1 /* -----------------------------------------------------------------------------
3 * (c) The GHC Team 1998-1999
5 * MegaBlock Allocator Interface. This file contains all the dirty
6 * architecture-dependent hackery required to get a chunk of aligned
7 * memory from the operating system.
9 * ---------------------------------------------------------------------------*/
11 /* This is non-posix compliant. */
12 /* #include "PosixSource.h" */
18 #include "BlockAlloc.h"
29 #ifdef HAVE_SYS_TYPES_H
30 #include <sys/types.h>
32 #ifndef mingw32_HOST_OS
33 # ifdef HAVE_SYS_MMAN_H
34 # include <sys/mman.h>
44 #include <mach/vm_map.h>
49 lnat mblocks_allocated = 0;
51 /* -----------------------------------------------------------------------------
52 The MBlock Map: provides our implementation of HEAP_ALLOCED()
53 -------------------------------------------------------------------------- */
55 #if SIZEOF_VOID_P == 4
56 StgWord8 mblock_map[MBLOCK_MAP_SIZE]; // initially all zeros
57 #elif SIZEOF_VOID_P == 8
58 static MBlockMap dummy_mblock_map;
59 MBlockMap *mblock_cache = &dummy_mblock_map;
60 int mblock_map_count = 0;
61 MBlockMap **mblock_maps = NULL;
64 findMBlockMap(void *p)
67 StgWord32 hi = (StgWord32) (((StgWord)p) >> 32);
68 for( i = 0; i < mblock_map_count; i++ )
70 if(mblock_maps[i]->addrHigh32 == hi)
72 return mblock_maps[i];
79 slowIsHeapAlloced(void *p)
81 MBlockMap *map = findMBlockMap(p);
85 return map->mblocks[MBLOCK_MAP_ENTRY(p)];
93 markHeapAlloced(void *p)
95 #if SIZEOF_VOID_P == 4
96 mblock_map[MBLOCK_MAP_ENTRY(p)] = 1;
97 #elif SIZEOF_VOID_P == 8
98 MBlockMap *map = findMBlockMap(p);
102 mblock_maps = realloc(mblock_maps,
103 sizeof(MBlockMap*) * mblock_map_count);
104 map = mblock_maps[mblock_map_count-1] = calloc(1,sizeof(MBlockMap));
105 map->addrHigh32 = (StgWord32) (((StgWord)p) >> 32);
107 map->mblocks[MBLOCK_MAP_ENTRY(p)] = 1;
112 /* -----------------------------------------------------------------------------
113 Allocate new mblock(s)
114 -------------------------------------------------------------------------- */
119 return getMBlocks(1);
122 /* -----------------------------------------------------------------------------
125 On Unix-like systems, we use mmap() to allocate our memory. We
126 want memory in chunks of MBLOCK_SIZE, and aligned on an MBLOCK_SIZE
127 boundary. The mmap() interface doesn't give us this level of
128 control, so we have to use some heuristics.
130 In the general case, if we want a block of n megablocks, then we
131 allocate n+1 and trim off the slop from either side (using
132 munmap()) to get an aligned chunk of size n. However, the next
133 time we'll try to allocate directly after the previously allocated
134 chunk, on the grounds that this is aligned and likely to be free.
135 If it turns out that we were wrong, we have to munmap() and try
136 again using the general method.
138 Note on posix_memalign(): this interface is available on recent
139 systems and appears to provide exactly what we want. However, it
140 turns out not to be as good as our mmap() implementation, because
141 it wastes extra space (using double the address space, in a test on
142 x86_64/Linux). The problem seems to be that posix_memalign()
143 returns memory that can be free()'d, so the library must store
144 extra information along with the allocated block, thus messing up
145 the alignment. Hence, we don't use posix_memalign() for now.
147 -------------------------------------------------------------------------- */
149 #if !defined(mingw32_HOST_OS) && !defined(cygwin32_HOST_OS)
151 // A wrapper around mmap(), to abstract away from OS differences in
152 // the mmap() interface.
155 my_mmap (void *addr, lnat size)
159 #if defined(solaris2_HOST_OS) || defined(irix_HOST_OS)
161 int fd = open("/dev/zero",O_RDONLY);
162 ret = mmap(addr, size, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
166 ret = mmap(addr, size, PROT_READ | PROT_WRITE,
167 MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
169 // Without MAP_FIXED, Apple's mmap ignores addr.
170 // With MAP_FIXED, it overwrites already mapped regions, whic
171 // mmap(0, ... MAP_FIXED ...) is worst of all: It unmaps the program text
172 // and replaces it with zeroes, causing instant death.
173 // This behaviour seems to be conformant with IEEE Std 1003.1-2001.
174 // Let's just use the underlying Mach Microkernel calls directly,
175 // they're much nicer.
179 if(addr) // try to allocate at adress
180 err = vm_allocate(mach_task_self(),(vm_address_t*) &ret, size, FALSE);
181 if(!addr || err) // try to allocate anywhere
182 err = vm_allocate(mach_task_self(),(vm_address_t*) &ret, size, TRUE);
185 // don't know what the error codes mean exactly, assume it's
186 // not our problem though.
187 errorBelch("memory allocation failed (requested %lu bytes)", size);
188 stg_exit(EXIT_FAILURE);
190 vm_protect(mach_task_self(),ret,size,FALSE,VM_PROT_READ|VM_PROT_WRITE);
193 ret = mmap(addr, size, PROT_READ | PROT_WRITE | PROT_EXEC,
194 MAP_ANON | MAP_PRIVATE, -1, 0);
197 if (ret == (void *)-1) {
198 if (errno == ENOMEM ||
199 (errno == EINVAL && sizeof(void*)==4 && size >= 0xc0000000)) {
200 // If we request more than 3Gig, then we get EINVAL
201 // instead of ENOMEM (at least on Linux).
202 errorBelch("out of memory (requested %lu bytes)", size);
203 stg_exit(EXIT_FAILURE);
205 barf("getMBlock: mmap: %s", strerror(errno));
212 // Implements the general case: allocate a chunk of memory of 'size'
216 gen_map_mblocks (lnat size)
221 // Try to map a larger block, and take the aligned portion from
222 // it (unmap the rest).
224 ret = my_mmap(0, size);
226 // unmap the slop bits around the chunk we allocated
227 slop = (W_)ret & MBLOCK_MASK;
229 if (munmap(ret, MBLOCK_SIZE - slop) == -1) {
230 barf("gen_map_mblocks: munmap failed");
232 if (slop > 0 && munmap(ret+size-slop, slop) == -1) {
233 barf("gen_map_mblocks: munmap failed");
236 // ToDo: if we happened to get an aligned block, then don't
237 // unmap the excess, just use it. For this to work, you
238 // need to keep in mind the following:
239 // * Calling my_mmap() with an 'addr' arg pointing to
240 // already my_mmap()ed space is OK and won't fail.
241 // * If my_mmap() can't satisfy the request at the
242 // given 'next_request' address in getMBlocks(), that
243 // you unmap the extra mblock mmap()ed here (or simply
244 // satisfy yourself that the slop introduced isn't worth
248 // next time, try after the block we just got.
249 ret += MBLOCK_SIZE - slop;
254 // The external interface: allocate 'n' mblocks, and return the
260 static caddr_t next_request = (caddr_t)HEAP_BASE;
262 lnat size = MBLOCK_SIZE * n;
265 if (next_request == 0) {
266 // use gen_map_mblocks the first time.
267 ret = gen_map_mblocks(size);
269 ret = my_mmap(next_request, size);
271 if (((W_)ret & MBLOCK_MASK) != 0) {
273 #if 0 // defined(DEBUG)
274 errorBelch("warning: getMBlock: misaligned block %p returned when allocating %d megablock(s) at %p", ret, n, next_request);
277 // unmap this block...
278 if (munmap(ret, size) == -1) {
279 barf("getMBlock: munmap failed");
281 // and do it the hard way
282 ret = gen_map_mblocks(size);
286 // Next time, we'll try to allocate right after the block we just got.
287 // ToDo: check that we haven't already grabbed the memory at next_request
288 next_request = ret + size;
290 IF_DEBUG(gc,debugBelch("Allocated %d megablock(s) at %p\n",n,ret));
293 for (i = 0; i < n; i++) {
294 markHeapAlloced( ret + i * MBLOCK_SIZE );
297 mblocks_allocated += n;
302 #else /* defined(mingw32_HOST_OS) || defined(cygwin32_HOST_OS) */
305 On Win32 platforms we make use of the two-phased virtual memory API
306 to allocate mega blocks. We proceed as follows:
308 Reserve a large chunk of VM (256M at the time, or what the user asked
309 for via the -M option), but don't supply a base address that's aligned on
310 a MB boundary. Instead we round up to the nearest mblock from the chunk of
311 VM we're handed back from the OS (at the moment we just leave the 'slop' at
312 the beginning of the reserved chunk unused - ToDo: reuse it .)
314 Reserving memory doesn't allocate physical storage (not even in the
315 page file), this is done later on by committing pages (or mega-blocks in
319 char* base_non_committed = (char*)0;
320 char* end_non_committed = (char*)0;
322 /* Default is to reserve 256M of VM to minimise the slop cost. */
323 #define SIZE_RESERVED_POOL ( 256 * 1024 * 1024 )
325 /* Number of bytes reserved */
326 static unsigned long size_reserved_pool = SIZE_RESERVED_POOL;
331 static char* base_mblocks = (char*)0;
332 static char* next_request = (char*)0;
333 void* ret = (void*)0;
336 lnat size = MBLOCK_SIZE * n;
338 if ( (base_non_committed == 0) || (next_request + size > end_non_committed) ) {
339 if (base_non_committed) {
340 /* Tacky, but if no user-provided -M option is in effect,
341 * set it to the default (==256M) in time for the heap overflow PSA.
343 if (RtsFlags.GcFlags.maxHeapSize == 0) {
344 RtsFlags.GcFlags.maxHeapSize = size_reserved_pool / BLOCK_SIZE;
348 if (RtsFlags.GcFlags.maxHeapSize != 0) {
349 size_reserved_pool = BLOCK_SIZE * RtsFlags.GcFlags.maxHeapSize;
350 if (size_reserved_pool < MBLOCK_SIZE) {
351 size_reserved_pool = 2*MBLOCK_SIZE;
354 base_non_committed = VirtualAlloc ( NULL
359 if ( base_non_committed == 0 ) {
360 errorBelch("getMBlocks: VirtualAlloc failed with: %ld\n", GetLastError());
363 end_non_committed = (char*)base_non_committed + (unsigned long)size_reserved_pool;
364 /* The returned pointer is not aligned on a mega-block boundary. Make it. */
365 base_mblocks = (char*)((unsigned long)base_non_committed & (unsigned long)~MBLOCK_MASK) + MBLOCK_SIZE;
367 debugBelch("getMBlocks: Dropping %d bytes off of 256M chunk\n",
368 (unsigned)base_mblocks - (unsigned)base_non_committed);
371 if ( ((char*)base_mblocks + size) > end_non_committed ) {
372 debugBelch("getMBlocks: oops, committed too small a region to start with.");
375 next_request = base_mblocks;
379 /* Commit the mega block(s) to phys mem */
380 if ( ret != (void*)-1 ) {
381 ret = VirtualAlloc(next_request, size, MEM_COMMIT, PAGE_READWRITE);
383 debugBelch("getMBlocks: VirtualAlloc failed with: %ld\n", GetLastError());
388 if (((W_)ret & MBLOCK_MASK) != 0) {
389 barf("getMBlocks: misaligned block returned");
392 if (ret == (void*)-1) {
393 barf("getMBlocks: unknown memory allocation failure on Win32.");
396 IF_DEBUG(gc,debugBelch("Allocated %d megablock(s) at 0x%x\n",n,(nat)ret));
397 next_request = (char*)next_request + size;
399 mblocks_allocated += n;
402 for (i = 0; i < n; i++) {
403 markHeapAlloced( ret + i * MBLOCK_SIZE );
409 /* Hand back the physical memory that is allocated to a mega-block.
410 ToDo: chain the released mega block onto some list so that
411 getMBlocks() can get at it.
417 freeMBlock(void* p, nat n)
421 rc = VirtualFree(p, n * MBLOCK_SIZE , MEM_DECOMMIT );
425 debugBelch("freeMBlocks: VirtualFree failed with: %d\n", GetLastError());