mono/utils/mono-mmap.c

   1 /*
   2  * mono-mmap.c: Support for mapping code into the process address space
   3  *
   4  * Author:
   5  *   Mono Team (mono-list@lists.ximian.com)
   6  *
   7  * Copyright 2001-2008 Novell, Inc.
   8  * Licensed under the MIT license. See LICENSE file in the project root for full license information.
   9  */
  10
  11 #include <config.h>
  12
  13 #ifndef HOST_WIN32
  14 #include <sys/types.h>
  15 #if HAVE_SYS_STAT_H
  16 #include <sys/stat.h>
  17 #endif
  18 #if HAVE_SYS_MMAN_H
  19 #include <sys/mman.h>
  20 #endif
  21 #ifdef HAVE_SIGNAL_H
  22 #include <signal.h>
  23 #endif
  24 #include <fcntl.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <stdlib.h>
  28 #include <errno.h>
  29 #endif /* !HOST_WIN32 */
  30
  31 #include "mono-mmap.h"
  32 #include "mono-mmap-internals.h"
  33 #include "mono-proclib.h"
  34 #include <mono/utils/mono-threads.h>
  35 #include <mono/utils/atomic.h>
  36 #include <mono/utils/mono-counters.h>
  37
  38 #define BEGIN_CRITICAL_SECTION do { \
  39         MonoThreadInfo *__info = mono_thread_info_current_unchecked (); \
  40         if (__info) __info->inside_critical_region = TRUE;      \
  41
  42 #define END_CRITICAL_SECTION \
  43         if (__info) __info->inside_critical_region = FALSE;     \
  44 } while (0)     \
  45
  46 #ifndef MAP_ANONYMOUS
  47 #define MAP_ANONYMOUS MAP_ANON
  48 #endif
  49
  50 #ifndef MAP_32BIT
  51 #define MAP_32BIT 0
  52 #endif
  53
  54 typedef struct {
  55         int size;
  56         int pid;
  57         int reserved;
  58         short stats_start;
  59         short stats_end;
  60 } SAreaHeader;
  61
  62 void*
  63 malloc_shared_area (int pid)
  64 {
  65         int size = mono_pagesize ();
  66         SAreaHeader *sarea = (SAreaHeader *) g_malloc0 (size);
  67         sarea->size = size;
  68         sarea->pid = pid;
  69         sarea->stats_start = sizeof (SAreaHeader);
  70         sarea->stats_end = sizeof (SAreaHeader);
  71
  72         return sarea;
  73 }
  74
  75 char*
  76 aligned_address (char *mem, size_t size, size_t alignment)
  77 {
  78         char *aligned = (char*)((size_t)(mem + (alignment - 1)) & ~(alignment - 1));
  79         g_assert (aligned >= mem && aligned + size <= mem + size + alignment && !((size_t)aligned & (alignment - 1)));
  80         return aligned;
  81 }
  82
  83 static volatile size_t allocation_count [MONO_MEM_ACCOUNT_MAX];
  84
  85 void
  86 account_mem (MonoMemAccountType type, ssize_t size)
  87 {
  88 #if SIZEOF_VOID_P == 4
  89         InterlockedAdd ((volatile gint32*)&allocation_count [type], (gint32)size);
  90 #else
  91         InterlockedAdd64 ((volatile gint64*)&allocation_count [type], (gint64)size);
  92 #endif
  93 }
  94
  95 const char*
  96 mono_mem_account_type_name (MonoMemAccountType type)
  97 {
  98         static const char *names[] = {
  99                 "code",
 100                 "hazard pointers",
 101                 "domain",
 102                 "SGen internal",
 103                 "SGen nursery",
 104                 "SGen LOS",
 105                 "SGen mark&sweep",
 106                 "SGen card table",
 107                 "SGen shadow card table",
 108                 "SGen debugging",
 109                 "SGen binary protocol",
 110                 "exceptions",
 111                 "profiler",
 112                 "other"
 113         };
 114
 115         return names [type];
 116 }
 117
 118 void
 119 mono_mem_account_register_counters (void)
 120 {
 121         for (int i = 0; i < MONO_MEM_ACCOUNT_MAX; ++i) {
 122                 const char *prefix = "Valloc ";
 123                 const char *name = mono_mem_account_type_name (i);
 124                 char descr [128];
 125                 g_assert (strlen (prefix) + strlen (name) < sizeof (descr));
 126                 sprintf (descr, "%s%s", prefix, name);
 127                 mono_counters_register (descr, MONO_COUNTER_WORD | MONO_COUNTER_RUNTIME | MONO_COUNTER_BYTES | MONO_COUNTER_VARIABLE, (void*)&allocation_count [i]);
 128         }
 129 }
 130
 131 #ifdef HOST_WIN32
 132 // Windows specific implementation in mono-mmap-windows.c
 133 #define HAVE_VALLOC_ALIGNED
 134
 135 #else
 136
 137 static void* malloced_shared_area = NULL;
 138 #if defined(HAVE_MMAP)
 139
 140 /**
 141  * mono_pagesize:
 142  * Get the page size in use on the system. Addresses and sizes in the
 143  * mono_mmap(), mono_munmap() and mono_mprotect() calls must be pagesize
 144  * aligned.
 145  *
 146  * Returns: the page size in bytes.
 147  */
 148 int
 149 mono_pagesize (void)
 150 {
 151         static int saved_pagesize = 0;
 152         if (saved_pagesize)
 153                 return saved_pagesize;
 154         saved_pagesize = getpagesize ();
 155         return saved_pagesize;
 156 }
 157
 158 int
 159 mono_valloc_granule (void)
 160 {
 161         return mono_pagesize ();
 162 }
 163
 164 static int
 165 prot_from_flags (int flags)
 166 {
 167         int prot = PROT_NONE;
 168         /* translate the protection bits */
 169         if (flags & MONO_MMAP_READ)
 170                 prot |= PROT_READ;
 171         if (flags & MONO_MMAP_WRITE)
 172                 prot |= PROT_WRITE;
 173         if (flags & MONO_MMAP_EXEC)
 174                 prot |= PROT_EXEC;
 175         return prot;
 176 }
 177
 178 /**
 179  * mono_valloc:
 180  * @addr: memory address
 181  * @length: memory area size
 182  * @flags: protection flags
 183  *
 184  * Allocates @length bytes of virtual memory with the @flags
 185  * protection. @addr can be a preferred memory address or a
 186  * mandatory one if MONO_MMAP_FIXED is set in @flags.
 187  * @addr must be pagesize aligned and can be NULL.
 188  * @length must be a multiple of pagesize.
 189  *
 190  * Returns: NULL on failure, the address of the memory area otherwise
 191  */
 192 void*
 193 mono_valloc (void *addr, size_t length, int flags, MonoMemAccountType type)
 194 {
 195         void *ptr;
 196         int mflags = 0;
 197         int prot = prot_from_flags (flags);
 198         /* translate the flags */
 199         if (flags & MONO_MMAP_FIXED)
 200                 mflags |= MAP_FIXED;
 201         if (flags & MONO_MMAP_32BIT)
 202                 mflags |= MAP_32BIT;
 203
 204         mflags |= MAP_ANONYMOUS;
 205         mflags |= MAP_PRIVATE;
 206
 207         BEGIN_CRITICAL_SECTION;
 208         ptr = mmap (addr, length, prot, mflags, -1, 0);
 209         if (ptr == MAP_FAILED) {
 210                 int fd = open ("/dev/zero", O_RDONLY);
 211                 if (fd != -1) {
 212                         ptr = mmap (addr, length, prot, mflags, fd, 0);
 213                         close (fd);
 214                 }
 215         }
 216         END_CRITICAL_SECTION;
 217
 218         if (ptr == MAP_FAILED)
 219                 return NULL;
 220
 221         account_mem (type, (ssize_t)length);
 222
 223         return ptr;
 224 }
 225
 226 /**
 227  * mono_vfree:
 228  * @addr: memory address returned by mono_valloc ()
 229  * @length: size of memory area
 230  *
 231  * Remove the memory mapping at the address @addr.
 232  *
 233  * Returns: 0 on success.
 234  */
 235 int
 236 mono_vfree (void *addr, size_t length, MonoMemAccountType type)
 237 {
 238         int res;
 239         BEGIN_CRITICAL_SECTION;
 240         res = munmap (addr, length);
 241         END_CRITICAL_SECTION;
 242
 243         account_mem (type, -(ssize_t)length);
 244
 245         return res;
 246 }
 247
 248 /**
 249  * mono_file_map:
 250  * @length: size of data to map
 251  * @flags: protection flags
 252  * @fd: file descriptor
 253  * @offset: offset in the file
 254  * @ret_handle: pointer to storage for returning a handle for the map
 255  *
 256  * Map the area of the file pointed to by the file descriptor @fd, at offset
 257  * @offset and of size @length in memory according to the protection flags
 258  * @flags.
 259  * @offset and @length must be multiples of the page size.
 260  * @ret_handle must point to a void*: this value must be used when unmapping
 261  * the memory area using mono_file_unmap ().
 262  *
 263  */
 264 void*
 265 mono_file_map (size_t length, int flags, int fd, guint64 offset, void **ret_handle)
 266 {
 267         void *ptr;
 268         int mflags = 0;
 269         int prot = prot_from_flags (flags);
 270         /* translate the flags */
 271         if (flags & MONO_MMAP_PRIVATE)
 272                 mflags |= MAP_PRIVATE;
 273         if (flags & MONO_MMAP_SHARED)
 274                 mflags |= MAP_SHARED;
 275         if (flags & MONO_MMAP_FIXED)
 276                 mflags |= MAP_FIXED;
 277         if (flags & MONO_MMAP_32BIT)
 278                 mflags |= MAP_32BIT;
 279
 280         BEGIN_CRITICAL_SECTION;
 281         ptr = mmap (0, length, prot, mflags, fd, offset);
 282         END_CRITICAL_SECTION;
 283         if (ptr == MAP_FAILED)
 284                 return NULL;
 285         *ret_handle = (void*)length;
 286         return ptr;
 287 }
 288
 289 /**
 290  * mono_file_unmap:
 291  * @addr: memory address returned by mono_file_map ()
 292  * @handle: handle of memory map
 293  *
 294  * Remove the memory mapping at the address @addr.
 295  * @handle must be the value returned in ret_handle by mono_file_map ().
 296  *
 297  * Returns: 0 on success.
 298  */
 299 int
 300 mono_file_unmap (void *addr, void *handle)
 301 {
 302         int res;
 303
 304         BEGIN_CRITICAL_SECTION;
 305         res = munmap (addr, (size_t)handle);
 306         END_CRITICAL_SECTION;
 307
 308         return res;
 309 }
 310
 311 /**
 312  * mono_mprotect:
 313  * @addr: memory address
 314  * @length: size of memory area
 315  * @flags: new protection flags
 316  *
 317  * Change the protection for the memory area at @addr for @length bytes
 318  * to matche the supplied @flags.
 319  * If @flags includes MON_MMAP_DISCARD the pages are discarded from memory
 320  * and the area is cleared to zero.
 321  * @addr must be aligned to the page size.
 322  * @length must be a multiple of the page size.
 323  *
 324  * Returns: 0 on success.
 325  */
 326 #if defined(__native_client__)
 327 int
 328 mono_mprotect (void *addr, size_t length, int flags)
 329 {
 330         int prot = prot_from_flags (flags);
 331         void *new_addr;
 332
 333         if (flags & MONO_MMAP_DISCARD) memset (addr, 0, length);
 334
 335         new_addr = mmap(addr, length, prot, MAP_PRIVATE | MAP_FIXED | MAP_ANONYMOUS, -1, 0);
 336         if (new_addr == addr) return 0;
 337         return -1;
 338 }
 339 #else
 340 int
 341 mono_mprotect (void *addr, size_t length, int flags)
 342 {
 343         int prot = prot_from_flags (flags);
 344
 345         if (flags & MONO_MMAP_DISCARD) {
 346                 /* on non-linux the pages are not guaranteed to be zeroed (*bsd, osx at least) */
 347 #ifdef __linux__
 348                 if (madvise (addr, length, MADV_DONTNEED))
 349                         memset (addr, 0, length);
 350 #else
 351                 memset (addr, 0, length);
 352 #ifdef HAVE_MADVISE
 353                 madvise (addr, length, MADV_DONTNEED);
 354                 madvise (addr, length, MADV_FREE);
 355 #else
 356                 posix_madvise (addr, length, POSIX_MADV_DONTNEED);
 357 #endif
 358 #endif
 359         }
 360         return mprotect (addr, length, prot);
 361 }
 362 #endif // __native_client__
 363
 364 #else
 365
 366 /* dummy malloc-based implementation */
 367 int
 368 mono_pagesize (void)
 369 {
 370         return 4096;
 371 }
 372
 373 int
 374 mono_valloc_granule (void)
 375 {
 376         return mono_pagesize ();
 377 }
 378
 379 void*
 380 mono_valloc (void *addr, size_t length, int flags, MonoMemAccountType type)
 381 {
 382         return g_malloc (length);
 383 }
 384
 385 void*
 386 mono_valloc_aligned (size_t size, size_t alignment, int flags, MonoMemAccountType type)
 387 {
 388         g_assert_not_reached ();
 389 }
 390
 391 #define HAVE_VALLOC_ALIGNED
 392
 393 int
 394 mono_vfree (void *addr, size_t length, MonoMemAccountType type)
 395 {
 396         g_free (addr);
 397         return 0;
 398 }
 399
 400 int
 401 mono_mprotect (void *addr, size_t length, int flags)
 402 {
 403         if (flags & MONO_MMAP_DISCARD) {
 404                 memset (addr, 0, length);
 405         }
 406         return 0;
 407 }
 408
 409 #endif // HAVE_MMAP
 410
 411 #if defined(HAVE_SHM_OPEN) && !defined (DISABLE_SHARED_PERFCOUNTERS)
 412
 413 static int use_shared_area;
 414
 415 static gboolean
 416 shared_area_disabled (void)
 417 {
 418         if (!use_shared_area) {
 419                 if (g_getenv ("MONO_DISABLE_SHARED_AREA"))
 420                         use_shared_area = -1;
 421                 else
 422                         use_shared_area = 1;
 423         }
 424         return use_shared_area == -1;
 425 }
 426
 427 static int
 428 mono_shared_area_instances_slow (void **array, int count, gboolean cleanup)
 429 {
 430         int i, j = 0;
 431         int num;
 432         void *data;
 433         gpointer *processes = mono_process_list (&num);
 434         for (i = 0; i < num; ++i) {
 435                 data = mono_shared_area_for_pid (processes [i]);
 436                 if (!data)
 437                         continue;
 438                 mono_shared_area_unload (data);
 439                 if (!cleanup) {
 440                         if (j < count)
 441                                 array [j++] = processes [i];
 442                         else
 443                                 break;
 444                 }
 445         }
 446         g_free (processes);
 447         return j;
 448 }
 449
 450 static int
 451 mono_shared_area_instances_helper (void **array, int count, gboolean cleanup)
 452 {
 453         const char *name;
 454         int i = 0;
 455         int curpid = getpid ();
 456         GDir *dir = g_dir_open ("/dev/shm/", 0, NULL);
 457         if (!dir)
 458                 return mono_shared_area_instances_slow (array, count, cleanup);
 459         while ((name = g_dir_read_name (dir))) {
 460                 int pid;
 461                 char *nend;
 462                 if (strncmp (name, "mono.", 5))
 463                         continue;
 464                 pid = strtol (name + 5, &nend, 10);
 465                 if (pid <= 0 || nend == name + 5 || *nend)
 466                         continue;
 467                 if (!cleanup) {
 468                         if (i < count)
 469                                 array [i++] = GINT_TO_POINTER (pid);
 470                         else
 471                                 break;
 472                 }
 473                 if (curpid != pid && kill (pid, 0) == -1 && (errno == ESRCH || errno == ENOMEM)) {
 474                         char buf [128];
 475                         g_snprintf (buf, sizeof (buf), "/mono.%d", pid);
 476                         shm_unlink (buf);
 477                 }
 478         }
 479         g_dir_close (dir);
 480         return i;
 481 }
 482
 483 void*
 484 mono_shared_area (void)
 485 {
 486         int fd;
 487         int pid = getpid ();
 488         /* we should allow the user to configure the size */
 489         int size = mono_pagesize ();
 490         char buf [128];
 491         void *res;
 492         SAreaHeader *header;
 493
 494         if (shared_area_disabled ()) {
 495                 if (!malloced_shared_area)
 496                         malloced_shared_area = malloc_shared_area (0);
 497                 /* get the pid here */
 498                 return malloced_shared_area;
 499         }
 500
 501         /* perform cleanup of segments left over from dead processes */
 502         mono_shared_area_instances_helper (NULL, 0, TRUE);
 503
 504         g_snprintf (buf, sizeof (buf), "/mono.%d", pid);
 505
 506         fd = shm_open (buf, O_CREAT|O_EXCL|O_RDWR, S_IRUSR|S_IWUSR|S_IRGRP);
 507         if (fd == -1 && errno == EEXIST) {
 508                 /* leftover */
 509                 shm_unlink (buf);
 510                 fd = shm_open (buf, O_CREAT|O_EXCL|O_RDWR, S_IRUSR|S_IWUSR|S_IRGRP);
 511         }
 512         /* in case of failure we try to return a memory area anyway,
 513          * even if it means the data can't be read by other processes
 514          */
 515         if (fd == -1)
 516                 return malloc_shared_area (pid);
 517         if (ftruncate (fd, size) != 0) {
 518                 shm_unlink (buf);
 519                 close (fd);
 520         }
 521         BEGIN_CRITICAL_SECTION;
 522         res = mmap (NULL, size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 523         END_CRITICAL_SECTION;
 524
 525         if (res == MAP_FAILED) {
 526                 shm_unlink (buf);
 527                 close (fd);
 528                 return malloc_shared_area (pid);
 529         }
 530         /* we don't need the file descriptor anymore */
 531         close (fd);
 532         header = (SAreaHeader *) res;
 533         header->size = size;
 534         header->pid = pid;
 535         header->stats_start = sizeof (SAreaHeader);
 536         header->stats_end = sizeof (SAreaHeader);
 537
 538         mono_atexit (mono_shared_area_remove);
 539         return res;
 540 }
 541
 542 void
 543 mono_shared_area_remove (void)
 544 {
 545         char buf [128];
 546
 547         if (shared_area_disabled ()) {
 548                 if (malloced_shared_area)
 549                         g_free (malloced_shared_area);
 550                 return;
 551         }
 552
 553         g_snprintf (buf, sizeof (buf), "/mono.%d", getpid ());
 554         shm_unlink (buf);
 555         if (malloced_shared_area)
 556                 g_free (malloced_shared_area);
 557 }
 558
 559 void*
 560 mono_shared_area_for_pid (void *pid)
 561 {
 562         int fd;
 563         /* we should allow the user to configure the size */
 564         int size = mono_pagesize ();
 565         char buf [128];
 566         void *res;
 567
 568         if (shared_area_disabled ())
 569                 return NULL;
 570
 571         g_snprintf (buf, sizeof (buf), "/mono.%d", GPOINTER_TO_INT (pid));
 572
 573         fd = shm_open (buf, O_RDONLY, S_IRUSR|S_IRGRP);
 574         if (fd == -1)
 575                 return NULL;
 576         BEGIN_CRITICAL_SECTION;
 577         res = mmap (NULL, size, PROT_READ, MAP_SHARED, fd, 0);
 578         END_CRITICAL_SECTION;
 579
 580         if (res == MAP_FAILED) {
 581                 close (fd);
 582                 return NULL;
 583         }
 584         /* FIXME: validate the area */
 585         /* we don't need the file descriptor anymore */
 586         close (fd);
 587         return res;
 588 }
 589
 590 void
 591 mono_shared_area_unload  (void *area)
 592 {
 593         /* FIXME: currently we load only a page */
 594         BEGIN_CRITICAL_SECTION;
 595         munmap (area, mono_pagesize ());
 596         END_CRITICAL_SECTION;
 597 }
 598
 599 int
 600 mono_shared_area_instances (void **array, int count)
 601 {
 602         return mono_shared_area_instances_helper (array, count, FALSE);
 603 }
 604 #else
 605 void*
 606 mono_shared_area (void)
 607 {
 608         if (!malloced_shared_area)
 609                 malloced_shared_area = malloc_shared_area (getpid ());
 610         /* get the pid here */
 611         return malloced_shared_area;
 612 }
 613
 614 void
 615 mono_shared_area_remove (void)
 616 {
 617         if (malloced_shared_area)
 618                 g_free (malloced_shared_area);
 619         malloced_shared_area = NULL;
 620 }
 621
 622 void*
 623 mono_shared_area_for_pid (void *pid)
 624 {
 625         return NULL;
 626 }
 627
 628 void
 629 mono_shared_area_unload (void *area)
 630 {
 631 }
 632
 633 int
 634 mono_shared_area_instances (void **array, int count)
 635 {
 636         return 0;
 637 }
 638
 639 #endif // HAVE_SHM_OPEN
 640
 641 #endif // HOST_WIN32
 642
 643 #ifndef HAVE_VALLOC_ALIGNED
 644 void*
 645 mono_valloc_aligned (size_t size, size_t alignment, int flags, MonoMemAccountType type)
 646 {
 647         /* Allocate twice the memory to be able to put the block on an aligned address */
 648         char *mem = (char *) mono_valloc (NULL, size + alignment, flags, type);
 649         char *aligned;
 650
 651         if (!mem)
 652                 return NULL;
 653
 654         aligned = aligned_address (mem, size, alignment);
 655
 656         if (aligned > mem)
 657                 mono_vfree (mem, aligned - mem, type);
 658         if (aligned + size < mem + size + alignment)
 659                 mono_vfree (aligned + size, (mem + size + alignment) - (aligned + size), type);
 660
 661         return aligned;
 662 }
 663 #endif
 664
 665 int
 666 mono_pages_not_faulted (void *addr, size_t size)
 667 {
 668 #ifdef HAVE_MINCORE
 669         int i;
 670         gint64 count;
 671         int pagesize = mono_pagesize ();
 672         int npages = (size + pagesize - 1) / pagesize;
 673         char *faulted = (char *) g_malloc0 (sizeof (char*) * npages);
 674
 675         /*
 676          * We cast `faulted` to void* because Linux wants an unsigned
 677          * char* while BSD wants a char*.
 678          */
 679 #ifdef __linux__
 680         if (mincore (addr, size, (unsigned char *)faulted) != 0) {
 681 #else
 682         if (mincore (addr, size, (char *)faulted) != 0) {
 683 #endif
 684                 count = -1;
 685         } else {
 686                 count = 0;
 687                 for (i = 0; i < npages; ++i) {
 688                         if (faulted [i] != 0)
 689                                 ++count;
 690                 }
 691         }
 692
 693         g_free (faulted);
 694
 695         return count;
 696 #else
 697         return -1;
 698 #endif
 699 }