FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_domainset.c
1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2017, Jeffrey Roberson <jeff@freebsd.org>
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 */
29
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32
33 #include "opt_vm.h"
34
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/bitset.h>
38 #include <sys/domainset.h>
39 #include <sys/proc.h>
40 #include <sys/lock.h>
41 #include <sys/mutex.h>
42 #include <sys/malloc.h>
43 #include <sys/rwlock.h>
44 #include <sys/vmmeter.h>
45
46 #include <vm/vm.h>
47 #include <vm/vm_param.h>
48 #include <vm/vm_domainset.h>
49 #include <vm/vm_object.h>
50 #include <vm/vm_page.h>
51 #include <vm/vm_phys.h>
52
53 #ifdef NUMA
54 /*
55 * Iterators are written such that the first nowait pass has as short a
56 * codepath as possible to eliminate bloat from the allocator. It is
57 * assumed that most allocations are successful.
58 */
59
60 static int vm_domainset_default_stride = 64;
61
62 /*
63 * Determine which policy is to be used for this allocation.
64 */
65 static void
66 vm_domainset_iter_init(struct vm_domainset_iter *di, struct domainset *ds,
67 int *iter, struct vm_object *obj, vm_pindex_t pindex)
68 {
69
70 di->di_domain = ds;
71 di->di_iter = iter;
72 di->di_policy = ds->ds_policy;
73 if (di->di_policy == DOMAINSET_POLICY_INTERLEAVE) {
74 #if VM_NRESERVLEVEL > 0
75 if (vm_object_reserv(obj)) {
76 /*
77 * Color the pindex so we end up on the correct
78 * reservation boundary.
79 */
80 pindex += obj->pg_color;
81 pindex >>= VM_LEVEL_0_ORDER;
82 } else
83 #endif
84 pindex /= vm_domainset_default_stride;
85 /*
86 * Offset pindex so the first page of each object does
87 * not end up in domain 0.
88 */
89 if (obj != NULL)
90 pindex += (((uintptr_t)obj) / sizeof(*obj));
91 di->di_offset = pindex;
92 }
93 /* Skip domains below min on the first pass. */
94 di->di_minskip = true;
95 }
96
97 static void
98 vm_domainset_iter_rr(struct vm_domainset_iter *di, int *domain)
99 {
100
101 *domain = di->di_domain->ds_order[
102 ++(*di->di_iter) % di->di_domain->ds_cnt];
103 }
104
105 static void
106 vm_domainset_iter_prefer(struct vm_domainset_iter *di, int *domain)
107 {
108 int d;
109
110 do {
111 d = di->di_domain->ds_order[
112 ++(*di->di_iter) % di->di_domain->ds_cnt];
113 } while (d == di->di_domain->ds_prefer);
114 *domain = d;
115 }
116
117 static void
118 vm_domainset_iter_interleave(struct vm_domainset_iter *di, int *domain)
119 {
120 int d;
121
122 d = di->di_offset % di->di_domain->ds_cnt;
123 *di->di_iter = d;
124 *domain = di->di_domain->ds_order[d];
125 }
126
127 static void
128 vm_domainset_iter_next(struct vm_domainset_iter *di, int *domain)
129 {
130
131 KASSERT(di->di_n > 0,
132 ("vm_domainset_iter_first: Invalid n %d", di->di_n));
133 switch (di->di_policy) {
134 case DOMAINSET_POLICY_FIRSTTOUCH:
135 /*
136 * To prevent impossible allocations we convert an invalid
137 * first-touch to round-robin.
138 */
139 /* FALLTHROUGH */
140 case DOMAINSET_POLICY_INTERLEAVE:
141 /* FALLTHROUGH */
142 case DOMAINSET_POLICY_ROUNDROBIN:
143 vm_domainset_iter_rr(di, domain);
144 break;
145 case DOMAINSET_POLICY_PREFER:
146 vm_domainset_iter_prefer(di, domain);
147 break;
148 default:
149 panic("vm_domainset_iter_first: Unknown policy %d",
150 di->di_policy);
151 }
152 KASSERT(*domain < vm_ndomains,
153 ("vm_domainset_iter_next: Invalid domain %d", *domain));
154 }
155
156 static void
157 vm_domainset_iter_first(struct vm_domainset_iter *di, int *domain)
158 {
159
160 switch (di->di_policy) {
161 case DOMAINSET_POLICY_FIRSTTOUCH:
162 *domain = PCPU_GET(domain);
163 if (DOMAINSET_ISSET(*domain, &di->di_domain->ds_mask)) {
164 /*
165 * Add an extra iteration because we will visit the
166 * current domain a second time in the rr iterator.
167 */
168 di->di_n = di->di_domain->ds_cnt + 1;
169 break;
170 }
171 /*
172 * To prevent impossible allocations we convert an invalid
173 * first-touch to round-robin.
174 */
175 /* FALLTHROUGH */
176 case DOMAINSET_POLICY_ROUNDROBIN:
177 di->di_n = di->di_domain->ds_cnt;
178 vm_domainset_iter_rr(di, domain);
179 break;
180 case DOMAINSET_POLICY_PREFER:
181 *domain = di->di_domain->ds_prefer;
182 di->di_n = di->di_domain->ds_cnt;
183 break;
184 case DOMAINSET_POLICY_INTERLEAVE:
185 vm_domainset_iter_interleave(di, domain);
186 di->di_n = di->di_domain->ds_cnt;
187 break;
188 default:
189 panic("vm_domainset_iter_first: Unknown policy %d",
190 di->di_policy);
191 }
192 KASSERT(di->di_n > 0,
193 ("vm_domainset_iter_first: Invalid n %d", di->di_n));
194 KASSERT(*domain < vm_ndomains,
195 ("vm_domainset_iter_first: Invalid domain %d", *domain));
196 }
197
198 void
199 vm_domainset_iter_page_init(struct vm_domainset_iter *di, struct vm_object *obj,
200 vm_pindex_t pindex, int *domain, int *req)
201 {
202 struct domainset_ref *dr;
203
204 /*
205 * Object policy takes precedence over thread policy. The policies
206 * are immutable and unsynchronized. Updates can race but pointer
207 * loads are assumed to be atomic.
208 */
209 if (obj != NULL && obj->domain.dr_policy != NULL)
210 dr = &obj->domain;
211 else
212 dr = &curthread->td_domain;
213 vm_domainset_iter_init(di, dr->dr_policy, &dr->dr_iter, obj, pindex);
214 di->di_flags = *req;
215 *req = (di->di_flags & ~(VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) |
216 VM_ALLOC_NOWAIT;
217 vm_domainset_iter_first(di, domain);
218 if (vm_page_count_min_domain(*domain))
219 vm_domainset_iter_page(di, obj, domain);
220 }
221
222 int
223 vm_domainset_iter_page(struct vm_domainset_iter *di, struct vm_object *obj,
224 int *domain)
225 {
226
227 /* If there are more domains to visit we run the iterator. */
228 while (--di->di_n != 0) {
229 vm_domainset_iter_next(di, domain);
230 if (!di->di_minskip || !vm_page_count_min_domain(*domain))
231 return (0);
232 }
233
234 /* If we skipped domains below min restart the search. */
235 if (di->di_minskip) {
236 di->di_minskip = false;
237 vm_domainset_iter_first(di, domain);
238 return (0);
239 }
240
241 /* If we visited all domains and this was a NOWAIT we return error. */
242 if ((di->di_flags & (VM_ALLOC_WAITOK | VM_ALLOC_WAITFAIL)) == 0)
243 return (ENOMEM);
244
245 /* Wait for one of the domains to accumulate some free pages. */
246 if (obj != NULL)
247 VM_OBJECT_WUNLOCK(obj);
248 vm_wait_doms(&di->di_domain->ds_mask, 0);
249 if (obj != NULL)
250 VM_OBJECT_WLOCK(obj);
251 if ((di->di_flags & VM_ALLOC_WAITFAIL) != 0)
252 return (ENOMEM);
253
254 /* Restart the search. */
255 vm_domainset_iter_first(di, domain);
256
257 return (0);
258 }
259
260 static void
261 _vm_domainset_iter_policy_init(struct vm_domainset_iter *di, int *domain,
262 int *flags)
263 {
264
265 di->di_flags = *flags;
266 *flags = (di->di_flags & ~M_WAITOK) | M_NOWAIT;
267 vm_domainset_iter_first(di, domain);
268 if (vm_page_count_min_domain(*domain))
269 vm_domainset_iter_policy(di, domain);
270 }
271
272 void
273 vm_domainset_iter_policy_init(struct vm_domainset_iter *di,
274 struct domainset *ds, int *domain, int *flags)
275 {
276
277 vm_domainset_iter_init(di, ds, &curthread->td_domain.dr_iter, NULL, 0);
278 _vm_domainset_iter_policy_init(di, domain, flags);
279 }
280
281 void
282 vm_domainset_iter_policy_ref_init(struct vm_domainset_iter *di,
283 struct domainset_ref *dr, int *domain, int *flags)
284 {
285
286 vm_domainset_iter_init(di, dr->dr_policy, &dr->dr_iter, NULL, 0);
287 _vm_domainset_iter_policy_init(di, domain, flags);
288 }
289
290 int
291 vm_domainset_iter_policy(struct vm_domainset_iter *di, int *domain)
292 {
293
294 /* If there are more domains to visit we run the iterator. */
295 while (--di->di_n != 0) {
296 vm_domainset_iter_next(di, domain);
297 if (!di->di_minskip || !vm_page_count_min_domain(*domain))
298 return (0);
299 }
300
301 /* If we skipped domains below min restart the search. */
302 if (di->di_minskip) {
303 di->di_minskip = false;
304 vm_domainset_iter_first(di, domain);
305 return (0);
306 }
307
308 /* If we visited all domains and this was a NOWAIT we return error. */
309 if ((di->di_flags & M_WAITOK) == 0)
310 return (ENOMEM);
311
312 /* Wait for one of the domains to accumulate some free pages. */
313 vm_wait_doms(&di->di_domain->ds_mask, 0);
314
315 /* Restart the search. */
316 vm_domainset_iter_first(di, domain);
317
318 return (0);
319 }
320
321 #else /* !NUMA */
322
323 int
324 vm_domainset_iter_page(struct vm_domainset_iter *di, struct vm_object *obj,
325 int *domain)
326 {
327
328 return (EJUSTRETURN);
329 }
330
331 void
332 vm_domainset_iter_page_init(struct vm_domainset_iter *di, struct vm_object *obj,
333 vm_pindex_t pindex, int *domain, int *flags)
334 {
335
336 *domain = 0;
337 }
338
339 int
340 vm_domainset_iter_policy(struct vm_domainset_iter *di, int *domain)
341 {
342
343 return (EJUSTRETURN);
344 }
345
346 void
347 vm_domainset_iter_policy_init(struct vm_domainset_iter *di,
348 struct domainset *ds, int *domain, int *flags)
349 {
350
351 *domain = 0;
352 }
353
354 void
355 vm_domainset_iter_policy_ref_init(struct vm_domainset_iter *di,
356 struct domainset_ref *dr, int *domain, int *flags)
357 {
358
359 *domain = 0;
360 }
361
362 #endif /* NUMA */
Cache object: be21facc371eebf04724452e2a1d9b46
|