emacs/src/w32heap.c

/* Heap management routines for GNU Emacs on the Microsoft Windows API.
   Copyright (C) 1994, 2001-2015 Free Software Foundation, Inc.

   This file is part of GNU Emacs.

   GNU Emacs is free software: you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
   the Free Software Foundation, either version 3 of the License, or
   (at your option) any later version.

   GNU Emacs is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   GNU General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with GNU Emacs.  If not, see <http://www.gnu.org/licenses/>. */

/*
  Geoff Voelker (voelker@cs.washington.edu)                          7-29-94
*/

/*
  Heavily modified by Fabrice Popineau (fabrice.popineau@gmail.com) 28-02-2014
*/

/*
  Memory allocation scheme for w32/w64:

  - Buffers are mmap'ed using a very simple emulation of mmap/munmap
  - During the temacs phase:
    * we use a private heap declared to be stored into the `dumped_data'
    * unfortunately, this heap cannot be made growable, so the size of
      blocks it can allocate is limited to (0x80000 - pagesize)
    * the blocks that are larger than this are allocated from the end
      of the `dumped_data' array; there are not so many of them.
      We use a very simple first-fit scheme to reuse those blocks.
    * we check that the private heap does not cross the area used
      by the bigger chunks.
  - During the emacs phase:
    * we create a private heap for new memory blocks
    * we make sure that we never free a block that has been dumped.
      Freeing a dumped block could work in principle, but may prove
      unreliable if we distribute binaries of emacs.exe: MS does not
      guarantee that the heap data structures are the same across all
      versions of their OS, even though the API is available since XP.  */

#include <config.h>
#include <stdio.h>
#include <errno.h>

#include <sys/mman.h>
#include "w32common.h"
#include "w32heap.h"
#include "lisp.h"  /* for VALMASK */

/* We chose to leave those declarations here.  They are used only in
   this file.  The RtlCreateHeap is available since XP.  It is located
   in ntdll.dll and is available with the DDK.  People often
   complained that HeapCreate doesn't offer the ability to create a
   heap at a given place, which we need here, and which RtlCreateHeap
   provides.  We reproduce here the definitions available with the
   DDK.  */

typedef PVOID (WINAPI * RtlCreateHeap_Proc) (
                                             /* _In_ */      ULONG Flags,
                                             /* _In_opt_ */  PVOID HeapBase,
                                             /* _In_opt_ */  SIZE_T ReserveSize,
                                             /* _In_opt_ */  SIZE_T CommitSize,
                                             /* _In_opt_ */  PVOID Lock,
                                             /* _In_opt_ */  PVOID Parameters
                                             );

typedef LONG NTSTATUS;

typedef NTSTATUS
(NTAPI * PRTL_HEAP_COMMIT_ROUTINE)(
                                   IN PVOID Base,
                                   IN OUT PVOID *CommitAddress,
                                   IN OUT PSIZE_T CommitSize
                                   );

typedef struct _RTL_HEAP_PARAMETERS {
  ULONG Length;
  SIZE_T SegmentReserve;
  SIZE_T SegmentCommit;
  SIZE_T DeCommitFreeBlockThreshold;
  SIZE_T DeCommitTotalFreeThreshold;
  SIZE_T MaximumAllocationSize;
  SIZE_T VirtualMemoryThreshold;
  SIZE_T InitialCommit;
  SIZE_T InitialReserve;
  PRTL_HEAP_COMMIT_ROUTINE CommitRoutine;
  SIZE_T Reserved[ 2 ];
} RTL_HEAP_PARAMETERS, *PRTL_HEAP_PARAMETERS;

/* We reserve space for dumping emacs lisp byte-code inside a static
   array.  By storing it in an array, the generic mechanism in
   unexecw32.c will be able to dump it without the need to add a
   special segment to the executable.  In order to be able to do this
   without losing too much space, we need to create a Windows heap at
   the specific address of the static array.  The RtlCreateHeap
   available inside the NT kernel since XP will do this.  It allows to
   create a non-growable heap at a specific address.  So before
   dumping, we create a non-growable heap at the address of the
   dumped_data[] array.  After dumping, we reuse memory allocated
   there without being able to free it (but most of it is not meant to
   be freed anyway), and we use a new private heap for all new
   allocations.  */

/* FIXME: Most of the space reserved for dumped_data[] is only used by
   the 1st bootstrap-emacs.exe built while bootstrapping.  Once the
   preloaded Lisp files are byte-compiled, the next loadup uses less
   than half of the size stated below.  It would be nice to find a way
   to build only the first bootstrap-emacs.exe with the large size,
   and reset that to a lower value afterwards.  */
#if defined _WIN64 || defined WIDE_EMACS_INT
# define DUMPED_HEAP_SIZE (18*1024*1024)
#else
# define DUMPED_HEAP_SIZE (11*1024*1024)
#endif

static unsigned char dumped_data[DUMPED_HEAP_SIZE];

/* Info for keeping track of our dynamic heap used after dumping. */
unsigned char *data_region_base = NULL;
unsigned char *data_region_end = NULL;
static DWORD_PTR committed = 0;

/* The maximum block size that can be handled by a non-growable w32
   heap is limited by the MaxBlockSize value below.

   This point deserves and explanation.

   The W32 heap allocator can be used for a growable
   heap or a non-growable one.

   A growable heap is not compatible with a fixed base address for the
   heap.  Only a non-growable one is.  One drawback of non-growable
   heaps is that they can hold only objects smaller than a certain
   size (the one defined below).  Most of the largest blocks are GC'ed
   before dumping.  In any case and to be safe, we implement a simple
   first-fit allocation algorithm starting at the end of the
   dumped_data[] array like depicted below:

  ----------------------------------------------
  |               |              |             |
  | Private heap  |->          <-|  Big chunks |
  |               |              |             |
  ----------------------------------------------
  ^               ^              ^
  dumped_data     dumped_data    bc_limit
                  + committed

*/

/* Info for managing our preload heap, which is essentially a fixed size
   data area in the executable. */
#define PAGE_SIZE 0x1000
#define MaxBlockSize (0x80000 - PAGE_SIZE)

#define MAX_BLOCKS 0x40

static struct
{
  unsigned char *address;
  size_t size;
  DWORD occupied;
} blocks[MAX_BLOCKS];

static DWORD          blocks_number = 0;
static unsigned char *bc_limit;

/* Handle for the private heap:
    - inside the dumped_data[] array before dump,
    - outside of it after dump.
*/
HANDLE heap = NULL;

/* We redirect the standard allocation functions.  */
malloc_fn the_malloc_fn;
realloc_fn the_realloc_fn;
free_fn the_free_fn;

/* It doesn't seem to be useful to allocate from a file mapping.
   It would be if the memory was shared.
     http://stackoverflow.com/questions/307060/what-is-the-purpose-of-allocating-pages-in-the-pagefile-with-createfilemapping  */

/* This is the function to commit memory when the heap allocator
   claims for new memory.  Before dumping, we allocate space
   from the fixed size dumped_data[] array.
*/
NTSTATUS NTAPI
dumped_data_commit (PVOID Base, PVOID *CommitAddress, PSIZE_T CommitSize)
{
  /* This is used before dumping.

     The private heap is stored at dumped_data[] address.
     We commit contiguous areas of the dumped_data array
     as requests arrive.  */
  *CommitAddress = data_region_base + committed;
  committed += *CommitSize;
  /* Check that the private heap area does not overlap the big chunks area.  */
  if (((unsigned char *)(*CommitAddress)) + *CommitSize >= bc_limit)
    {
      fprintf (stderr,
	       "dumped_data_commit: memory exhausted.\nEnlarge dumped_data[]!\n");
      exit (-1);
    }
  return 0;
}

/* Heap creation.  */

/* We want to turn on Low Fragmentation Heap for XP and older systems.
   MinGW32 lacks those definitions.  */
#ifndef MINGW_W64
typedef enum _HEAP_INFORMATION_CLASS {
  HeapCompatibilityInformation
} HEAP_INFORMATION_CLASS;

typedef WINBASEAPI BOOL (WINAPI * HeapSetInformation_Proc)(HANDLE,HEAP_INFORMATION_CLASS,PVOID,SIZE_T);
#endif

void
init_heap (void)
{
  if (using_dynamic_heap)
    {
      unsigned long enable_lfh = 2;

      /* After dumping, use a new private heap.  We explicitly enable
         the low fragmentation heap (LFH) here, for the sake of pre
         Vista versions.  Note: this will harmlessly fail on Vista and
         later, where the low-fragmentation heap is enabled by
         default.  It will also fail on pre-Vista versions when Emacs
         is run under a debugger; set _NO_DEBUG_HEAP=1 in the
         environment before starting GDB to get low fragmentation heap
         on XP and older systems, for the price of losing "certain
         heap debug options"; for the details see
         http://msdn.microsoft.com/en-us/library/windows/desktop/aa366705%28v=vs.85%29.aspx.  */
      data_region_end = data_region_base;

      /* Create the private heap.  */
      heap = HeapCreate (0, 0, 0);

#ifndef MINGW_W64
      /* Set the low-fragmentation heap for OS before Vista.  */
      HMODULE hm_kernel32dll = LoadLibrary ("kernel32.dll");
      HeapSetInformation_Proc s_pfn_Heap_Set_Information = (HeapSetInformation_Proc) GetProcAddress (hm_kernel32dll, "HeapSetInformation");
      if (s_pfn_Heap_Set_Information != NULL)
	{
	  if (s_pfn_Heap_Set_Information ((PVOID) heap,
					  HeapCompatibilityInformation,
					  &enable_lfh, sizeof(enable_lfh)) == 0)
	    DebPrint (("Enabling Low Fragmentation Heap failed: error %ld\n",
		       GetLastError ()));
	}
#endif

      the_malloc_fn = malloc_after_dump;
      the_realloc_fn = realloc_after_dump;
      the_free_fn = free_after_dump;
    }
  else
    {
      /* Find the RtlCreateHeap function.  Headers for this function
         are provided with the w32 ddk, but the function is available
         in ntdll.dll since XP.  */
      HMODULE hm_ntdll = LoadLibrary ("ntdll.dll");
      RtlCreateHeap_Proc s_pfn_Rtl_Create_Heap
	= (RtlCreateHeap_Proc) GetProcAddress (hm_ntdll, "RtlCreateHeap");
      /* Specific parameters for the private heap.  */
      RTL_HEAP_PARAMETERS params;
      ZeroMemory (&params, sizeof(params));
      params.Length = sizeof(RTL_HEAP_PARAMETERS);

      data_region_base = (unsigned char *)ROUND_UP (dumped_data, 0x1000);
      data_region_end = bc_limit = dumped_data + DUMPED_HEAP_SIZE;

      params.InitialCommit = committed = 0x1000;
      params.InitialReserve = sizeof(dumped_data);
      /* Use our own routine to commit memory from the dumped_data
         array.  */
      params.CommitRoutine = &dumped_data_commit;

      /* Create the private heap.  */
      if (s_pfn_Rtl_Create_Heap == NULL)
	{
	  fprintf (stderr, "Cannot build Emacs without RtlCreateHeap being available; exiting.\n");
	  exit (-1);
	}
      heap = s_pfn_Rtl_Create_Heap (0, data_region_base, 0, 0, NULL, &params);
      the_malloc_fn = malloc_before_dump;
      the_realloc_fn = realloc_before_dump;
      the_free_fn = free_before_dump;
    }

  /* Update system version information to match current system.  */
  cache_system_info ();
}

#undef malloc
#undef realloc
#undef free

/* FREEABLE_P checks if the block can be safely freed.  */
#define FREEABLE_P(addr)                                        \
    ((unsigned char *)(addr) < dumped_data                      \
     || (unsigned char *)(addr) >= dumped_data + DUMPED_HEAP_SIZE)

void *
malloc_after_dump (size_t size)
{
  /* Use the new private heap.  */
  void *p = HeapAlloc (heap, 0, size);

  /* After dump, keep track of the "brk value" for sbrk(0).  */
  if (p)
    {
      unsigned char *new_brk = (unsigned char *)p + size;

      if (new_brk > data_region_end)
	data_region_end = new_brk;
    }
  else
    errno = ENOMEM;
  return p;
}

void *
malloc_before_dump (size_t size)
{
  void *p;

  /* Before dumping.  The private heap can handle only requests for
     less than MaxBlockSize.  */
  if (size < MaxBlockSize)
    {
      /* Use the private heap if possible.  */
      p = HeapAlloc (heap, 0, size);
      if (!p)
	errno = ENOMEM;
    }
  else
    {
      /* Find the first big chunk that can hold the requested size.  */
      int i = 0;

      for (i = 0; i < blocks_number; i++)
	{
	  if (blocks[i].occupied == 0 && blocks[i].size >= size)
	    break;
	}
      if (i < blocks_number)
	{
	  /* If found, use it.  */
	  p = blocks[i].address;
	  blocks[i].occupied = TRUE;
	}
      else
	{
	  /* Allocate a new big chunk from the end of the dumped_data
	     array.  */
	  if (blocks_number >= MAX_BLOCKS)
	    {
	      fprintf (stderr,
		       "malloc_before_dump: no more big chunks available.\nEnlarge MAX_BLOCKS!\n");
	      exit (-1);
	    }
	  bc_limit -= size;
	  bc_limit = (unsigned char *)ROUND_DOWN (bc_limit, 0x10);
	  p = bc_limit;
	  blocks[blocks_number].address = p;
	  blocks[blocks_number].size = size;
	  blocks[blocks_number].occupied = TRUE;
	  blocks_number++;
	  /* Check that areas do not overlap.  */
	  if (bc_limit < dumped_data + committed)
	    {
	      fprintf (stderr,
		       "malloc_before_dump: memory exhausted.\nEnlarge dumped_data[]!\n");
	      exit (-1);
	    }
	}
    }
  return p;
}

/* Re-allocate the previously allocated block in ptr, making the new
   block SIZE bytes long.  */
void *
realloc_after_dump (void *ptr, size_t size)
{
  void *p;

  /* After dumping.  */
  if (FREEABLE_P (ptr))
    {
      /* Reallocate the block since it lies in the new heap.  */
      p = HeapReAlloc (heap, 0, ptr, size);
      if (!p)
	errno = ENOMEM;
    }
  else
    {
      /* If the block lies in the dumped data, do not free it.  Only
         allocate a new one.  */
      p = HeapAlloc (heap, 0, size);
      if (p)
	CopyMemory (p, ptr, size);
      else
	errno = ENOMEM;
    }
  /* After dump, keep track of the "brk value" for sbrk(0).  */
  if (p)
    {
      unsigned char *new_brk = (unsigned char *)p + size;

      if (new_brk > data_region_end)
	data_region_end = new_brk;
    }
  return p;
}

void *
realloc_before_dump (void *ptr, size_t size)
{
  void *p;

  /* Before dumping.  */
  if (dumped_data < (unsigned char *)ptr
      && (unsigned char *)ptr < bc_limit && size <= MaxBlockSize)
    {
      p = HeapReAlloc (heap, 0, ptr, size);
      if (!p)
	errno = ENOMEM;
    }
  else
    {
      /* In this case, either the new block is too large for the heap,
         or the old block was already too large.  In both cases,
         malloc_before_dump() and free_before_dump() will take care of
         reallocation.  */
      p = malloc_before_dump (size);
      /* If SIZE is below MaxBlockSize, malloc_before_dump will try to
	 allocate it in the fixed heap.  If that fails, we could have
	 kept the block in its original place, above bc_limit, instead
	 of failing the call as below.  But this doesn't seem to be
	 worth the added complexity, as loadup allocates only a very
	 small number of large blocks, and never reallocates them.  */
      if (p)
	{
	  CopyMemory (p, ptr, size);
	  free_before_dump (ptr);
	}
    }
  return p;
}

/* Free a block allocated by `malloc', `realloc' or `calloc'.  */
void
free_after_dump (void *ptr)
{
  /* After dumping.  */
  if (FREEABLE_P (ptr))
    {
      /* Free the block if it is in the new private heap.  */
      HeapFree (heap, 0, ptr);
    }
}

void
free_before_dump (void *ptr)
{
  /* Before dumping.  */
  if (dumped_data < (unsigned char *)ptr
      && (unsigned char *)ptr < bc_limit)
    {
      /* Free the block if it is allocated in the private heap.  */
      HeapFree (heap, 0, ptr);
    }
  else
    {
      /* Look for the big chunk.  */
      int i;

      for (i = 0; i < blocks_number; i++)
	{
	  if (blocks[i].address == ptr)
	    {
	      /* Reset block occupation if found.  */
	      blocks[i].occupied = 0;
	      break;
	    }
	  /* What if the block is not found?  We should trigger an
	     error here.  */
	  eassert (i < blocks_number);
	}
    }
}

#ifdef ENABLE_CHECKING
void
report_temacs_memory_usage (void)
{
  DWORD blocks_used = 0;
  size_t large_mem_used = 0;
  int i;

  for (i = 0; i < blocks_number; i++)
    if (blocks[i].occupied)
      {
	blocks_used++;
	large_mem_used += blocks[i].size;
      }

  /* Emulate 'message', which writes to stderr in non-interactive
     sessions.  */
  fprintf (stderr,
	   "Dump memory usage: Heap: %" PRIu64 "  Large blocks(%lu/%lu): %" PRIu64 "/%" PRIu64 "\n",
	   (unsigned long long)committed, blocks_used, blocks_number,
	   (unsigned long long)large_mem_used,
	   (unsigned long long)(dumped_data + DUMPED_HEAP_SIZE - bc_limit));
}
#endif

/* Emulate getpagesize. */
int
getpagesize (void)
{
  return sysinfo_cache.dwPageSize;
}

void *
sbrk (ptrdiff_t increment)
{
  /* data_region_end is the address beyond the last allocated byte.
     The sbrk() function is not emulated at all, except for a 0 value
     of its parameter.  This is needed by the Emacs Lisp function
     `memory-limit'.  */
  eassert (increment == 0);
  return data_region_end;
}

#define MAX_BUFFER_SIZE (512 * 1024 * 1024)

/* MMAP allocation for buffers.  */
void *
mmap_alloc (void **var, size_t nbytes)
{
  void *p = NULL;

  /* We implement amortized allocation.  We start by reserving twice
     the size requested and commit only the size requested.  Then
     realloc could proceed and use the reserved pages, reallocating
     only if needed.  Buffer shrink would happen only so that we stay
     in the 2x range.  This is a big win when visiting compressed
     files, where the final size of the buffer is not known in
     advance, and the buffer is enlarged several times as the data is
     decompressed on the fly.  */
  if (nbytes < MAX_BUFFER_SIZE)
    p = VirtualAlloc (NULL, (nbytes * 2), MEM_RESERVE, PAGE_READWRITE);

  /* If it fails, or if the request is above 512MB, try with the
     requested size.  */
  if (p == NULL)
    p = VirtualAlloc (NULL, nbytes, MEM_RESERVE, PAGE_READWRITE);

  if (p != NULL)
    {
      /* Now, commit pages for NBYTES.  */
      *var = VirtualAlloc (p, nbytes, MEM_COMMIT, PAGE_READWRITE);
    }

  if (!p)
    {
      if (GetLastError () == ERROR_NOT_ENOUGH_MEMORY)
	errno = ENOMEM;
      else
	{
	  DebPrint (("mmap_alloc: error %ld\n", GetLastError ()));
	  errno = EINVAL;
	}
    }

  return *var = p;
}

void
mmap_free (void **var)
{
  if (*var)
    {
      if (VirtualFree (*var, 0, MEM_RELEASE) == 0)
        DebPrint (("mmap_free: error %ld\n", GetLastError ()));
      *var = NULL;
    }
}

void *
mmap_realloc (void **var, size_t nbytes)
{
  MEMORY_BASIC_INFORMATION memInfo, m2;

  if (*var == NULL)
    return mmap_alloc (var, nbytes);

  /* This case happens in init_buffer().  */
  if (nbytes == 0)
    {
      mmap_free (var);
      return mmap_alloc (var, nbytes);
    }

  if (VirtualQuery (*var, &memInfo, sizeof (memInfo)) == 0)
    DebPrint (("mmap_realloc: VirtualQuery error = %ld\n", GetLastError ()));

  /* We need to enlarge the block.  */
  if (memInfo.RegionSize < nbytes)
    {
      if (VirtualQuery (*var + memInfo.RegionSize, &m2, sizeof(m2)) == 0)
        DebPrint (("mmap_realloc: VirtualQuery error = %ld\n",
		   GetLastError ()));
      /* If there is enough room in the current reserved area, then
	 commit more pages as needed.  */
      if (m2.State == MEM_RESERVE
	  && nbytes <= memInfo.RegionSize + m2.RegionSize)
	{
	  void *p;

	  p = VirtualAlloc (*var + memInfo.RegionSize,
			    nbytes - memInfo.RegionSize,
			    MEM_COMMIT, PAGE_READWRITE);
	  if (!p /* && GetLastError() != ERROR_NOT_ENOUGH_MEMORY */)
	    {
	      DebPrint (("realloc enlarge: VirtualAlloc error %ld\n",
			 GetLastError ()));
	      errno = ENOMEM;
	    }
	  return *var;
	}
      else
	{
	  /* Else we must actually enlarge the block by allocating a
	     new one and copying previous contents from the old to the
	     new one.  */
	  void *old_ptr = *var;

	  if (mmap_alloc (var, nbytes))
	    {
	      CopyMemory (*var, old_ptr, memInfo.RegionSize);
	      mmap_free (&old_ptr);
	      return *var;
	    }
	  else
	    {
	      /* We failed to enlarge the buffer.  */
	      *var = old_ptr;
	      return NULL;
	    }
	}
    }

  /* If we are shrinking by more than one page...  */
  if (memInfo.RegionSize  > nbytes + getpagesize())
    {
      /* If we are shrinking a lot...  */
      if ((memInfo.RegionSize / 2) > nbytes)
        {
          /* Let's give some memory back to the system and release
	     some pages.  */
          void *old_ptr = *var;

	  if (mmap_alloc (var, nbytes))
            {
              CopyMemory (*var, old_ptr, nbytes);
              mmap_free (&old_ptr);
              return *var;
            }
          else
	    {
	      /* In case we fail to shrink, try to go on with the old block.
		 But that means there is a lot of memory pressure.
		 We could also decommit pages.  */
	      *var = old_ptr;
	      return *var;
	    }
        }

      /* We still can decommit pages.  */
      if (VirtualFree (*var + nbytes + get_page_size(),
		       memInfo.RegionSize - nbytes - get_page_size(),
		       MEM_DECOMMIT) == 0)
        DebPrint (("mmap_realloc: VirtualFree error %ld\n", GetLastError ()));
      return *var;
    }

  /* Not enlarging, not shrinking by more than one page.  */
  return *var;
}