MBlock.c 9.78 KB
Newer Older
1
/* -----------------------------------------------------------------------------
2
 * $Id: MBlock.c,v 1.36 2002/11/05 09:31:37 simonmar Exp $
3
4
 *
 * (c) The GHC Team 1998-1999
5
6
7
8
9
10
11
 *
 * MegaBlock Allocator Interface.  This file contains all the dirty
 * architecture-dependent hackery required to get a chunk of aligned
 * memory from the operating system.
 *
 * ---------------------------------------------------------------------------*/

12
13
/* This is non-posix compliant. */
/* #include "PosixSource.h" */
14
15
16
17
18
19
20

#include "Rts.h"
#include "RtsUtils.h"
#include "RtsFlags.h"
#include "MBlock.h"
#include "BlockAlloc.h"

21
22
23
#ifdef HAVE_STDLIB_H
#include <stdlib.h>
#endif
24
25
26
27
28
29
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif
#ifdef HAVE_SYS_TYPES_H
#include <sys/types.h>
#endif
sof's avatar
sof committed
30
31
32
33
#ifndef mingw32_TARGET_OS
# ifdef HAVE_SYS_MMAN_H
# include <sys/mman.h>
# endif
34
35
36
37
#endif
#ifdef HAVE_FCNTL_H
#include <fcntl.h>
#endif
sof's avatar
sof committed
38
#if HAVE_WINDOWS_H
39
40
#include <windows.h>
#endif
41
42
43
#if darwin_TARGET_OS
#include <mach/vm_map.h>
#endif
44

45
46
#include <errno.h>

47
48
lnat mblocks_allocated = 0;

49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
/* -----------------------------------------------------------------------------
   The MBlock Map: provides our implementation of HEAP_ALLOCED()
   -------------------------------------------------------------------------- */

StgWord8 mblock_map[4096]; // initially all zeros

static void
mblockIsHeap (void *p)
{
    mblock_map[((StgWord)p & ~MBLOCK_MASK) >> MBLOCK_SHIFT] = 1;
}

/* -----------------------------------------------------------------------------
   Allocate new mblock(s)
   -------------------------------------------------------------------------- */

65
66
67
68
69
70
void *
getMBlock(void)
{
  return getMBlocks(1);
}

71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
/* -----------------------------------------------------------------------------
   The mmap() method

   On Unix-like systems, we use mmap() to allocate our memory.  We
   want memory in chunks of MBLOCK_SIZE, and aligned on an MBLOCK_SIZE
   boundary.  The mmap() interface doesn't give us this level of
   control, so we have to use some heuristics.

   In the general case, if we want a block of n megablocks, then we
   allocate n+1 and trim off the slop from either side (using
   munmap()) to get an aligned chunk of size n.  However, the next
   time we'll try to allocate directly after the previously allocated
   chunk, on the grounds that this is aligned and likely to be free.
   If it turns out that we were wrong, we have to munmap() and try
   again using the general method.
   -------------------------------------------------------------------------- */

sof's avatar
sof committed
88
#if !defined(mingw32_TARGET_OS) && !defined(cygwin32_TARGET_OS)
89
90
91
92
93
94
95
96
97
98
99
100

// A wrapper around mmap(), to abstract away from OS differences in
// the mmap() interface.

static void *
my_mmap (void *addr, int size)
{
    void *ret;

#ifdef solaris2_TARGET_OS
    { 
	int fd = open("/dev/zero",O_RDONLY);
101
	ret = mmap(addr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
102
103
104
105
106
107
	close(fd);
    }
#elif hpux_TARGET_OS
    ret = mmap(addr, size, PROT_READ | PROT_WRITE, 
	       MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
#elif darwin_TARGET_OS
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
    // Without MAP_FIXED, Apple's mmap ignores addr.
    // With MAP_FIXED, it overwrites already mapped regions, whic
    // mmap(0, ... MAP_FIXED ...) is worst of all: It unmaps the program text
    // and replaces it with zeroes, causing instant death.
    // This behaviour seems to be conformant with IEEE Std 1003.1-2001.
    // Let's just use the underlying Mach Microkernel calls directly,
    // they're much nicer.
    
    kern_return_t err;
    ret = addr;
    if(addr)	// try to allocate at adress
	err = vm_allocate(mach_task_self(),(vm_address_t*) &ret, size, FALSE);
    if(!addr || err)	// try to allocate anywhere
	err = vm_allocate(mach_task_self(),(vm_address_t*) &ret, size, TRUE);
	
    if(err)
	ret = (void*) -1;
    else
	vm_protect(mach_task_self(),ret,size,FALSE,VM_PROT_READ|VM_PROT_WRITE);
127
128
129
130
131
#else
    ret = mmap(addr, size, PROT_READ | PROT_WRITE, 
	       MAP_ANON | MAP_PRIVATE, -1, 0);
#endif

132
133
134
135
136
137
138
139
    if (ret == (void *)-1) {
	if (errno == ENOMEM) {
	    barf("out of memory (requested %d bytes)", size);
	} else {
	    barf("getMBlock: mmap failed");
	}
    }

140
    return ret;
141
}
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166

// Implements the general case: allocate a chunk of memory of 'size'
// mblocks.

static void *
gen_map_mblocks (int size)
{
    int slop;
    void *ret;

    // Try to map a larger block, and take the aligned portion from
    // it (unmap the rest).
    size += MBLOCK_SIZE;
    ret = my_mmap(0, size);
    
    // unmap the slop bits around the chunk we allocated
    slop = (W_)ret & MBLOCK_MASK;
	
    if (munmap(ret, MBLOCK_SIZE - slop) == -1) {
	barf("gen_map_mblocks: munmap failed");
    }
    if (slop > 0 && munmap(ret+size-slop, slop) == -1) {
	barf("gen_map_mblocks: munmap failed");
    }
    
167
168
169
    // ToDo: if we happened to get an aligned block, then don't
    // unmap the excess, just use it.

170
171
172
173
174
175
176
177
178
    // next time, try after the block we just got.
    ret += MBLOCK_SIZE - slop;
    return ret;
}


// The external interface: allocate 'n' mblocks, and return the
// address.

179
180
181
void *
getMBlocks(nat n)
{
182
  static caddr_t next_request = (caddr_t)HEAP_BASE;
183
184
  caddr_t ret;
  lnat size = MBLOCK_SIZE * n;
185
  nat i;
186
 
187
188
189
190
191
  if (next_request == 0) {
      // use gen_map_mblocks the first time.
      ret = gen_map_mblocks(size);
  } else {
      ret = my_mmap(next_request, size);
192

193
194
195
196
197
198
199
200
201
202
203
204
205
      if (((W_)ret & MBLOCK_MASK) != 0) {
	  // misaligned block!
#ifdef DEBUG
	  belch("getMBlock: misaligned block %p returned when allocating %d megablock(s) at %p", ret, n, next_request);
#endif
	  
	  // unmap this block...
	  if (munmap(ret, size) == -1) {
	      barf("getMBlock: munmap failed");
	  }
	  // and do it the hard way
	  ret = gen_map_mblocks(size);
      }
206
207
  }

208
  // Next time, we'll try to allocate right after the block we just got.
209
  // ToDo: check that we haven't already grabbed the memory at next_request
210
211
  next_request = ret + size;

ken's avatar
ken committed
212
  IF_DEBUG(gc,fprintf(stderr,"Allocated %d megablock(s) at %p\n",n,ret));
213

214
215
216
217
  // fill in the table
  for (i = 0; i < n; i++) {
      mblockIsHeap( ret + i * MBLOCK_SIZE );
  }
218

219
  mblocks_allocated += n;
220

221
222
  return ret;
}
sof's avatar
sof committed
223

sof's avatar
sof committed
224
#else /* defined(mingw32_TARGET_OS) || defined(cygwin32_TARGET_OS) */
sof's avatar
sof committed
225
226
227
228
229

/*
 On Win32 platforms we make use of the two-phased virtual memory API
 to allocate mega blocks. We proceed as follows:

sof's avatar
sof committed
230
231
232
233
234
 Reserve a large chunk of VM (256M at the time, or what the user asked
 for via the -M option), but don't supply a base address that's aligned on
 a MB boundary. Instead we round up to the nearest mblock from the chunk of
 VM we're handed back from the OS (at the moment we just leave the 'slop' at
 the beginning of the reserved chunk unused - ToDo: reuse it .)
sof's avatar
sof committed
235
236

 Reserving memory doesn't allocate physical storage (not even in the
sof's avatar
sof committed
237
 page file), this is done later on by committing pages (or mega-blocks in
sof's avatar
sof committed
238
239
240
 our case).
*/

sof's avatar
sof committed
241
char* base_non_committed = (char*)0;
sof's avatar
sof committed
242
char* end_non_committed = (char*)0;
sof's avatar
sof committed
243

sof's avatar
sof committed
244
/* Default is to reserve 256M of VM to minimise the slop cost. */
245
#define SIZE_RESERVED_POOL  ( 256 * 1024 * 1024 )
sof's avatar
sof committed
246

sof's avatar
sof committed
247
248
249
/* Number of bytes reserved */
static unsigned long size_reserved_pool = SIZE_RESERVED_POOL;

sof's avatar
sof committed
250
251
252
253
254
255
void *
getMBlocks(nat n)
{
  static char* base_mblocks       = (char*)0;
  static char* next_request       = (char*)0;
  void* ret                       = (void*)0;
256
  int i;
sof's avatar
sof committed
257
258

  lnat size = MBLOCK_SIZE * n;
sof's avatar
sof committed
259
260
261
262
263
264
265
266
267
268
269
  
  if ( (base_non_committed == 0) || (next_request + size > end_non_committed) ) {
    if (base_non_committed) {
      barf("RTS exhausted max heap size (%d bytes)\n", size_reserved_pool);
    }
    if (RtsFlags.GcFlags.maxHeapSize != 0) {
      size_reserved_pool = BLOCK_SIZE * RtsFlags.GcFlags.maxHeapSize;
      if (size_reserved_pool < MBLOCK_SIZE) {
	size_reserved_pool = 2*MBLOCK_SIZE;
      }
    }
sof's avatar
sof committed
270
    base_non_committed = VirtualAlloc ( NULL
sof's avatar
sof committed
271
                                      , size_reserved_pool
sof's avatar
sof committed
272
273
274
275
				      , MEM_RESERVE
				      , PAGE_READWRITE
				      );
    if ( base_non_committed == 0 ) {
276
         fprintf(stderr, "getMBlocks: VirtualAlloc failed with: %ld\n", GetLastError());
sof's avatar
sof committed
277
278
         ret=(void*)-1;
    } else {
sof's avatar
sof committed
279
280
      end_non_committed = (char*)base_non_committed + (unsigned long)size_reserved_pool;
      /* The returned pointer is not aligned on a mega-block boundary. Make it. */
281
      base_mblocks = (char*)((unsigned long)base_non_committed & (unsigned long)~MBLOCK_MASK) + MBLOCK_SIZE;
282
283
#      if 0
       fprintf(stderr, "getMBlocks: Dropping %d bytes off of 256M chunk\n", 
sof's avatar
sof committed
284
	               (unsigned)base_mblocks - (unsigned)base_non_committed);
285
#      endif
sof's avatar
sof committed
286

sof's avatar
sof committed
287
       if ( ((char*)base_mblocks + size) > end_non_committed ) {
288
          fprintf(stderr, "getMBlocks: oops, committed too small a region to start with.");
sof's avatar
sof committed
289
290
291
292
293
294
295
296
297
298
	  ret=(void*)-1;
       } else {
          next_request = base_mblocks;
       }
    }
  }
  /* Commit the mega block(s) to phys mem */
  if ( ret != (void*)-1 ) {
     ret = VirtualAlloc(next_request, size, MEM_COMMIT, PAGE_READWRITE);
     if (ret == NULL) {
299
        fprintf(stderr, "getMBlocks: VirtualAlloc failed with: %ld\n", GetLastError());
sof's avatar
sof committed
300
301
302
303
304
        ret=(void*)-1;
     }
  }

  if (((W_)ret & MBLOCK_MASK) != 0) {
305
    barf("getMBlocks: misaligned block returned");
sof's avatar
sof committed
306
307
  }

308
309
310
  if (ret == (void*)-1) {
     barf("getMBlocks: unknown memory allocation failure on Win32.");
  }
sof's avatar
sof committed
311

312
  IF_DEBUG(gc,fprintf(stderr,"Allocated %d megablock(s) at 0x%x\n",n,(nat)ret));
sof's avatar
sof committed
313
314
315
316
  next_request = (char*)next_request + size;

  mblocks_allocated += n;
  
317
318
319
320
321
  // fill in the table
  for (i = 0; i < n; i++) {
      mblockIsHeap( ret + i * MBLOCK_SIZE );
  }

sof's avatar
sof committed
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
  return ret;
}

/* Hand back the physical memory that is allocated to a mega-block. 
   ToDo: chain the released mega block onto some list so that
         getMBlocks() can get at it.

   Currently unused.
*/
#if 0
void
freeMBlock(void* p, nat n)
{
  BOOL rc;

  rc = VirtualFree(p, n * MBLOCK_SIZE , MEM_DECOMMIT );
  
  if (rc == FALSE) {
340
#    ifdef DEBUG
sof's avatar
sof committed
341
     fprintf(stderr, "freeMBlocks: VirtualFree failed with: %d\n", GetLastError());
342
#    endif
sof's avatar
sof committed
343
344
345
346
347
348
  }

}
#endif

#endif