpu != NULL);
144 		rcpu = &rcpu_storage[l->l_cpu-&rump_cpus[0]];
145 		if (rcpu->rcpu_flags & RCPU_BUSY) {
146 			KASSERT((rcpu->rcpu_flags & RCPU_FREELIST) == 0);
147 			while (rcpu->rcpu_flags & RCPU_BUSY) {
148 				rcpu->rcpu_flags |= RCPU_WANTED;
149 				rumpuser_cv_wait_nowrap(rcpu->rcpu_cv,
150 				    schedmtx);
151 			}
152 			rcpu->rcpu_flags &= ~RCPU_WANTED;
153 		} else {
154 			KASSERT(rcpu->rcpu_flags & (RCPU_FREELIST|RCPU_WANTED));
155 		}
156 		if (rcpu->rcpu_flags & RCPU_FREELIST) {
157 			LIST_REMOVE(rcpu, rcpu_entries);
158 			rcpu->rcpu_flags &= ~RCPU_FREELIST;
159 		}
160 	} else {
161 		while ((rcpu = LIST_FIRST(&cpu_freelist)) == NULL) {
162 			rumpuser_cv_wait_nowrap(schedcv, schedmtx);
163 		}
164 		KASSERT(rcpu->rcpu_flags & RCPU_FREELIST);
165 		LIST_REMOVE(rcpu, rcpu_entries);
166 		rcpu->rcpu_flags &= ~RCPU_FREELIST;
167 		KASSERT(l->l_cpu == NULL);
168 		l->l_cpu = rcpu->rcpu_ci;
169 	}
170 	rcpu->rcpu_flags |= RCPU_BUSY;
171 	rumpuser_mutex_exit(schedmtx);
172 	l->l_mutex = rcpu->rcpu_ci->ci_schedstate.spc_mutex;
173 }
174 
175 void
176 rump_unschedule()
177 {
178 	struct lwp *l;
179 
180 	l = rumpuser_get_curlwp();
181 	KASSERT(l->l_mutex == l->l_cpu->ci_schedstate.spc_mutex);
182 	rump_unschedule_cpu(l);
183 	l->l_mutex = NULL;
184 
185 	/*
186 	 * If we're using a temp lwp, need to take lwp0 for rump_lwp_free().
187 	 * (we could maybe cache idle lwp's to avoid constant bouncing)
188 	 */
189 	if (l->l_flag & LW_WEXIT) {
190 		rumpuser_set_curlwp(NULL);
191 
192 		/* busy lwp0 */
193 		rumpuser_mutex_enter_nowrap(schedmtx);
194 		while (lwp0busy)
195 			rumpuser_cv_wait_nowrap(lwp0cv, schedmtx);
196 		lwp0busy = true;
197 		rumpuser_mutex_exit(schedmtx);
198 
199 		rump_schedule_cpu(&lwp0);
200 		rumpuser_set_curlwp(&lwp0);
201 		rump_lwp_free(l);
202 		rump_unschedule_cpu(&lwp0);
203 		rumpuser_set_curlwp(NULL);
204 
205 		rumpuser_mutex_enter_nowrap(schedmtx);
206 		lwp0busy = false;
207 		rumpuser_cv_signal(lwp0cv);
208 		rumpuser_mutex_exit(schedmtx);
209 	}
210 }
211 
212 void
213 rump_unschedule_cpu(struct lwp *l)
214 {
215 
216 	if ((l->l_pflag & LP_INTR) == 0)
217 		rump_softint_run(l->l_cpu);
218 	rump_unschedule_cpu1(l);
219 }
220 
221 void
222 rump_unschedule_cpu1(struct lwp *l)
223 {
224 	struct rumpcpu *rcpu;
225 	struct cpu_info *ci;
226 
227 	ci = l->l_cpu;
228 	if ((l->l_pflag & LP_BOUND) == 0) {
229 		l->l_cpu = NULL;
230 	}
231 	rcpu = &rcpu_storage[ci-&rump_cpus[0]];
232 	KASSERT(rcpu->rcpu_ci == ci);
233 	KASSERT(rcpu->rcpu_flags & RCPU_BUSY);
234 
235 	rumpuser_mutex_enter_nowrap(schedmtx);
236 	if (rcpu->rcpu_flags & RCPU_WANTED) {
237 		/*
238 		 * The assumption is that there will usually be max 1
239 		 * thread waiting on the rcpu_cv, so broadcast is fine.
240 		 * (and the current structure requires it because of
241 		 * only a bitmask being used for wanting).
242 		 */
243 		rumpuser_cv_broadcast(rcpu->rcpu_cv);
244 	} else {
245 		LIST_INSERT_HEAD(&cpu_freelist, rcpu, rcpu_entries);
246 		rcpu->rcpu_flags |= RCPU_FREELIST;
247 		rumpuser_cv_signal(schedcv);
248 	}
249 	rcpu->rcpu_flags &= ~RCPU_BUSY;
250 	rumpuser_mutex_exit(schedmtx);
251 }
252 
253 /* Give up and retake CPU (perhaps a different one) */
254 void
255 yield()
256 {
257 	struct lwp *l = curlwp;
258 	int nlocks;
259 
260 	KERNEL_UNLOCK_ALL(l, &nlocks);
261 	rump_unschedule_cpu(l);
262 	rump_schedule_cpu(l);
263 	KERNEL_LOCK(nlocks, l);
264 }
265 
266 void
267 preempt()
268 {
269 
270 	yield();
271 }
272