subr_localcount.c revision 1.1.6.4 1 /* $NetBSD: subr_localcount.c,v 1.1.6.4 2017/05/11 21:31:12 pgoyette Exp $ */
2
3 /*-
4 * Copyright (c) 2016 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Taylor R. Campbell.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
30 */
31
32 /*
33 * CPU-local reference counts
34 *
35 * localcount(9) is a reference-counting scheme that involves no
36 * interprocessor synchronization most of the time, at the cost of
37 * eight bytes of memory per CPU per object and at the cost of
38 * expensive interprocessor synchronization to drain references.
39 *
40 * localcount(9) references may be held across sleeps, may be
41 * transferred from CPU to CPU or thread to thread: they behave
42 * semantically like typical reference counts, with different
43 * pragmatic performance characteristics.
44 */
45
46 #include <sys/cdefs.h>
47 __KERNEL_RCSID(0, "$NetBSD: subr_localcount.c,v 1.1.6.4 2017/05/11 21:31:12 pgoyette Exp $");
48
49 #include <sys/param.h>
50 #include <sys/localcount.h>
51 #include <sys/types.h>
52 #include <sys/condvar.h>
53 #include <sys/errno.h>
54 #include <sys/mutex.h>
55 #include <sys/percpu.h>
56 #include <sys/xcall.h>
57
58 /*
59 * localcount_init(lc)
60 *
61 * Initialize a localcount object. Returns 0 on success, error
62 * code on failure. May fail to allocate memory for percpu(9).
63 *
64 * The caller must call localcount_drain and then localcount_fini
65 * when done with lc.
66 */
67 static void localcount_xc(void *, void *);
68
69 int
70 localcount_init(struct localcount *lc)
71 {
72
73 lc->lc_totalp = NULL;
74 lc->lc_percpu = percpu_alloc(sizeof(int64_t));
75 if (lc->lc_percpu == NULL)
76 return ENOMEM;
77
78 return 0;
79 }
80
81 /*
82 * localcount_drain(lc, cv, interlock)
83 *
84 * Wait for all acquired references to lc to drain. Caller must
85 * hold interlock; localcount_drain releases it during cross-calls
86 * and waits on cv. The cv and interlock passed here must be the
87 * same as are passed to localcount_release for this lc.
88 *
89 * Caller must guarantee that no new references can be acquired
90 * with localcount_acquire before calling localcount_drain. For
91 * example, any object that may be found in a list and acquired
92 * must be removed from the list before localcount_drain.
93 *
94 * The localcount object lc may be used only with localcount_fini
95 * after this, unless reinitialized after localcount_fini with
96 * localcount_init.
97 */
98 void
99 localcount_drain(struct localcount *lc, kcondvar_t *cv, kmutex_t *interlock)
100 {
101 int64_t total = 0;
102
103 KASSERT(mutex_owned(interlock));
104 KASSERT(lc->lc_totalp == NULL);
105
106 /* Mark it draining. */
107 lc->lc_totalp = &total;
108
109 /*
110 * Count up all references on all CPUs.
111 *
112 * This serves as a global memory barrier: after xc_wait, all
113 * CPUs will have witnessed the nonnull value of lc->lc_totalp,
114 * so that it is safe to wait on the cv for them.
115 */
116 mutex_exit(interlock);
117 xc_wait(xc_broadcast(0, &localcount_xc, lc, interlock));
118 mutex_enter(interlock);
119
120 /* Wait for remaining references to drain. */
121 while (total != 0) {
122 /*
123 * At this point, now that we have added up all
124 * references on all CPUs, the total had better be
125 * nonnegative.
126 */
127 KASSERTMSG((0 < total),
128 "negatively referenced localcount: %p, %"PRId64,
129 lc, total);
130 cv_wait(cv, interlock);
131 }
132
133 /* Paranoia: Cause any further use of lc->lc_totalp to crash. */
134 lc->lc_totalp = (void *)(uintptr_t)1;
135 }
136
137 /*
138 * localcount_fini(lc)
139 *
140 * Finalize a localcount object, releasing any memory allocated
141 * for it. Caller must have already called localcount_drain.
142 */
143 void
144 localcount_fini(struct localcount *lc)
145 {
146
147 KASSERT(lc->lc_totalp == (void *)(uintptr_t)1);
148 percpu_free(lc->lc_percpu, sizeof(uint64_t));
149 }
150
151 static void
152 localcount_xc(void *cookie0, void *cookie1)
153 {
154 struct localcount *lc = cookie0;
155 kmutex_t *interlock = cookie1;
156 int64_t *localp;
157
158 mutex_enter(interlock);
159 localp = percpu_getref(lc->lc_percpu);
160 *lc->lc_totalp += *localp;
161 percpu_putref(lc->lc_percpu);
162 mutex_exit(interlock);
163 }
164
165 static void
166 localcount_adjust(struct localcount *lc, int delta)
167 {
168 int64_t *localp;
169
170 localp = percpu_getref(lc->lc_percpu);
171 *localp += delta;
172 percpu_putref(lc->lc_percpu);
173 }
174
175 /*
176 * localcount_acquire(lc)
177 *
178 * Acquire a reference to lc.
179 *
180 * The reference may be held across sleeps and may be migrated
181 * from CPU to CPU, or even thread to thread -- it is only
182 * counted, not associated with a particular concrete owner.
183 *
184 * Involves no interprocessor synchronization. May be used in any
185 * context: while a lock is held, within a pserialize(9) read
186 * section, in hard interrupt context (provided other users block
187 * hard interrupts), in soft interrupt context, in thread context,
188 * &c.
189 *
190 * Caller must guarantee that there is no concurrent
191 * localcount_drain. For example, any object that may be found in
192 * a list and acquired must be removed from the list before
193 * localcount_drain.
194 */
195 void
196 localcount_acquire(struct localcount *lc)
197 {
198
199 KASSERT(lc->lc_totalp == NULL);
200 localcount_adjust(lc, +1);
201 }
202
203 /*
204 * localcount_release(lc, cv, interlock)
205 *
206 * Release a reference to lc. If there is a concurrent
207 * localcount_drain and this may be the last reference, notify
208 * localcount_drain by acquiring interlock, waking cv, and
209 * releasing interlock. The cv and interlock passed here must be
210 * the same as are passed to localcount_drain for this lc.
211 *
212 * Involves no interprocessor synchronization unless there is a
213 * concurrent localcount_drain in progress.
214 */
215 void
216 localcount_release(struct localcount *lc, kcondvar_t *cv, kmutex_t *interlock)
217 {
218
219 /*
220 * Block xcall so that if someone begins draining after we see
221 * lc->lc_totalp as null, then they won't start cv_wait until
222 * after they have counted this CPU's contributions.
223 *
224 * Otherwise, localcount_drain may notice an extant reference
225 * from this CPU and cv_wait for it, but having seen
226 * lc->lc_totalp as null, this CPU will not wake
227 * localcount_drain.
228 */
229 kpreempt_disable();
230
231 KDASSERT(mutex_ownable(interlock));
232 if (__predict_false(lc->lc_totalp != NULL)) {
233 /*
234 * Slow path -- wake localcount_drain in case this is
235 * the last reference.
236 */
237 mutex_enter(interlock);
238 localcount_adjust(lc, -1);
239 *lc->lc_totalp -= 1;
240 if (*lc->lc_totalp == 0)
241 cv_broadcast(cv);
242 mutex_exit(interlock);
243 goto out;
244 }
245
246 localcount_adjust(lc, -1);
247 out: kpreempt_enable();
248 }
249