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 void     segnf_badop(void);
  70 static int      segnf_nop(void);
  71 static int      segnf_getprot(struct seg *seg, caddr_t addr,
  72                     size_t len, uint_t *protv);
  73 static u_offset_t segnf_getoffset(struct seg *seg, caddr_t addr);
  74 static int      segnf_gettype(struct seg *seg, caddr_t addr);
  75 static int      segnf_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp);
  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 
  79 
  80 struct seg_ops segnf_ops = {
  81         .dup            = segnf_dup,
  82         .unmap          = segnf_unmap,
  83         .free           = segnf_free,
  84         .fault          = (faultcode_t (*)(struct hat *, struct seg *, caddr_t,
  85             size_t, enum fault_type, enum seg_rw))segnf_nomap,
  86         .faulta         = (faultcode_t (*)(struct seg *, caddr_t)) segnf_nomap,
  87         .setprot        = segnf_setprot,
  88         .checkprot      = segnf_checkprot,
  89         .kluster        = (int (*)())segnf_badop,
  90         .sync           = (int (*)(struct seg *, caddr_t, size_t, int, uint_t))
  91                 segnf_nop,
  92         .incore         = (size_t (*)(struct seg *, caddr_t, size_t, char *))
  93                 segnf_nop,
  94         .lockop         = (int (*)(struct seg *, caddr_t, size_t, int, int,
  95             ulong_t *, size_t))segnf_nop,
  96         .getprot        = segnf_getprot,
  97         .getoffset      = segnf_getoffset,
  98         .gettype        = segnf_gettype,
  99         .getvp          = segnf_getvp,
 100         .advise         = (int (*)(struct seg *, caddr_t, size_t, uint_t))
 101                 segnf_nop,
 102         .pagelock       = segnf_pagelock,
 103 };
 104 
 105 /*
 106  * vnode and page for the page of zeros we use for the nf mappings.
 107  */
 108 static kmutex_t segnf_lock;
 109 static struct vnode nfvp;
 110 static struct page **nfpp;
 111 
 112 #define addr_to_vcolor(addr)                                            \
 113         (shm_alignment) ?                                               \
 114         ((int)(((uintptr_t)(addr) & (shm_alignment - 1)) >> PAGESHIFT)) : 0
 115 
 116 /*
 117  * We try to limit the number of Non-fault segments created.
 118  * Non fault segments are created to optimize sparc V9 code which uses
 119  * the sparc nonfaulting load ASI (ASI_PRIMARY_NOFAULT).
 120  *
 121  * There are several reasons why creating too many non-fault segments
 122  * could cause problems.
 123  *
 124  *      First, excessive allocation of kernel resources for the seg
 125  *      structures and the HAT data to map the zero pages.
 126  *
 127  *      Secondly, creating nofault segments actually uses up user virtual
 128  *      address space. This makes it unavailable for subsequent mmap(0, ...)
 129  *      calls which use as_gap() to find empty va regions.  Creation of too
 130  *      many nofault segments could thus interfere with the ability of the
 131  *      runtime linker to load a shared object.
 132  */
 133 #define MAXSEGFORNF     (10000)
 134 #define MAXNFSEARCH     (5)
 135 
 136 
 137 /*
 138  * Must be called from startup()
 139  */
 140 void
 141 segnf_init()
 142 {
 143         mutex_init(&segnf_lock, NULL, MUTEX_DEFAULT, NULL);
 144 }
 145 
 146 
 147 /*
 148  * Create a no-fault segment.
 149  *
 150  * The no-fault segment is not technically necessary, as the code in
 151  * nfload() in trap.c will emulate the SPARC instruction and load
 152  * a value of zero in the destination register.
 153  *
 154  * However, this code tries to put a page of zero's at the nofault address
 155  * so that subsequent non-faulting loads to the same page will not
 156  * trap with a tlb miss.
 157  *
 158  * In order to help limit the number of segments we merge adjacent nofault
 159  * segments into a single segment.  If we get a large number of segments
 160  * we'll also try to delete a random other nf segment.
 161  */
 162 /* ARGSUSED */
 163 int
 164 segnf_create(struct seg *seg, void *argsp)
 165 {
 166         uint_t prot;
 167         pgcnt_t vacpgs;
 168         u_offset_t off = 0;
 169         caddr_t vaddr = NULL;
 170         int i, color;
 171         struct seg *s1;
 172         struct seg *s2;
 173         size_t size;
 174         struct as *as = seg->s_as;
 175 
 176         ASSERT(as && AS_WRITE_HELD(as, &as->a_lock));
 177 
 178         /*
 179          * Need a page per virtual color or just 1 if no vac.
 180          */
 181         mutex_enter(&segnf_lock);
 182         if (nfpp == NULL) {
 183                 struct seg kseg;
 184 
 185                 vacpgs = 1;
 186                 if (shm_alignment > PAGESIZE) {
 187                         vacpgs = shm_alignment >> PAGESHIFT;
 188                 }
 189 
 190                 nfpp = kmem_alloc(sizeof (*nfpp) * vacpgs, KM_SLEEP);
 191 
 192                 kseg.s_as = &kas;
 193                 for (i = 0; i < vacpgs; i++, off += PAGESIZE,
 194                     vaddr += PAGESIZE) {
 195                         nfpp[i] = page_create_va(&nfvp, off, PAGESIZE,
 196                             PG_WAIT | PG_NORELOC, &kseg, vaddr);
 197                         page_io_unlock(nfpp[i]);
 198                         page_downgrade(nfpp[i]);
 199                         pagezero(nfpp[i], 0, PAGESIZE);
 200                 }
 201         }
 202         mutex_exit(&segnf_lock);
 203 
 204         hat_map(as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
 205 
 206         /*
 207          * s_data can't be NULL because of ASSERTS in the common vm code.
 208          */
 209         seg->s_ops = &segnf_ops;
 210         seg->s_data = seg;
 211         seg->s_flags |= S_PURGE;
 212 
 213         mutex_enter(&as->a_contents);
 214         as->a_flags |= AS_NEEDSPURGE;
 215         mutex_exit(&as->a_contents);
 216 
 217         prot = PROT_READ;
 218         color = addr_to_vcolor(seg->s_base);
 219         if (as != &kas)
 220                 prot |= PROT_USER;
 221         hat_memload(as->a_hat, seg->s_base, nfpp[color],
 222             prot | HAT_NOFAULT, HAT_LOAD);
 223 
 224         /*
 225          * At this point see if we can concatenate a segment to
 226          * a non-fault segment immediately before and/or after it.
 227          */
 228         if ((s1 = AS_SEGPREV(as, seg)) != NULL &&
 229             s1->s_ops == &segnf_ops &&
 230             s1->s_base + s1->s_size == seg->s_base) {
 231                 size = s1->s_size;
 232                 seg_free(s1);
 233                 seg->s_base -= size;
 234                 seg->s_size += size;
 235         }
 236 
 237         if ((s2 = AS_SEGNEXT(as, seg)) != NULL &&
 238             s2->s_ops == &segnf_ops &&
 239             seg->s_base + seg->s_size == s2->s_base) {
 240                 size = s2->s_size;
 241                 seg_free(s2);
 242                 seg->s_size += size;
 243         }
 244 
 245         /*
 246          * if we already have a lot of segments, try to delete some other
 247          * nofault segment to reduce the probability of uncontrolled segment
 248          * creation.
 249          *
 250          * the code looks around quickly (no more than MAXNFSEARCH segments
 251          * each way) for another NF segment and then deletes it.
 252          */
 253         if (avl_numnodes(&as->a_segtree) > MAXSEGFORNF) {
 254                 size = 0;
 255                 s2 = NULL;
 256                 s1 = AS_SEGPREV(as, seg);
 257                 while (size++ < MAXNFSEARCH && s1 != NULL) {
 258                         if (s1->s_ops == &segnf_ops)
 259                                 s2 = s1;
 260                         s1 = AS_SEGPREV(s1->s_as, seg);
 261                 }
 262                 if (s2 == NULL) {
 263                         s1 = AS_SEGNEXT(as, seg);
 264                         while (size-- > 0 && s1 != NULL) {
 265                                 if (s1->s_ops == &segnf_ops)
 266                                         s2 = s1;
 267                                 s1 = AS_SEGNEXT(as, seg);
 268                         }
 269                 }
 270                 if (s2 != NULL)
 271                         seg_unmap(s2);
 272         }
 273 
 274         return (0);
 275 }
 276 
 277 /*
 278  * Never really need "No fault" segments, so they aren't dup'd.
 279  */
 280 /* ARGSUSED */
 281 static int
 282 segnf_dup(struct seg *seg, struct seg *newseg)
 283 {
 284         panic("segnf_dup");
 285         return (0);
 286 }
 287 
 288 /*
 289  * Split a segment at addr for length len.
 290  */
 291 static int
 292 segnf_unmap(struct seg *seg, caddr_t addr, size_t len)
 293 {
 294         ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
 295 
 296         /*
 297          * Check for bad sizes.
 298          */
 299         if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
 300             (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
 301                 cmn_err(CE_PANIC, "segnf_unmap: bad unmap size");
 302         }
 303 
 304         /*
 305          * Unload any hardware translations in the range to be taken out.
 306          */
 307         hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD_UNMAP);
 308 
 309         if (addr == seg->s_base && len == seg->s_size) {
 310                 /*
 311                  * Freeing entire segment.
 312                  */
 313                 seg_free(seg);
 314         } else if (addr == seg->s_base) {
 315                 /*
 316                  * Freeing the beginning of the segment.
 317                  */
 318                 seg->s_base += len;
 319                 seg->s_size -= len;
 320         } else if (addr + len == seg->s_base + seg->s_size) {
 321                 /*
 322                  * Freeing the end of the segment.
 323                  */
 324                 seg->s_size -= len;
 325         } else {
 326                 /*
 327                  * The section to go is in the middle of the segment, so we
 328                  * have to cut it into two segments.  We shrink the existing
 329                  * "seg" at the low end, and create "nseg" for the high end.
 330                  */
 331                 caddr_t nbase = addr + len;
 332                 size_t nsize = (seg->s_base + seg->s_size) - nbase;
 333                 struct seg *nseg;
 334 
 335                 /*
 336                  * Trim down "seg" before trying to stick "nseg" into the as.
 337                  */
 338                 seg->s_size = addr - seg->s_base;
 339                 nseg = seg_alloc(seg->s_as, nbase, nsize);
 340                 if (nseg == NULL)
 341                         cmn_err(CE_PANIC, "segnf_unmap: seg_alloc failed");
 342 
 343                 /*
 344                  * s_data can't be NULL because of ASSERTs in common VM code.
 345                  */
 346                 nseg->s_ops = seg->s_ops;
 347                 nseg->s_data = nseg;
 348                 nseg->s_flags |= S_PURGE;
 349                 mutex_enter(&seg->s_as->a_contents);
 350                 seg->s_as->a_flags |= AS_NEEDSPURGE;
 351                 mutex_exit(&seg->s_as->a_contents);
 352         }
 353 
 354         return (0);
 355 }
 356 
 357 /*
 358  * Free a segment.
 359  */
 360 static void
 361 segnf_free(struct seg *seg)
 362 {
 363         ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
 364 }
 365 
 366 /*
 367  * No faults allowed on segnf.
 368  */
 369 static faultcode_t
 370 segnf_nomap(void)
 371 {
 372         return (FC_NOMAP);
 373 }
 374 
 375 /* ARGSUSED */
 376 static int
 377 segnf_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
 378 {
 379         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 380         return (EACCES);
 381 }
 382 
 383 /* ARGSUSED */
 384 static int
 385 segnf_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
 386 {
 387         uint_t sprot;
 388         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 389 
 390         sprot = seg->s_as == &kas ?  PROT_READ : PROT_READ|PROT_USER;
 391         return ((prot & sprot) == prot ? 0 : EACCES);
 392 }
 393 
 394 static void
 395 segnf_badop(void)
 396 {
 397         panic("segnf_badop");
 398         /*NOTREACHED*/
 399 }
 400 
 401 static int
 402 segnf_nop(void)
 403 {
 404         return (0);
 405 }
 406 
 407 static int
 408 segnf_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
 409 {
 410         size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
 411         size_t p;
 412         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 413 
 414         for (p = 0; p < pgno; ++p)
 415                 protv[p] = PROT_READ;
 416         return (0);
 417 }
 418 
 419 /* ARGSUSED */
 420 static u_offset_t
 421 segnf_getoffset(struct seg *seg, caddr_t addr)
 422 {
 423         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 424 
 425         return ((u_offset_t)0);
 426 }
 427 
 428 /* ARGSUSED */
 429 static int
 430 segnf_gettype(struct seg *seg, caddr_t addr)
 431 {
 432         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 433 
 434         return (MAP_SHARED);
 435 }
 436 
 437 /* ARGSUSED */
 438 static int
 439 segnf_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
 440 {
 441         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 442 
 443         *vpp = &nfvp;
 444         return (0);
 445 }
 446 
 447 /*ARGSUSED*/
 448 static int
 449 segnf_pagelock(struct seg *seg, caddr_t addr, size_t len,
 450     struct page ***ppp, enum lock_type type, enum seg_rw rw)
 451 {
 452         return (ENOTSUP);
 453 }