1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2006 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26 /* Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T */
27 /* All Rights Reserved */
28
29 /*
30 * Portions of this source code were derived from Berkeley 4.3 BSD
31 * under license from the Regents of the University of California.
32 */
33
34 /*
35 * VM - segment for non-faulting loads.
36 */
37
38 #include <sys/types.h>
39 #include <sys/t_lock.h>
40 #include <sys/param.h>
41 #include <sys/mman.h>
42 #include <sys/errno.h>
43 #include <sys/kmem.h>
44 #include <sys/cmn_err.h>
45 #include <sys/vnode.h>
46 #include <sys/proc.h>
47 #include <sys/conf.h>
48 #include <sys/debug.h>
49 #include <sys/archsystm.h>
50 #include <sys/lgrp.h>
51
52 #include <vm/page.h>
53 #include <vm/hat.h>
54 #include <vm/as.h>
55 #include <vm/seg.h>
56 #include <vm/vpage.h>
57
58 /*
59 * Private seg op routines.
60 */
61 static int segnf_dup(struct seg *seg, struct seg *newseg);
62 static int segnf_unmap(struct seg *seg, caddr_t addr, size_t len);
63 static void segnf_free(struct seg *seg);
64 static faultcode_t segnf_nomap(void);
65 static int segnf_setprot(struct seg *seg, caddr_t addr,
66 size_t len, uint_t prot);
67 static int segnf_checkprot(struct seg *seg, caddr_t addr,
68 size_t len, uint_t prot);
69 static int segnf_nop(void);
70 static int segnf_getprot(struct seg *seg, caddr_t addr,
71 size_t len, uint_t *protv);
72 static u_offset_t segnf_getoffset(struct seg *seg, caddr_t addr);
73 static int segnf_gettype(struct seg *seg, caddr_t addr);
74 static int segnf_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp);
75 static void segnf_dump(struct seg *seg);
76 static int segnf_pagelock(struct seg *seg, caddr_t addr, size_t len,
77 struct page ***ppp, enum lock_type type, enum seg_rw rw);
78 static int segnf_setpagesize(struct seg *seg, caddr_t addr, size_t len,
79 uint_t szc);
80 static int segnf_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp);
81 static lgrp_mem_policy_info_t *segnf_getpolicy(struct seg *seg,
82 caddr_t addr);
83
84
85 struct seg_ops segnf_ops = {
86 .dup = segnf_dup,
87 .unmap = segnf_unmap,
88 .free = segnf_free,
89 .fault = (faultcode_t (*)(struct hat *, struct seg *, caddr_t,
90 size_t, enum fault_type, enum seg_rw))segnf_nomap,
91 .faulta = (faultcode_t (*)(struct seg *, caddr_t)) segnf_nomap,
92 .setprot = segnf_setprot,
93 .checkprot = segnf_checkprot,
94 .sync = (int (*)(struct seg *, caddr_t, size_t, int, uint_t))
95 segnf_nop,
96 .incore = (size_t (*)(struct seg *, caddr_t, size_t, char *))
97 segnf_nop,
98 .lockop = (int (*)(struct seg *, caddr_t, size_t, int, int,
99 ulong_t *, size_t))segnf_nop,
100 .getprot = segnf_getprot,
101 .getoffset = segnf_getoffset,
102 .gettype = segnf_gettype,
103 .getvp = segnf_getvp,
104 .advise = (int (*)(struct seg *, caddr_t, size_t, uint_t))
105 segnf_nop,
106 .dump = segnf_dump,
107 .pagelock = segnf_pagelock,
108 .setpagesize = segnf_setpagesize,
109 .getmemid = segnf_getmemid,
110 .getpolicy = segnf_getpolicy,
111 };
112
113 /*
114 * vnode and page for the page of zeros we use for the nf mappings.
115 */
116 static kmutex_t segnf_lock;
117 static struct vnode nfvp;
118 static struct page **nfpp;
119
120 #define addr_to_vcolor(addr) \
121 (shm_alignment) ? \
122 ((int)(((uintptr_t)(addr) & (shm_alignment - 1)) >> PAGESHIFT)) : 0
123
124 /*
125 * We try to limit the number of Non-fault segments created.
126 * Non fault segments are created to optimize sparc V9 code which uses
127 * the sparc nonfaulting load ASI (ASI_PRIMARY_NOFAULT).
128 *
129 * There are several reasons why creating too many non-fault segments
130 * could cause problems.
131 *
132 * First, excessive allocation of kernel resources for the seg
133 * structures and the HAT data to map the zero pages.
134 *
135 * Secondly, creating nofault segments actually uses up user virtual
136 * address space. This makes it unavailable for subsequent mmap(0, ...)
137 * calls which use as_gap() to find empty va regions. Creation of too
138 * many nofault segments could thus interfere with the ability of the
139 * runtime linker to load a shared object.
140 */
141 #define MAXSEGFORNF (10000)
142 #define MAXNFSEARCH (5)
143
144
145 /*
146 * Must be called from startup()
147 */
148 void
149 segnf_init()
150 {
151 mutex_init(&segnf_lock, NULL, MUTEX_DEFAULT, NULL);
152 }
153
154
155 /*
156 * Create a no-fault segment.
157 *
158 * The no-fault segment is not technically necessary, as the code in
159 * nfload() in trap.c will emulate the SPARC instruction and load
160 * a value of zero in the destination register.
161 *
162 * However, this code tries to put a page of zero's at the nofault address
163 * so that subsequent non-faulting loads to the same page will not
164 * trap with a tlb miss.
165 *
166 * In order to help limit the number of segments we merge adjacent nofault
167 * segments into a single segment. If we get a large number of segments
168 * we'll also try to delete a random other nf segment.
169 */
170 /* ARGSUSED */
171 int
172 segnf_create(struct seg *seg, void *argsp)
173 {
174 uint_t prot;
175 pgcnt_t vacpgs;
176 u_offset_t off = 0;
177 caddr_t vaddr = NULL;
178 int i, color;
179 struct seg *s1;
180 struct seg *s2;
181 size_t size;
182 struct as *as = seg->s_as;
183
184 ASSERT(as && AS_WRITE_HELD(as, &as->a_lock));
185
186 /*
187 * Need a page per virtual color or just 1 if no vac.
188 */
189 mutex_enter(&segnf_lock);
190 if (nfpp == NULL) {
191 struct seg kseg;
192
193 vacpgs = 1;
194 if (shm_alignment > PAGESIZE) {
195 vacpgs = shm_alignment >> PAGESHIFT;
196 }
197
198 nfpp = kmem_alloc(sizeof (*nfpp) * vacpgs, KM_SLEEP);
199
200 kseg.s_as = &kas;
201 for (i = 0; i < vacpgs; i++, off += PAGESIZE,
202 vaddr += PAGESIZE) {
203 nfpp[i] = page_create_va(&nfvp, off, PAGESIZE,
204 PG_WAIT | PG_NORELOC, &kseg, vaddr);
205 page_io_unlock(nfpp[i]);
206 page_downgrade(nfpp[i]);
207 pagezero(nfpp[i], 0, PAGESIZE);
208 }
209 }
210 mutex_exit(&segnf_lock);
211
212 hat_map(as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
213
214 /*
215 * s_data can't be NULL because of ASSERTS in the common vm code.
216 */
217 seg->s_ops = &segnf_ops;
218 seg->s_data = seg;
219 seg->s_flags |= S_PURGE;
220
221 mutex_enter(&as->a_contents);
222 as->a_flags |= AS_NEEDSPURGE;
223 mutex_exit(&as->a_contents);
224
225 prot = PROT_READ;
226 color = addr_to_vcolor(seg->s_base);
227 if (as != &kas)
228 prot |= PROT_USER;
229 hat_memload(as->a_hat, seg->s_base, nfpp[color],
230 prot | HAT_NOFAULT, HAT_LOAD);
231
232 /*
233 * At this point see if we can concatenate a segment to
234 * a non-fault segment immediately before and/or after it.
235 */
236 if ((s1 = AS_SEGPREV(as, seg)) != NULL &&
237 s1->s_ops == &segnf_ops &&
238 s1->s_base + s1->s_size == seg->s_base) {
239 size = s1->s_size;
240 seg_free(s1);
241 seg->s_base -= size;
242 seg->s_size += size;
243 }
244
245 if ((s2 = AS_SEGNEXT(as, seg)) != NULL &&
246 s2->s_ops == &segnf_ops &&
247 seg->s_base + seg->s_size == s2->s_base) {
248 size = s2->s_size;
249 seg_free(s2);
250 seg->s_size += size;
251 }
252
253 /*
254 * if we already have a lot of segments, try to delete some other
255 * nofault segment to reduce the probability of uncontrolled segment
256 * creation.
257 *
258 * the code looks around quickly (no more than MAXNFSEARCH segments
259 * each way) for another NF segment and then deletes it.
260 */
261 if (avl_numnodes(&as->a_segtree) > MAXSEGFORNF) {
262 size = 0;
263 s2 = NULL;
264 s1 = AS_SEGPREV(as, seg);
265 while (size++ < MAXNFSEARCH && s1 != NULL) {
266 if (s1->s_ops == &segnf_ops)
267 s2 = s1;
268 s1 = AS_SEGPREV(s1->s_as, seg);
269 }
270 if (s2 == NULL) {
271 s1 = AS_SEGNEXT(as, seg);
272 while (size-- > 0 && s1 != NULL) {
273 if (s1->s_ops == &segnf_ops)
274 s2 = s1;
275 s1 = AS_SEGNEXT(as, seg);
276 }
277 }
278 if (s2 != NULL)
279 seg_unmap(s2);
280 }
281
282 return (0);
283 }
284
285 /*
286 * Never really need "No fault" segments, so they aren't dup'd.
287 */
288 /* ARGSUSED */
289 static int
290 segnf_dup(struct seg *seg, struct seg *newseg)
291 {
292 panic("segnf_dup");
293 return (0);
294 }
295
296 /*
297 * Split a segment at addr for length len.
298 */
299 static int
300 segnf_unmap(struct seg *seg, caddr_t addr, size_t len)
301 {
302 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
303
304 /*
305 * Check for bad sizes.
306 */
307 if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
308 (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
309 cmn_err(CE_PANIC, "segnf_unmap: bad unmap size");
310 }
311
312 /*
313 * Unload any hardware translations in the range to be taken out.
314 */
315 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD_UNMAP);
316
317 if (addr == seg->s_base && len == seg->s_size) {
318 /*
319 * Freeing entire segment.
320 */
321 seg_free(seg);
322 } else if (addr == seg->s_base) {
323 /*
324 * Freeing the beginning of the segment.
325 */
326 seg->s_base += len;
327 seg->s_size -= len;
328 } else if (addr + len == seg->s_base + seg->s_size) {
329 /*
330 * Freeing the end of the segment.
331 */
332 seg->s_size -= len;
333 } else {
334 /*
335 * The section to go is in the middle of the segment, so we
336 * have to cut it into two segments. We shrink the existing
337 * "seg" at the low end, and create "nseg" for the high end.
338 */
339 caddr_t nbase = addr + len;
340 size_t nsize = (seg->s_base + seg->s_size) - nbase;
341 struct seg *nseg;
342
343 /*
344 * Trim down "seg" before trying to stick "nseg" into the as.
345 */
346 seg->s_size = addr - seg->s_base;
347 nseg = seg_alloc(seg->s_as, nbase, nsize);
348 if (nseg == NULL)
349 cmn_err(CE_PANIC, "segnf_unmap: seg_alloc failed");
350
351 /*
352 * s_data can't be NULL because of ASSERTs in common VM code.
353 */
354 nseg->s_ops = seg->s_ops;
355 nseg->s_data = nseg;
356 nseg->s_flags |= S_PURGE;
357 mutex_enter(&seg->s_as->a_contents);
358 seg->s_as->a_flags |= AS_NEEDSPURGE;
359 mutex_exit(&seg->s_as->a_contents);
360 }
361
362 return (0);
363 }
364
365 /*
366 * Free a segment.
367 */
368 static void
369 segnf_free(struct seg *seg)
370 {
371 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
372 }
373
374 /*
375 * No faults allowed on segnf.
376 */
377 static faultcode_t
378 segnf_nomap(void)
379 {
380 return (FC_NOMAP);
381 }
382
383 /* ARGSUSED */
384 static int
385 segnf_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
386 {
387 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
388 return (EACCES);
389 }
390
391 /* ARGSUSED */
392 static int
393 segnf_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
394 {
395 uint_t sprot;
396 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
397
398 sprot = seg->s_as == &kas ? PROT_READ : PROT_READ|PROT_USER;
399 return ((prot & sprot) == prot ? 0 : EACCES);
400 }
401
402 static int
403 segnf_nop(void)
404 {
405 return (0);
406 }
407
408 static int
409 segnf_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
410 {
411 size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
412 size_t p;
413 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
414
415 for (p = 0; p < pgno; ++p)
416 protv[p] = PROT_READ;
417 return (0);
418 }
419
420 /* ARGSUSED */
421 static u_offset_t
422 segnf_getoffset(struct seg *seg, caddr_t addr)
423 {
424 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
425
426 return ((u_offset_t)0);
427 }
428
429 /* ARGSUSED */
430 static int
431 segnf_gettype(struct seg *seg, caddr_t addr)
432 {
433 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
434
435 return (MAP_SHARED);
436 }
437
438 /* ARGSUSED */
439 static int
440 segnf_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
441 {
442 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
443
444 *vpp = &nfvp;
445 return (0);
446 }
447
448 /*
449 * segnf pages are not dumped, so we just return
450 */
451 /* ARGSUSED */
452 static void
453 segnf_dump(struct seg *seg)
454 {}
455
456 /*ARGSUSED*/
457 static int
458 segnf_pagelock(struct seg *seg, caddr_t addr, size_t len,
459 struct page ***ppp, enum lock_type type, enum seg_rw rw)
460 {
461 return (ENOTSUP);
462 }
463
464 /*ARGSUSED*/
465 static int
466 segnf_setpagesize(struct seg *seg, caddr_t addr, size_t len,
467 uint_t szc)
468 {
469 return (ENOTSUP);
470 }
471
472 /*ARGSUSED*/
473 static int
474 segnf_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
475 {
476 return (ENODEV);
477 }
478
479 /*ARGSUSED*/
480 static lgrp_mem_policy_info_t *
481 segnf_getpolicy(struct seg *seg, caddr_t addr)
482 {
483 return (NULL);
484 }