Print this page
use NULL dump segop as a shorthand for no-op
Instead of forcing every segment driver to implement a dummy function that
does nothing, handle NULL dump segop function pointer as a no-op 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 *
↓ open down ↓ |
94 lines elided |
↑ open up ↑ |
95 95 */
96 96 int kpm_enable = 1;
97 97 int kpm_smallpages = 0;
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 -static void segkpm_dump(struct seg *);
106 105 static int segkpm_pagelock(struct seg *seg, caddr_t addr, size_t len,
107 106 struct page ***page, enum lock_type type,
108 107 enum seg_rw rw);
109 108
110 109 static const struct seg_ops segkpm_ops = {
111 110 .fault = segkpm_fault,
112 - .dump = segkpm_dump,
113 111 .pagelock = segkpm_pagelock,
114 112 //#ifndef SEGKPM_SUPPORT
115 113 #if 0
116 114 #error FIXME: define nop
117 115 .dup = nop,
118 116 .unmap = nop,
119 117 .free = nop,
120 118 .faulta = nop,
121 119 .setprot = nop,
122 120 .checkprot = nop,
123 121 .kluster = nop,
124 122 .sync = nop,
125 123 .incore = nop,
126 124 .lockop = nop,
127 125 .getprot = nop,
128 126 .getoffset = nop,
129 127 .gettype = nop,
130 128 .getvp = nop,
131 129 .advise = nop,
132 130 .getpolicy = nop,
133 131 #endif
134 132 };
135 133
136 134 /*
137 135 * kpm_pgsz and kpm_pgshft are set by platform layer.
138 136 */
139 137 size_t kpm_pgsz; /* kpm page size */
140 138 uint_t kpm_pgshft; /* kpm page shift */
141 139 u_offset_t kpm_pgoff; /* kpm page offset mask */
142 140 uint_t kpmp2pshft; /* kpm page to page shift */
143 141 pgcnt_t kpmpnpgs; /* how many pages per kpm page */
144 142
145 143
146 144 #ifdef SEGKPM_SUPPORT
147 145
148 146 int
149 147 segkpm_create(struct seg *seg, void *argsp)
150 148 {
151 149 struct segkpm_data *skd;
152 150 struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
153 151 ushort_t *p;
154 152 int i, j;
155 153
156 154 ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
157 155 ASSERT(btokpmp(seg->s_size) >= 1 &&
158 156 kpmpageoff((uintptr_t)seg->s_base) == 0 &&
159 157 kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
160 158
161 159 skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
162 160
163 161 seg->s_data = (void *)skd;
164 162 seg->s_ops = &segkpm_ops;
165 163 skd->skd_prot = b->prot;
166 164
167 165 /*
168 166 * (1) Segkpm virtual addresses are based on physical adresses.
169 167 * From this and in opposite to other segment drivers it is
170 168 * often required to allocate a page first to be able to
171 169 * calculate the final segkpm virtual address.
172 170 * (2) Page allocation is done by calling page_create_va(),
173 171 * one important input argument is a virtual address (also
174 172 * expressed by the "va" in the function name). This function
175 173 * is highly optimized to select the right page for an optimal
176 174 * processor and platform support (e.g. virtual addressed
177 175 * caches (VAC), physical addressed caches, NUMA).
178 176 *
179 177 * Because of (1) the approach is to generate a faked virtual
180 178 * address for calling page_create_va(). In order to exploit
181 179 * the abilities of (2), especially to utilize the cache
182 180 * hierarchy (3) and to avoid VAC alias conflicts (4) the
183 181 * selection has to be done carefully. For each virtual color
184 182 * a separate counter is provided (4). The count values are
185 183 * used for the utilization of all cache lines (3) and are
186 184 * corresponding to the cache bins.
187 185 */
188 186 skd->skd_nvcolors = b->nvcolors;
189 187
190 188 p = skd->skd_va_select =
191 189 kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
192 190
193 191 for (i = 0; i < NCPU; i++)
194 192 for (j = 0; j < b->nvcolors; j++, p++)
195 193 *p = j;
196 194
197 195 return (0);
198 196 }
199 197
200 198 /*
201 199 * This routine is called via a machine specific fault handling
202 200 * routine.
203 201 */
204 202 /* ARGSUSED */
205 203 faultcode_t
206 204 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
207 205 enum fault_type type, enum seg_rw rw)
208 206 {
209 207 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
210 208
211 209 switch (type) {
212 210 case F_INVAL:
213 211 return (hat_kpm_fault(hat, addr));
214 212 case F_SOFTLOCK:
215 213 case F_SOFTUNLOCK:
216 214 return (0);
217 215 default:
218 216 return (FC_NOSUPPORT);
219 217 }
220 218 /*NOTREACHED*/
221 219 }
222 220
223 221 #define addr_to_vcolor(addr, vcolors) \
224 222 ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
225 223
226 224 /*
227 225 * Create a virtual address that can be used for invocations of
228 226 * page_create_va. Goal is to utilize the cache hierarchy (round
229 227 * robin bins) and to select the right color for virtual indexed
230 228 * caches. It isn't exact since we also increment the bin counter
231 229 * when the caller uses VOP_GETPAGE and gets a hit in the page
232 230 * cache, but we keep the bins turning for cache distribution
233 231 * (see also segkpm_create block comment).
234 232 */
235 233 caddr_t
236 234 segkpm_create_va(u_offset_t off)
237 235 {
238 236 int vcolor;
239 237 ushort_t *p;
240 238 struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
241 239 int nvcolors = skd->skd_nvcolors;
242 240 caddr_t va;
243 241
244 242 vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
245 243 p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
246 244 va = (caddr_t)ptob(*p);
247 245
248 246 atomic_add_16(p, nvcolors);
249 247
250 248 return (va);
251 249 }
252 250
253 251 /*
254 252 * Unload mapping if the instance has an active kpm mapping.
255 253 */
256 254 void
257 255 segkpm_mapout_validkpme(struct kpme *kpme)
258 256 {
259 257 caddr_t vaddr;
260 258 page_t *pp;
261 259
262 260 retry:
263 261 if ((pp = kpme->kpe_page) == NULL) {
264 262 return;
265 263 }
266 264
267 265 if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
268 266 goto retry;
269 267
270 268 /*
271 269 * Check if segkpm mapping is not unloaded in the meantime
272 270 */
273 271 if (kpme->kpe_page == NULL) {
274 272 page_unlock(pp);
275 273 return;
276 274 }
277 275
278 276 vaddr = hat_kpm_page2va(pp, 1);
279 277 hat_kpm_mapout(pp, kpme, vaddr);
280 278 page_unlock(pp);
281 279 }
282 280
283 281 #else /* SEGKPM_SUPPORT */
284 282
285 283 /* segkpm stubs */
286 284
287 285 /*ARGSUSED*/
288 286 int segkpm_create(struct seg *seg, void *argsp)
289 287 {
290 288 return (0);
291 289 }
292 290
293 291 /* ARGSUSED */
294 292 faultcode_t
295 293 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
296 294 enum fault_type type, enum seg_rw rw)
297 295 {
298 296 return (0);
299 297 }
300 298
301 299 /* ARGSUSED */
302 300 caddr_t segkpm_create_va(u_offset_t off)
303 301 {
304 302 return (NULL);
305 303 }
306 304
307 305 /* ARGSUSED */
308 306 void segkpm_mapout_validkpme(struct kpme *kpme)
309 307 {
↓ open down ↓ |
187 lines elided |
↑ open up ↑ |
310 308 }
311 309
312 310 #endif /* SEGKPM_SUPPORT */
313 311
314 312 /* ARGSUSED */
315 313 static int
316 314 segkpm_pagelock(struct seg *seg, caddr_t addr, size_t len,
317 315 struct page ***page, enum lock_type type, enum seg_rw rw)
318 316 {
319 317 return (ENOTSUP);
320 -}
321 -
322 -/*
323 - * segkpm pages are not dumped, so we just return
324 - */
325 -/*ARGSUSED*/
326 -static void
327 -segkpm_dump(struct seg *seg)
328 -{
329 318 }
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX