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 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 static void reserve_area( void *addr, void *end )
 247 {
 248     size_t size = (char *)end - (char *)addr;
 249
 250 #if (defined(__svr4__) || defined(__NetBSD__)) && !defined(MAP_TRYFIXED)
 251     /* try_mmap_fixed is inefficient when using vfork, so we need a different algorithm here */
 252     /* we assume no other thread is running at this point */
 253     size_t i, pagesize = getpagesize();
 254     char vec;
 255
 256     while (size)
 257     {
 258         for (i = 0; i < size; i += pagesize)
 259             if (mincore( (caddr_t)addr + i, pagesize, &vec ) != -1) break;
 260
 261         i &= ~granularity_mask;
 262         if (i && mmap( addr, i, PROT_NONE, MAP_FIXED | MAP_PRIVATE | MAP_ANON | MAP_NORESERVE,
 263                        get_fdzero(), 0 ) != (void *)-1)
 264             wine_mmap_add_reserved_area( addr, i );
 265
 266         i += granularity_mask + 1;
 267         if ((char *)addr + i < (char *)addr) break;  /* overflow */
 268         addr = (char *)addr + i;
 269         if (addr >= end) break;
 270         size = (char *)end - (char *)addr;
 271     }
 272 #else
 273     if (!size) return;
 274
 275     if (mmap_reserve( addr, size ))
 276     {
 277         wine_mmap_add_reserved_area( addr, size );
 278         return;
 279     }
 280     if (size > granularity_mask + 1)
 281     {
 282         size_t new_size = (size / 2) & ~granularity_mask;
 283         reserve_area( addr, (char *)addr + new_size );
 284         reserve_area( (char *)addr + new_size, end );
 285     }
 286 #endif
 287 }
 288
 289
 290 /***********************************************************************
 291  *           reserve_dos_area
 292  *
 293  * Reserve the DOS area (0x00000000-0x00110000).
 294  */
 295 static void reserve_dos_area(void)
 296 {
 297     const size_t page_size = getpagesize();
 298     const size_t dos_area_size = 0x110000;
 299     void *ptr;
 300
 301     /* first page has to be handled specially */
 302     ptr = wine_anon_mmap( (void *)page_size, dos_area_size - page_size, PROT_NONE, MAP_NORESERVE );
 303     if (ptr != (void *)page_size)
 304     {
 305         if (ptr != (void *)-1) munmap( ptr, dos_area_size - page_size );
 306         return;
 307     }
 308     /* now add first page with MAP_FIXED */
 309     wine_anon_mmap( NULL, page_size, PROT_NONE, MAP_NORESERVE|MAP_FIXED );
 310     wine_mmap_add_reserved_area( NULL, dos_area_size );
 311 }
 312
 313
 314 /***********************************************************************
 315  *           mmap_init
 316  */
 317 void mmap_init(void)
 318 {
 319     struct reserved_area *area;
 320     struct list *ptr;
 321 #if defined(__i386__) && !defined(__FreeBSD__) && !defined(__FreeBSD_kernel__)  /* commented out until FreeBSD gets fixed */
 322     char stack;
 323     char * const stack_ptr = &stack;
 324     char *user_space_limit = (char *)0x7ffe0000;
 325
 326     /* check for a reserved area starting at the user space limit */
 327     /* to avoid wasting time trying to allocate it again */
 328     LIST_FOR_EACH( ptr, &reserved_areas )
 329     {
 330         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 331         if ((char *)area->base > user_space_limit) break;
 332         if ((char *)area->base + area->size > user_space_limit)
 333         {
 334             user_space_limit = (char *)area->base + area->size;
 335             break;
 336         }
 337     }
 338
 339     if (stack_ptr >= user_space_limit)
 340     {
 341         char *end = 0;
 342         char *base = stack_ptr - ((unsigned int)stack_ptr & granularity_mask) - (granularity_mask + 1);
 343         if (base > user_space_limit) reserve_area( user_space_limit, base );
 344         base = stack_ptr - ((unsigned int)stack_ptr & granularity_mask) + (granularity_mask + 1);
 345 #ifdef linux
 346         /* Linux heuristic: assume the stack is near the end of the address */
 347         /* space, this avoids a lot of futile allocation attempts */
 348         end = (char *)(((unsigned long)base + 0x0fffffff) & 0xf0000000);
 349 #endif
 350         reserve_area( base, end );
 351     }
 352     else reserve_area( user_space_limit, 0 );
 353 #endif /* __i386__ */
 354
 355     /* reserve the DOS area if not already done */
 356
 357     ptr = list_head( &reserved_areas );
 358     if (ptr)
 359     {
 360         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 361         if (!area->base) return;  /* already reserved */
 362     }
 363     reserve_dos_area();
 364 }
 365
 366 #else /* HAVE_MMAP */
 367
 368 void *wine_anon_mmap( void *start, size_t size, int prot, int flags )
 369 {
 370     return (void *)-1;
 371 }
 372
 373 static inline int munmap( void *ptr, size_t size )
 374 {
 375     return 0;
 376 }
 377
 378 void mmap_init(void)
 379 {
 380 }
 381
 382 #endif
 383
 384 /***********************************************************************
 385  *           wine_mmap_add_reserved_area
 386  *
 387  * Add an address range to the list of reserved areas.
 388  * Caller must have made sure the range is not used by anything else.
 389  *
 390  * Note: the reserved areas functions are not reentrant, caller is
 391  * responsible for proper locking.
 392  */
 393 void wine_mmap_add_reserved_area( void *addr, size_t size )
 394 {
 395     struct reserved_area *area;
 396     struct list *ptr;
 397
 398     if (!((char *)addr + size)) size--;  /* avoid wrap-around */
 399
 400     LIST_FOR_EACH( ptr, &reserved_areas )
 401     {
 402         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 403         if (area->base > addr)
 404         {
 405             /* try to merge with the next one */
 406             if ((char *)addr + size == (char *)area->base)
 407             {
 408                 area->base = addr;
 409                 area->size += size;
 410                 return;
 411             }
 412             break;
 413         }
 414         else if ((char *)area->base + area->size == (char *)addr)
 415         {
 416             /* merge with the previous one */
 417             area->size += size;
 418
 419             /* try to merge with the next one too */
 420             if ((ptr = list_next( &reserved_areas, ptr )))
 421             {
 422                 struct reserved_area *next = LIST_ENTRY( ptr, struct reserved_area, entry );
 423                 if ((char *)addr + size == (char *)next->base)
 424                 {
 425                     area->size += next->size;
 426                     list_remove( &next->entry );
 427                     free( next );
 428                 }
 429             }
 430             return;
 431         }
 432     }
 433
 434     if ((area = malloc( sizeof(*area) )))
 435     {
 436         area->base = addr;
 437         area->size = size;
 438         list_add_before( ptr, &area->entry );
 439     }
 440 }
 441
 442
 443 /***********************************************************************
 444  *           wine_mmap_remove_reserved_area
 445  *
 446  * Remove an address range from the list of reserved areas.
 447  * If 'unmap' is non-zero the range is unmapped too.
 448  *
 449  * Note: the reserved areas functions are not reentrant, caller is
 450  * responsible for proper locking.
 451  */
 452 void wine_mmap_remove_reserved_area( void *addr, size_t size, int unmap )
 453 {
 454     struct reserved_area *area;
 455     struct list *ptr;
 456
 457     if (!((char *)addr + size)) size--;  /* avoid wrap-around */
 458
 459     ptr = list_head( &reserved_areas );
 460     /* find the first area covering address */
 461     while (ptr)
 462     {
 463         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 464         if ((char *)area->base >= (char *)addr + size) break;  /* outside the range */
 465         if ((char *)area->base + area->size > (char *)addr)  /* overlaps range */
 466         {
 467             if (area->base >= addr)
 468             {
 469                 if ((char *)area->base + area->size > (char *)addr + size)
 470                 {
 471                     /* range overlaps beginning of area only -> shrink area */
 472                     if (unmap) munmap( area->base, (char *)addr + size - (char *)area->base );
 473                     area->size -= (char *)addr + size - (char *)area->base;
 474                     area->base = (char *)addr + size;
 475                     break;
 476                 }
 477                 else
 478                 {
 479                     /* range contains the whole area -> remove area completely */
 480                     ptr = list_next( &reserved_areas, ptr );
 481                     if (unmap) munmap( area->base, area->size );
 482                     list_remove( &area->entry );
 483                     free( area );
 484                     continue;
 485                 }
 486             }
 487             else
 488             {
 489                 if ((char *)area->base + area->size > (char *)addr + size)
 490                 {
 491                     /* range is in the middle of area -> split area in two */
 492                     struct reserved_area *new_area = malloc( sizeof(*new_area) );
 493                     if (new_area)
 494                     {
 495                         new_area->base = (char *)addr + size;
 496                         new_area->size = (char *)area->base + area->size - (char *)new_area->base;
 497                         list_add_after( ptr, &new_area->entry );
 498                     }
 499                     else size = (char *)area->base + area->size - (char *)addr;
 500                     area->size = (char *)addr - (char *)area->base;
 501                     if (unmap) munmap( addr, size );
 502                     break;
 503                 }
 504                 else
 505                 {
 506                     /* range overlaps end of area only -> shrink area */
 507                     if (unmap) munmap( addr, (char *)area->base + area->size - (char *)addr );
 508                     area->size = (char *)addr - (char *)area->base;
 509                 }
 510             }
 511         }
 512         ptr = list_next( &reserved_areas, ptr );
 513     }
 514 }
 515
 516
 517 /***********************************************************************
 518  *           wine_mmap_is_in_reserved_area
 519  *
 520  * Check if the specified range is included in a reserved area.
 521  * Returns 1 if range is fully included, 0 if range is not included
 522  * at all, and -1 if it is only partially included.
 523  *
 524  * Note: the reserved areas functions are not reentrant, caller is
 525  * responsible for proper locking.
 526  */
 527 int wine_mmap_is_in_reserved_area( void *addr, size_t size )
 528 {
 529     struct reserved_area *area;
 530     struct list *ptr;
 531
 532     LIST_FOR_EACH( ptr, &reserved_areas )
 533     {
 534         area = LIST_ENTRY( ptr, struct reserved_area, entry );
 535         if (area->base > addr) break;
 536         if ((char *)area->base + area->size <= (char *)addr) continue;
 537         /* area must contain block completely */
 538         if ((char *)area->base + area->size < (char *)addr + size) return -1;
 539         return 1;
 540     }
 541     return 0;
 542 }
 543
 544
 545 /***********************************************************************
 546  *           wine_mmap_enum_reserved_areas
 547  *
 548  * Enumerate the list of reserved areas, sorted by addresses.
 549  * If enum_func returns a non-zero value, enumeration is stopped and the value is returned.
 550  *
 551  * Note: the reserved areas functions are not reentrant, caller is
 552  * responsible for proper locking.
 553  */
 554 int wine_mmap_enum_reserved_areas( int (*enum_func)(void *base, size_t size, void *arg), void *arg,
 555                                    int top_down )
 556 {
 557     int ret = 0;
 558     struct list *ptr;
 559
 560     if (top_down)
 561     {
 562         for (ptr = reserved_areas.prev; ptr != &reserved_areas; ptr = ptr->prev)
 563         {
 564             struct reserved_area *area = LIST_ENTRY( ptr, struct reserved_area, entry );
 565             if ((ret = enum_func( area->base, area->size, arg ))) break;
 566         }
 567     }
 568     else
 569     {
 570         for (ptr = reserved_areas.next; ptr != &reserved_areas; ptr = ptr->next)
 571         {
 572             struct reserved_area *area = LIST_ENTRY( ptr, struct reserved_area, entry );
 573             if ((ret = enum_func( area->base, area->size, arg ))) break;
 574         }
 575     }
 576     return ret;
 577 }