Print this page
6149 use NULL capable segop as a shorthand for no-capabilities
Split |
Close |
Expand all |
Collapse all |
--- old/usr/src/uts/common/vm/seg_kpm.c
+++ new/usr/src/uts/common/vm/seg_kpm.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License, Version 1.0 only
6 6 * (the "License"). You may not use this file except in compliance
7 7 * with the License.
8 8 *
9 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 10 * or http://www.opensolaris.org/os/licensing.
11 11 * See the License for the specific language governing permissions
12 12 * and limitations under the License.
13 13 *
14 14 * When distributing Covered Code, include this CDDL HEADER in each
15 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 16 * If applicable, add the following below this CDDL HEADER, with the
17 17 * fields enclosed by brackets "[]" replaced with your own identifying
18 18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 19 *
20 20 * CDDL HEADER END
21 21 */
22 22 /*
23 23 * Copyright 2006 Sun Microsystems, Inc. All rights reserved.
24 24 * Use is subject to license terms.
25 25 */
26 26
27 27 /*
28 28 * Kernel Physical Mapping (kpm) segment driver (segkpm).
29 29 *
30 30 * This driver delivers along with the hat_kpm* interfaces an alternative
31 31 * mechanism for kernel mappings within the 64-bit Solaris operating system,
32 32 * which allows the mapping of all physical memory into the kernel address
33 33 * space at once. This is feasible in 64 bit kernels, e.g. for Ultrasparc II
34 34 * and beyond processors, since the available VA range is much larger than
35 35 * possible physical memory. Momentarily all physical memory is supported,
36 36 * that is represented by the list of memory segments (memsegs).
37 37 *
38 38 * Segkpm mappings have also very low overhead and large pages are used
39 39 * (when possible) to minimize the TLB and TSB footprint. It is also
40 40 * extentable for other than Sparc architectures (e.g. AMD64). Main
41 41 * advantage is the avoidance of the TLB-shootdown X-calls, which are
42 42 * normally needed when a kernel (global) mapping has to be removed.
43 43 *
44 44 * First example of a kernel facility that uses the segkpm mapping scheme
45 45 * is seg_map, where it is used as an alternative to hat_memload().
46 46 * See also hat layer for more information about the hat_kpm* routines.
47 47 * The kpm facilty can be turned off at boot time (e.g. /etc/system).
48 48 */
49 49
50 50 #include <sys/types.h>
51 51 #include <sys/param.h>
52 52 #include <sys/sysmacros.h>
53 53 #include <sys/systm.h>
54 54 #include <sys/vnode.h>
55 55 #include <sys/cmn_err.h>
56 56 #include <sys/debug.h>
57 57 #include <sys/thread.h>
58 58 #include <sys/cpuvar.h>
59 59 #include <sys/bitmap.h>
60 60 #include <sys/atomic.h>
61 61 #include <sys/lgrp.h>
62 62
63 63 #include <vm/seg_kmem.h>
64 64 #include <vm/seg_kpm.h>
65 65 #include <vm/hat.h>
66 66 #include <vm/as.h>
67 67 #include <vm/seg.h>
68 68 #include <vm/page.h>
69 69
70 70 /*
71 71 * Global kpm controls.
72 72 * See also platform and mmu specific controls.
73 73 *
74 74 * kpm_enable -- global on/off switch for segkpm.
75 75 * . Set by default on 64bit platforms that have kpm support.
76 76 * . Will be disabled from platform layer if not supported.
77 77 * . Can be disabled via /etc/system.
78 78 *
79 79 * kpm_smallpages -- use only regular/system pagesize for kpm mappings.
80 80 * . Can be useful for critical debugging of kpm clients.
81 81 * . Set to zero by default for platforms that support kpm large pages.
82 82 * The use of kpm large pages reduces the footprint of kpm meta data
83 83 * and has all the other advantages of using large pages (e.g TLB
84 84 * miss reduction).
85 85 * . Set by default for platforms that don't support kpm large pages or
86 86 * where large pages cannot be used for other reasons (e.g. there are
87 87 * only few full associative TLB entries available for large pages).
88 88 *
89 89 * segmap_kpm -- separate on/off switch for segmap using segkpm:
90 90 * . Set by default.
91 91 * . Will be disabled when kpm_enable is zero.
92 92 * . Will be disabled when MAXBSIZE != PAGESIZE.
93 93 * . Can be disabled via /etc/system.
94 94 *
95 95 */
96 96 int kpm_enable = 1;
97 97 int kpm_smallpages = 0;
↓ open down ↓ |
97 lines elided |
↑ open up ↑ |
98 98 int segmap_kpm = 1;
99 99
100 100 /*
101 101 * Private seg op routines.
102 102 */
103 103 faultcode_t segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr,
104 104 size_t len, enum fault_type type, enum seg_rw rw);
105 105 static void segkpm_dump(struct seg *);
106 106 static void segkpm_badop(void);
107 107 static int segkpm_notsup(void);
108 -static int segkpm_capable(struct seg *, segcapability_t);
109 108
110 109 #define SEGKPM_BADOP(t) (t(*)())segkpm_badop
111 110 #define SEGKPM_NOTSUP (int(*)())segkpm_notsup
112 111
113 112 static struct seg_ops segkpm_ops = {
114 113 .dup = SEGKPM_BADOP(int),
115 114 .unmap = SEGKPM_BADOP(int),
116 115 .free = SEGKPM_BADOP(void),
117 116 .fault = segkpm_fault,
118 117 .faulta = SEGKPM_BADOP(int),
119 118 .setprot = SEGKPM_BADOP(int),
120 119 .checkprot = SEGKPM_BADOP(int),
121 120 .kluster = SEGKPM_BADOP(int),
122 121 .swapout = SEGKPM_BADOP(size_t),
123 122 .sync = SEGKPM_BADOP(int),
124 123 .incore = SEGKPM_BADOP(size_t),
125 124 .lockop = SEGKPM_BADOP(int),
↓ open down ↓ |
7 lines elided |
↑ open up ↑ |
126 125 .getprot = SEGKPM_BADOP(int),
127 126 .getoffset = SEGKPM_BADOP(u_offset_t),
128 127 .gettype = SEGKPM_BADOP(int),
129 128 .getvp = SEGKPM_BADOP(int),
130 129 .advise = SEGKPM_BADOP(int),
131 130 .dump = segkpm_dump,
132 131 .pagelock = SEGKPM_NOTSUP,
133 132 .setpagesize = SEGKPM_BADOP(int),
134 133 .getmemid = SEGKPM_BADOP(int),
135 134 .getpolicy = SEGKPM_BADOP(lgrp_mem_policy_info_t *),
136 - .capable = segkpm_capable,
137 135 };
138 136
139 137 /*
140 138 * kpm_pgsz and kpm_pgshft are set by platform layer.
141 139 */
142 140 size_t kpm_pgsz; /* kpm page size */
143 141 uint_t kpm_pgshft; /* kpm page shift */
144 142 u_offset_t kpm_pgoff; /* kpm page offset mask */
145 143 uint_t kpmp2pshft; /* kpm page to page shift */
146 144 pgcnt_t kpmpnpgs; /* how many pages per kpm page */
147 145
148 146
149 147 #ifdef SEGKPM_SUPPORT
150 148
151 149 int
152 150 segkpm_create(struct seg *seg, void *argsp)
153 151 {
154 152 struct segkpm_data *skd;
155 153 struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
156 154 ushort_t *p;
157 155 int i, j;
158 156
159 157 ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
160 158 ASSERT(btokpmp(seg->s_size) >= 1 &&
161 159 kpmpageoff((uintptr_t)seg->s_base) == 0 &&
162 160 kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
163 161
164 162 skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
165 163
166 164 seg->s_data = (void *)skd;
167 165 seg->s_ops = &segkpm_ops;
168 166 skd->skd_prot = b->prot;
169 167
170 168 /*
171 169 * (1) Segkpm virtual addresses are based on physical adresses.
172 170 * From this and in opposite to other segment drivers it is
173 171 * often required to allocate a page first to be able to
174 172 * calculate the final segkpm virtual address.
175 173 * (2) Page allocation is done by calling page_create_va(),
176 174 * one important input argument is a virtual address (also
177 175 * expressed by the "va" in the function name). This function
178 176 * is highly optimized to select the right page for an optimal
179 177 * processor and platform support (e.g. virtual addressed
180 178 * caches (VAC), physical addressed caches, NUMA).
181 179 *
182 180 * Because of (1) the approach is to generate a faked virtual
183 181 * address for calling page_create_va(). In order to exploit
184 182 * the abilities of (2), especially to utilize the cache
185 183 * hierarchy (3) and to avoid VAC alias conflicts (4) the
186 184 * selection has to be done carefully. For each virtual color
187 185 * a separate counter is provided (4). The count values are
188 186 * used for the utilization of all cache lines (3) and are
189 187 * corresponding to the cache bins.
190 188 */
191 189 skd->skd_nvcolors = b->nvcolors;
192 190
193 191 p = skd->skd_va_select =
194 192 kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
195 193
196 194 for (i = 0; i < NCPU; i++)
197 195 for (j = 0; j < b->nvcolors; j++, p++)
198 196 *p = j;
199 197
200 198 return (0);
201 199 }
202 200
203 201 /*
204 202 * This routine is called via a machine specific fault handling
205 203 * routine.
206 204 */
207 205 /* ARGSUSED */
208 206 faultcode_t
209 207 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
210 208 enum fault_type type, enum seg_rw rw)
211 209 {
212 210 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
213 211
214 212 switch (type) {
215 213 case F_INVAL:
216 214 return (hat_kpm_fault(hat, addr));
217 215 case F_SOFTLOCK:
218 216 case F_SOFTUNLOCK:
219 217 return (0);
220 218 default:
221 219 return (FC_NOSUPPORT);
222 220 }
223 221 /*NOTREACHED*/
224 222 }
225 223
226 224 #define addr_to_vcolor(addr, vcolors) \
227 225 ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
228 226
229 227 /*
230 228 * Create a virtual address that can be used for invocations of
231 229 * page_create_va. Goal is to utilize the cache hierarchy (round
232 230 * robin bins) and to select the right color for virtual indexed
233 231 * caches. It isn't exact since we also increment the bin counter
234 232 * when the caller uses VOP_GETPAGE and gets a hit in the page
235 233 * cache, but we keep the bins turning for cache distribution
236 234 * (see also segkpm_create block comment).
237 235 */
238 236 caddr_t
239 237 segkpm_create_va(u_offset_t off)
240 238 {
241 239 int vcolor;
242 240 ushort_t *p;
243 241 struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
244 242 int nvcolors = skd->skd_nvcolors;
245 243 caddr_t va;
246 244
247 245 vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
248 246 p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
249 247 va = (caddr_t)ptob(*p);
250 248
251 249 atomic_add_16(p, nvcolors);
252 250
253 251 return (va);
254 252 }
255 253
256 254 /*
257 255 * Unload mapping if the instance has an active kpm mapping.
258 256 */
259 257 void
260 258 segkpm_mapout_validkpme(struct kpme *kpme)
261 259 {
262 260 caddr_t vaddr;
263 261 page_t *pp;
264 262
265 263 retry:
266 264 if ((pp = kpme->kpe_page) == NULL) {
267 265 return;
268 266 }
269 267
270 268 if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
271 269 goto retry;
272 270
273 271 /*
274 272 * Check if segkpm mapping is not unloaded in the meantime
275 273 */
276 274 if (kpme->kpe_page == NULL) {
277 275 page_unlock(pp);
278 276 return;
279 277 }
280 278
281 279 vaddr = hat_kpm_page2va(pp, 1);
282 280 hat_kpm_mapout(pp, kpme, vaddr);
283 281 page_unlock(pp);
284 282 }
285 283
286 284 static void
287 285 segkpm_badop()
288 286 {
289 287 panic("segkpm_badop");
290 288 }
291 289
292 290 #else /* SEGKPM_SUPPORT */
293 291
294 292 /* segkpm stubs */
295 293
296 294 /*ARGSUSED*/
297 295 int segkpm_create(struct seg *seg, void *argsp) { return (0); }
298 296
299 297 /* ARGSUSED */
300 298 faultcode_t
301 299 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
302 300 enum fault_type type, enum seg_rw rw)
303 301 {
304 302 return ((faultcode_t)0);
305 303 }
306 304
307 305 /* ARGSUSED */
308 306 caddr_t segkpm_create_va(u_offset_t off) { return (NULL); }
309 307
310 308 /* ARGSUSED */
311 309 void segkpm_mapout_validkpme(struct kpme *kpme) {}
312 310
313 311 static void
314 312 segkpm_badop() {}
315 313
316 314 #endif /* SEGKPM_SUPPORT */
317 315
318 316 static int
319 317 segkpm_notsup()
320 318 {
↓ open down ↓ |
174 lines elided |
↑ open up ↑ |
321 319 return (ENOTSUP);
322 320 }
323 321
324 322 /*
325 323 * segkpm pages are not dumped, so we just return
326 324 */
327 325 /*ARGSUSED*/
328 326 static void
329 327 segkpm_dump(struct seg *seg)
330 328 {}
331 -
332 -/*
333 - * We claim to have no special capabilities.
334 - */
335 -/*ARGSUSED*/
336 -static int
337 -segkpm_capable(struct seg *seg, segcapability_t capability)
338 -{
339 - return (0);
340 -}
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX