Print this page
use NULL capable segop as a shorthand for no-capabilities
Instead of forcing every segment driver to implement a dummy "return 0"
function, handle NULL capable segop function pointer as "no copabilities
supported" shorthand.
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;
98 98 int segmap_kpm = 1;
↓ open down ↓ |
98 lines elided |
↑ open up ↑ |
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 int segkpm_pagelock(struct seg *seg, caddr_t addr, size_t len,
107 107 struct page ***page, enum lock_type type,
108 108 enum seg_rw rw);
109 -static int segkpm_capable(struct seg *, segcapability_t);
110 109
111 110 static struct seg_ops segkpm_ops = {
112 111 .fault = segkpm_fault,
113 112 .dump = segkpm_dump,
114 113 .pagelock = segkpm_pagelock,
115 - .capable = segkpm_capable,
116 114 //#ifndef SEGKPM_SUPPORT
117 115 #if 0
118 116 #error FIXME: define nop
119 117 .dup = nop,
120 118 .unmap = nop,
121 119 .free = nop,
122 120 .faulta = nop,
123 121 .setprot = nop,
124 122 .checkprot = nop,
125 123 .kluster = nop,
126 124 .sync = nop,
127 125 .incore = nop,
128 126 .lockop = nop,
129 127 .getprot = nop,
130 128 .getoffset = nop,
131 129 .gettype = nop,
132 130 .getvp = nop,
133 131 .advise = nop,
134 132 .setpagesize = nop,
135 133 .getmemid = nop,
136 134 .getpolicy = nop,
137 135 #endif
138 136 };
139 137
140 138 /*
141 139 * kpm_pgsz and kpm_pgshft are set by platform layer.
142 140 */
143 141 size_t kpm_pgsz; /* kpm page size */
144 142 uint_t kpm_pgshft; /* kpm page shift */
145 143 u_offset_t kpm_pgoff; /* kpm page offset mask */
146 144 uint_t kpmp2pshft; /* kpm page to page shift */
147 145 pgcnt_t kpmpnpgs; /* how many pages per kpm page */
148 146
149 147
150 148 #ifdef SEGKPM_SUPPORT
151 149
152 150 int
153 151 segkpm_create(struct seg *seg, void *argsp)
154 152 {
155 153 struct segkpm_data *skd;
156 154 struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
157 155 ushort_t *p;
158 156 int i, j;
159 157
160 158 ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
161 159 ASSERT(btokpmp(seg->s_size) >= 1 &&
162 160 kpmpageoff((uintptr_t)seg->s_base) == 0 &&
163 161 kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
164 162
165 163 skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
166 164
167 165 seg->s_data = (void *)skd;
168 166 seg->s_ops = &segkpm_ops;
169 167 skd->skd_prot = b->prot;
170 168
171 169 /*
172 170 * (1) Segkpm virtual addresses are based on physical adresses.
173 171 * From this and in opposite to other segment drivers it is
174 172 * often required to allocate a page first to be able to
175 173 * calculate the final segkpm virtual address.
176 174 * (2) Page allocation is done by calling page_create_va(),
177 175 * one important input argument is a virtual address (also
178 176 * expressed by the "va" in the function name). This function
179 177 * is highly optimized to select the right page for an optimal
180 178 * processor and platform support (e.g. virtual addressed
181 179 * caches (VAC), physical addressed caches, NUMA).
182 180 *
183 181 * Because of (1) the approach is to generate a faked virtual
184 182 * address for calling page_create_va(). In order to exploit
185 183 * the abilities of (2), especially to utilize the cache
186 184 * hierarchy (3) and to avoid VAC alias conflicts (4) the
187 185 * selection has to be done carefully. For each virtual color
188 186 * a separate counter is provided (4). The count values are
189 187 * used for the utilization of all cache lines (3) and are
190 188 * corresponding to the cache bins.
191 189 */
192 190 skd->skd_nvcolors = b->nvcolors;
193 191
194 192 p = skd->skd_va_select =
195 193 kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
196 194
197 195 for (i = 0; i < NCPU; i++)
198 196 for (j = 0; j < b->nvcolors; j++, p++)
199 197 *p = j;
200 198
201 199 return (0);
202 200 }
203 201
204 202 /*
205 203 * This routine is called via a machine specific fault handling
206 204 * routine.
207 205 */
208 206 /* ARGSUSED */
209 207 faultcode_t
210 208 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
211 209 enum fault_type type, enum seg_rw rw)
212 210 {
213 211 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
214 212
215 213 switch (type) {
216 214 case F_INVAL:
217 215 return (hat_kpm_fault(hat, addr));
218 216 case F_SOFTLOCK:
219 217 case F_SOFTUNLOCK:
220 218 return (0);
221 219 default:
222 220 return (FC_NOSUPPORT);
223 221 }
224 222 /*NOTREACHED*/
225 223 }
226 224
227 225 #define addr_to_vcolor(addr, vcolors) \
228 226 ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
229 227
230 228 /*
231 229 * Create a virtual address that can be used for invocations of
232 230 * page_create_va. Goal is to utilize the cache hierarchy (round
233 231 * robin bins) and to select the right color for virtual indexed
234 232 * caches. It isn't exact since we also increment the bin counter
235 233 * when the caller uses VOP_GETPAGE and gets a hit in the page
236 234 * cache, but we keep the bins turning for cache distribution
237 235 * (see also segkpm_create block comment).
238 236 */
239 237 caddr_t
240 238 segkpm_create_va(u_offset_t off)
241 239 {
242 240 int vcolor;
243 241 ushort_t *p;
244 242 struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
245 243 int nvcolors = skd->skd_nvcolors;
246 244 caddr_t va;
247 245
248 246 vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
249 247 p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
250 248 va = (caddr_t)ptob(*p);
251 249
252 250 atomic_add_16(p, nvcolors);
253 251
254 252 return (va);
255 253 }
256 254
257 255 /*
258 256 * Unload mapping if the instance has an active kpm mapping.
259 257 */
260 258 void
261 259 segkpm_mapout_validkpme(struct kpme *kpme)
262 260 {
263 261 caddr_t vaddr;
264 262 page_t *pp;
265 263
266 264 retry:
267 265 if ((pp = kpme->kpe_page) == NULL) {
268 266 return;
269 267 }
270 268
271 269 if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
272 270 goto retry;
273 271
274 272 /*
275 273 * Check if segkpm mapping is not unloaded in the meantime
276 274 */
277 275 if (kpme->kpe_page == NULL) {
278 276 page_unlock(pp);
279 277 return;
280 278 }
281 279
282 280 vaddr = hat_kpm_page2va(pp, 1);
283 281 hat_kpm_mapout(pp, kpme, vaddr);
284 282 page_unlock(pp);
285 283 }
286 284
287 285 #else /* SEGKPM_SUPPORT */
288 286
289 287 /* segkpm stubs */
290 288
291 289 /*ARGSUSED*/
292 290 int segkpm_create(struct seg *seg, void *argsp)
293 291 {
294 292 return (0);
295 293 }
296 294
297 295 /* ARGSUSED */
298 296 faultcode_t
299 297 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
300 298 enum fault_type type, enum seg_rw rw)
301 299 {
302 300 return (0);
303 301 }
304 302
305 303 /* ARGSUSED */
306 304 caddr_t segkpm_create_va(u_offset_t off)
307 305 {
308 306 return (NULL);
309 307 }
310 308
311 309 /* ARGSUSED */
312 310 void segkpm_mapout_validkpme(struct kpme *kpme)
313 311 {
314 312 }
315 313
316 314 #endif /* SEGKPM_SUPPORT */
317 315
318 316 /* ARGSUSED */
319 317 static int
320 318 segkpm_pagelock(struct seg *seg, caddr_t addr, size_t len,
321 319 struct page ***page, enum lock_type type, enum seg_rw rw)
322 320 {
↓ open down ↓ |
197 lines elided |
↑ open up ↑ |
323 321 return (ENOTSUP);
324 322 }
325 323
326 324 /*
327 325 * segkpm pages are not dumped, so we just return
328 326 */
329 327 /*ARGSUSED*/
330 328 static void
331 329 segkpm_dump(struct seg *seg)
332 330 {
333 -}
334 -
335 -/*
336 - * We claim to have no special capabilities.
337 - */
338 -/*ARGSUSED*/
339 -static int
340 -segkpm_capable(struct seg *seg, segcapability_t capability)
341 -{
342 - return (0);
343 331 }
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX