libs/wine/mmap.c

   1 /*
   2  * Wine memory mappings support
   3  *
   4  * Copyright 2000, 2004 Alexandre Julliard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  19  */
  20
  21 #include "config.h"
  22 #include "wine/port.h"
  23
  24 #include <assert.h>
  25 #include <ctype.h>
  26 #include <fcntl.h>
  27 #include <stdlib.h>
  28 #include <stdio.h>
  29 #include <string.h>
  30 #include <sys/types.h>
  31 #ifdef HAVE_SYS_MMAN_H
  32 #include <sys/mman.h>
  33 #endif
  34 #ifdef HAVE_UNISTD_H
  35 # include <unistd.h>
  36 #endif
  37 #ifdef HAVE_STDINT_H
  38 # include <stdint.h>
  39 #endif
  40
  41 #include "wine/library.h"
  42 #include "wine/list.h"
  43
  44 struct reserved_area
  45 {
  46     struct list entry;
  47     void       *base;
  48     size_t      size;
  49 };
  50
  51 static struct list reserved_areas = LIST_INIT(reserved_areas);
  52 static const unsigned int granularity_mask = 0xffff;  /* reserved areas have 64k granularity */
  53
  54 #ifdef HAVE_MMAP
  55
  56 #ifndef MAP_NORESERVE
  57 #define MAP_NORESERVE 0
  58 #endif
  59 #ifndef MAP_PRIVATE
  60 #define MAP_PRIVATE 0
  61 #endif
  62 #ifndef MAP_ANON
  63 #define MAP_ANON 0
  64 #endif
  65
  66 static inline int get_fdzero(void)
  67 {
  68     static int fd = -1;
  69
  70     if (MAP_ANON == 0 && fd == -1)
  71     {
  72         if ((fd = open( "/dev/zero", O_RDONLY )) == -1)
  73         {
  74             perror( "/dev/zero: open" );
  75             exit(1);
  76         }
  77     }
  78     return fd;
  79 }
  80
  81 #if (defined(__svr4__) || defined(__NetBSD__)) && !defined(MAP_TRYFIXED)
  82 /***********************************************************************
  83  *             try_mmap_fixed
  84  *
  85  * The purpose of this routine is to emulate the behaviour of
  86  * the Linux mmap() routine if a non-NULL address is passed,
  87  * but the MAP_FIXED flag is not set.  Linux in this case tries
  88  * to place the mapping at the specified address, *unless* the
  89  * range is already in use.  Solaris, however, completely ignores
  90  * the address argument in this case.
  91  *
  92  * As Wine code occasionally relies on the Linux behaviour, e.g. to
  93  * be able to map non-relocatable PE executables to their proper
  94  * start addresses, or to map the DOS memory to 0, this routine
  95  * emulates the Linux behaviour by checking whether the desired
  96  * address range is still available, and placing the mapping there
  97  * using MAP_FIXED if so.
  98  */
  99 static int try_mmap_fixed (void *addr, size_t len, int prot, int flags,
 100                            int fildes, off_t off)
 101 {
 102     char * volatile result = NULL;
 103     int pagesize = getpagesize();
 104     pid_t pid;
 105
 106     /* We only try to map to a fixed address if
 107        addr is non-NULL and properly aligned,
 108        and MAP_FIXED isn't already specified. */
 109
 110     if ( !addr )
 111         return 0;
 112     if ( (uintptr_t)addr & (pagesize-1) )
 113         return 0;
 114     if ( flags & MAP_FIXED )
 115         return 0;
 116
 117     /* We use vfork() to freeze all threads of the
 118        current process.  This allows us to check without
 119        race condition whether the desired memory range is
 120        already in use.  Note that because vfork() shares
 121        the address spaces between parent and child, we
 122        can actually perform the mapping in the child. */
 123
 124     if ( (pid = vfork()) == -1 )
 125     {
 126         perror("try_mmap_fixed: vfork");
 127         exit(1);
 128     }
 129     if ( pid == 0 )
 130     {
 131         int i;
 132         char vec;
 133
 134         /* We call mincore() for every page in the desired range.
 135            If any of these calls succeeds, the page is already
 136            mapped and we must fail. */
 137         for ( i = 0; i < len; i += pagesize )
 138             if ( mincore( (caddr_t)addr + i, pagesize, &vec ) != -1 )
 139                _exit(1);
 140
 141         /* Perform the mapping with MAP_FIXED set.  This is safe
 142            now, as none of the pages is currently in use. */
 143         result = mmap( addr, len, prot, flags | MAP_FIXED, fildes, off );
 144         if ( result == addr )
 145             _exit(0);
 146
 147         if ( result != (void *) -1 ) /* This should never happen ... */
 148             munmap( result, len );
 149
 150        _exit(1);
 151     }
 152
 153     /* vfork() lets the parent continue only after the child
 154        has exited.  Furthermore, Wine sets SIGCHLD to SIG_IGN,
 155        so we don't need to wait for the child. */
 156
 157     return result == addr;
 158 }
 159
 160 #elif defined(__APPLE__)
 161
 162 #include <mach/mach_init.h>
 163 #include <mach/vm_map.h>
 164
 165 /*
 166  * On Darwin, we can use the Mach call vm_allocate to allocate
 167  * anonymous memory at the specified address, and then use mmap with
 168  * MAP_FIXED to replace the mapping.
 169  */
 170 static int try_mmap_fixed (void *addr, size_t len, int prot, int flags,
 171                            int fildes, off_t off)
 172 {
 173     vm_address_t result = (vm_address_t)addr;
 174
 175     if (!vm_allocate(mach_task_self(),&result,len,0))
 176     {
 177         if (mmap( (void *)result, len, prot, flags | MAP_FIXED, fildes, off ) != MAP_FAILED)
 178             return 1;
 179         vm_deallocate(mach_task_self(),result,len);
 180     }
 181     return 0;
 182 }
 183
 184 #endif  /* (__svr4__ || __NetBSD__) && !MAP_TRYFIXED */
 185
 186
 187 /***********************************************************************
 188  *              wine_anon_mmap
 189  *
 190  * Portable wrapper for anonymous mmaps
 191  */
 192 void *wine_anon_mmap( void *start, size_t size, int prot, int flags )
 193 {
 194 #ifdef MAP_SHARED
 195     flags &= ~MAP_SHARED;
 196 #endif
 197
 198     /* Linux EINVAL's on us if we don't pass MAP_PRIVATE to an anon mmap */
 199     flags |= MAP_PRIVATE | MAP_ANON;
 200
 201     if (!(flags & MAP_FIXED))
 202     {
 203 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
 204         /* Even FreeBSD 5.3 does not properly support NULL here. */
 205         if( start == NULL ) start = (void *)0x110000;
 206 #endif
 207
 208 #ifdef MAP_TRYFIXED
 209         /* If available, this will attempt a fixed mapping in-kernel */
 210         flags |= MAP_TRYFIXED;
 211 #elif defined(__svr4__) || defined(__NetBSD__) || defined(__APPLE__)
 212         if ( try_mmap_fixed( start, size, prot, flags, get_fdzero(), 0 ) )
 213             return start;
 214 #endif
 215     }
 216     return mmap( start, size, prot, flags, get_fdzero(), 0 );
 217 }
 218
 219
 220 /***********************************************************************
 221  *              mmap_reserve
 222  *
 223  * mmap wrapper used for reservations, only maps the specified address
 224  */
 225 static inline int mmap_reserve( void *addr, size_t size )
 226 {
 227     void *ptr;
 228     int flags = MAP_PRIVATE | MAP_ANON | MAP_NORESERVE;
 229
 230 #ifdef MAP_TRYFIXED
 231     flags |= MAP_TRYFIXED;
 232 #elif defined(__APPLE__)
 233     return try_mmap_fixed( addr, size, PROT_NONE, flags, get_fdzero(), 0 );
 234 #endif
 235     ptr = mmap( addr, size, PROT_NONE, flags, get_fdzero(), 0 );
 236     if (ptr != addr && ptr != (void *)-1)  munmap( ptr, size );
 237     return (ptr == addr);
 238 }
 239
 240
 241 /***********************************************************************
 242  *           reserve_area
 243  *
 244  * Reserve as much memory as possible in the given area.
 245  */
 246 #if defined(__i386__) && !defined(__FreeBSD__) && !defined(__FreeBSD_kernel__)  /* commented out until FreeBSD gets fixed */
 247 static void reserve_area( void *addr, void *end )
 248 {
 249     size_t size = (char *)end - (char *)addr;
 250
 251 #if (defined(__svr4__) || defined(__NetBSD__)) && !defined(MAP_TRYFIXED)
 252     /* try_mmap_fixed is inefficient when using vfork, so we need a different algorithm here */
 253     /* we assume no other thread is running at this point */
 254     size_t i, pagesize = getpagesize();
 255     char vec;
 256
 257     while (size)
 258     {
 259         for (i = 0; i < size; i += pagesize)
 260             if (mincore( (caddr_t)addr + i, pagesize, &vec ) != -1) break;
 261
 262         i &= ~granularity_mask;
 263         if (i && mmap( addr, i, PROT_NONE, MAP_FIXED | MAP_PRIVATE | MAP_ANON | MAP_NORESERVE,
 264                        get_fdzero(), 0 ) != (void *)-1)
 265             wine_mmap_add_reserved_area( addr, i );
 266
 267         i += granularity_mask + 1;
 268         if ((char *)addr + i < (char *)addr) break;  /* overflow */
 269         addr = (char *)addr + i;
 270         if (addr >= end) break;
 271         size = (char *)end - (char *)addr;
 272     }
 273 #else
 274     if (!size) return;
 275
 276     if (mmap_reserve( addr, size ))
 277     {
 278         wine_mmap_add_reserved_area( addr, size );
 279         return;
 280     }
 281     if (size > granularity_mask + 1)
 282     {
 283         size_t new_size = (size / 2) & ~granularity_mask;
 284         reserve_area( addr, (char *)addr + new_size );
 285         reserve_area( (char *)addr + new_size, end );
 286     }
 287 #endif
 288 }
 289 #endif
 290
 291
 292 /***********************************************************************
 293  *           reserve_dos_area
 294  *
 295  * Reserve the DOS area (0x00000000-0x00110000).
 296  */
 297 static void reserve_dos_area(void)
 298 {
 299     const size_t page_size = getpagesize();
 300     const size_t dos_area_size = 0x110000;
 301     void *ptr;
 302
 303     /* first page has to be handled specially */
 304     ptr = wine_anon_mmap( (void *)page_size, dos_area_size - page_size, PROT_NONE, MAP_NORESERVE );
 305     if (ptr != (void *)page_size)
 306     {
 307         if (ptr != (void *)-1) munmap( ptr, dos_area_size - page_size );
 308         return;
 309     }
 310     /* now add first page with MAP_FIXED */
 311     wine_anon_mmap( NULL, page_size, PROT_NONE, MAP_NORESERVE|MAP_FIXED );
 312     wine_mmap_add_reserved_area( NULL, dos_area_size );
 313 }
 314
 315
 316 /***********************************************************************
 317  *           mmap_init
 318  */
 319 void mmap_init(void)
 320 {
 321     struct reserved_area *area;
 322     struct list *ptr;
 323 #if defined(__i386__) && !defined(__FreeBSD__) && !defined(__FreeBSD_kernel__)  /* commented out until FreeBSD gets fixed */
 324     char stack;
 325     char * const stack_ptr = &stack;
 326     char *user_space_limit = (char *)0x7ffe0000;
 327
 328     /* check for a reserved area starting at the user space limit */
 329     /* to avoid wasting time trying to allocate it again */
 330     LIST_FOR_EACH( ptr, &reserved_areas )
 331     {
 332         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 333         if ((char *)area->base > user_space_limit) break;
 334         if ((char *)area->base + area->size > user_space_limit)
 335         {
 336             user_space_limit = (char *)area->base + area->size;
 337             break;
 338         }
 339     }
 340
 341     if (stack_ptr >= user_space_limit)
 342     {
 343         char *end = 0;
 344         char *base = stack_ptr - ((unsigned int)stack_ptr & granularity_mask) - (granularity_mask + 1);
 345         if (base > user_space_limit) reserve_area( user_space_limit, base );
 346         base = stack_ptr - ((unsigned int)stack_ptr & granularity_mask) + (granularity_mask + 1);
 347 #ifdef linux
 348         /* Linux heuristic: assume the stack is near the end of the address */
 349         /* space, this avoids a lot of futile allocation attempts */
 350         end = (char *)(((unsigned long)base + 0x0fffffff) & 0xf0000000);
 351 #endif
 352         reserve_area( base, end );
 353     }
 354     else reserve_area( user_space_limit, 0 );
 355 #endif /* __i386__ */
 356
 357     /* reserve the DOS area if not already done */
 358
 359     ptr = list_head( &reserved_areas );
 360     if (ptr)
 361     {
 362         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 363         if (!area->base) return;  /* already reserved */
 364     }
 365     reserve_dos_area();
 366 }
 367
 368 #else /* HAVE_MMAP */
 369
 370 void *wine_anon_mmap( void *start, size_t size, int prot, int flags )
 371 {
 372     return (void *)-1;
 373 }
 374
 375 static inline int munmap( void *ptr, size_t size )
 376 {
 377     return 0;
 378 }
 379
 380 void mmap_init(void)
 381 {
 382 }
 383
 384 #endif
 385
 386 /***********************************************************************
 387  *           wine_mmap_add_reserved_area
 388  *
 389  * Add an address range to the list of reserved areas.
 390  * Caller must have made sure the range is not used by anything else.
 391  *
 392  * Note: the reserved areas functions are not reentrant, caller is
 393  * responsible for proper locking.
 394  */
 395 void wine_mmap_add_reserved_area( void *addr, size_t size )
 396 {
 397     struct reserved_area *area;
 398     struct list *ptr;
 399
 400     if (!((char *)addr + size)) size--;  /* avoid wrap-around */
 401
 402     LIST_FOR_EACH( ptr, &reserved_areas )
 403     {
 404         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 405         if (area->base > addr)
 406         {
 407             /* try to merge with the next one */
 408             if ((char *)addr + size == (char *)area->base)
 409             {
 410                 area->base = addr;
 411                 area->size += size;
 412                 return;
 413             }
 414             break;
 415         }
 416         else if ((char *)area->base + area->size == (char *)addr)
 417         {
 418             /* merge with the previous one */
 419             area->size += size;
 420
 421             /* try to merge with the next one too */
 422             if ((ptr = list_next( &reserved_areas, ptr )))
 423             {
 424                 struct reserved_area *next = LIST_ENTRY( ptr, struct reserved_area, entry );
 425                 if ((char *)addr + size == (char *)next->base)
 426                 {
 427                     area->size += next->size;
 428                     list_remove( &next->entry );
 429                     free( next );
 430                 }
 431             }
 432             return;
 433         }
 434     }
 435
 436     if ((area = malloc( sizeof(*area) )))
 437     {
 438         area->base = addr;
 439         area->size = size;
 440         list_add_before( ptr, &area->entry );
 441     }
 442 }
 443
 444
 445 /***********************************************************************
 446  *           wine_mmap_remove_reserved_area
 447  *
 448  * Remove an address range from the list of reserved areas.
 449  * If 'unmap' is non-zero the range is unmapped too.
 450  *
 451  * Note: the reserved areas functions are not reentrant, caller is
 452  * responsible for proper locking.
 453  */
 454 void wine_mmap_remove_reserved_area( void *addr, size_t size, int unmap )
 455 {
 456     struct reserved_area *area;
 457     struct list *ptr;
 458
 459     if (!((char *)addr + size)) size--;  /* avoid wrap-around */
 460
 461     ptr = list_head( &reserved_areas );
 462     /* find the first area covering address */
 463     while (ptr)
 464     {
 465         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 466         if ((char *)area->base >= (char *)addr + size) break;  /* outside the range */
 467         if ((char *)area->base + area->size > (char *)addr)  /* overlaps range */
 468         {
 469             if (area->base >= addr)
 470             {
 471                 if ((char *)area->base + area->size > (char *)addr + size)
 472                 {
 473                     /* range overlaps beginning of area only -> shrink area */
 474                     if (unmap) munmap( area->base, (char *)addr + size - (char *)area->base );
 475                     area->size -= (char *)addr + size - (char *)area->base;
 476                     area->base = (char *)addr + size;
 477                     break;
 478                 }
 479                 else
 480                 {
 481                     /* range contains the whole area -> remove area completely */
 482                     ptr = list_next( &reserved_areas, ptr );
 483                     if (unmap) munmap( area->base, area->size );
 484                     list_remove( &area->entry );
 485                     free( area );
 486                     continue;
 487                 }
 488             }
 489             else
 490             {
 491                 if ((char *)area->base + area->size > (char *)addr + size)
 492                 {
 493                     /* range is in the middle of area -> split area in two */
 494                     struct reserved_area *new_area = malloc( sizeof(*new_area) );
 495                     if (new_area)
 496                     {
 497                         new_area->base = (char *)addr + size;
 498                         new_area->size = (char *)area->base + area->size - (char *)new_area->base;
 499                         list_add_after( ptr, &new_area->entry );
 500                     }
 501                     else size = (char *)area->base + area->size - (char *)addr;
 502                     area->size = (char *)addr - (char *)area->base;
 503                     if (unmap) munmap( addr, size );
 504                     break;
 505                 }
 506                 else
 507                 {
 508                     /* range overlaps end of area only -> shrink area */
 509                     if (unmap) munmap( addr, (char *)area->base + area->size - (char *)addr );
 510                     area->size = (char *)addr - (char *)area->base;
 511                 }
 512             }
 513         }
 514         ptr = list_next( &reserved_areas, ptr );
 515     }
 516 }
 517
 518
 519 /***********************************************************************
 520  *           wine_mmap_is_in_reserved_area
 521  *
 522  * Check if the specified range is included in a reserved area.
 523  * Returns 1 if range is fully included, 0 if range is not included
 524  * at all, and -1 if it is only partially included.
 525  *
 526  * Note: the reserved areas functions are not reentrant, caller is
 527  * responsible for proper locking.
 528  */
 529 int wine_mmap_is_in_reserved_area( void *addr, size_t size )
 530 {
 531     struct reserved_area *area;
 532     struct list *ptr;
 533
 534     LIST_FOR_EACH( ptr, &reserved_areas )
 535     {
 536         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 537         if (area->base > addr) break;
 538         if ((char *)area->base + area->size <= (char *)addr) continue;
 539         /* area must contain block completely */
 540         if ((char *)area->base + area->size < (char *)addr + size) return -1;
 541         return 1;
 542     }
 543     return 0;
 544 }
 545
 546
 547 /***********************************************************************
 548  *           wine_mmap_enum_reserved_areas
 549  *
 550  * Enumerate the list of reserved areas, sorted by addresses.
 551  * If enum_func returns a non-zero value, enumeration is stopped and the value is returned.
 552  *
 553  * Note: the reserved areas functions are not reentrant, caller is
 554  * responsible for proper locking.
 555  */
 556 int wine_mmap_enum_reserved_areas( int (*enum_func)(void *base, size_t size, void *arg), void *arg,
 557                                    int top_down )
 558 {
 559     int ret = 0;
 560     struct list *ptr;
 561
 562     if (top_down)
 563     {
 564         for (ptr = reserved_areas.prev; ptr != &reserved_areas; ptr = ptr->prev)
 565         {
 566             struct reserved_area *area = LIST_ENTRY( ptr, struct reserved_area, entry );
 567             if ((ret = enum_func( area->base, area->size, arg ))) break;
 568         }
 569     }
 570     else
 571     {
 572         for (ptr = reserved_areas.next; ptr != &reserved_areas; ptr = ptr->next)
 573         {
 574             struct reserved_area *area = LIST_ENTRY( ptr, struct reserved_area, entry );
 575             if ((ret = enum_func( area->base, area->size, arg ))) break;
 576         }
 577     }
 578     return ret;
 579 }