2 * sgen-pinned-allocator.c: Simple generational GC.
5 * Paolo Molaro (lupus@ximian.com)
7 * Copyright 2005-2010 Novell, Inc (http://www.novell.com)
9 * Thread start/stop adapted from Boehm's GC:
10 * Copyright (c) 1994 by Xerox Corporation. All rights reserved.
11 * Copyright (c) 1996 by Silicon Graphics. All rights reserved.
12 * Copyright (c) 1998 by Fergus Henderson. All rights reserved.
13 * Copyright (c) 2000-2004 by Hewlett-Packard Company. All rights reserved.
15 * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
16 * OR IMPLIED. ANY USE IS AT YOUR OWN RISK.
18 * Permission is hereby granted to use or copy this program
19 * for any purpose, provided the above notices are retained on all copies.
20 * Permission to modify the code and to distribute modified code is granted,
21 * provided the above notices are retained, and a notice that the code was
22 * modified is included with the above copyright notice.
25 * Copyright 2001-2003 Ximian, Inc
26 * Copyright 2003-2010 Novell, Inc.
28 * Permission is hereby granted, free of charge, to any person obtaining
29 * a copy of this software and associated documentation files (the
30 * "Software"), to deal in the Software without restriction, including
31 * without limitation the rights to use, copy, modify, merge, publish,
32 * distribute, sublicense, and/or sell copies of the Software, and to
33 * permit persons to whom the Software is furnished to do so, subject to
34 * the following conditions:
36 * The above copyright notice and this permission notice shall be
37 * included in all copies or substantial portions of the Software.
39 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
40 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
41 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
42 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
43 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
44 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
45 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
52 #include "utils/mono-counters.h"
53 #include "metadata/sgen-gc.h"
55 /* Pinned objects are allocated in the LOS space if bigger than half a page
56 * or from freelists otherwise. We assume that pinned objects are relatively few
57 * and they have a slow dying speed (like interned strings, thread objects).
58 * As such they will be collected only at major collections.
59 * free lists are not global: when we need memory we allocate a PinnedChunk.
60 * Each pinned chunk is made of several pages, the first of wich is used
61 * internally for bookeeping (here think of a page as 4KB). The bookeeping
62 * includes the freelists vectors and info about the object size of each page
63 * in the pinned chunk. So, when needed, a free page is found in a pinned chunk,
64 * a size is assigned to it, the page is divided in the proper chunks and each
65 * chunk is added to the freelist. To not waste space, the remaining space in the
66 * first page is used as objects of size 16 or 32 (need to measure which are more
68 * We use this same structure to allocate memory used internally by the GC, so
69 * we never use malloc/free if we need to alloc during collection: the world is stopped
70 * and malloc/free will deadlock.
71 * When we want to iterate over pinned objects, we just scan a page at a time
72 * linearly according to the size of objects in the page: the next pointer used to link
73 * the items in the freelist uses the same word as the vtable. Since we keep freelists
74 * for each pinned chunk, if the word points outside the pinned chunk it means
76 * We could avoid this expensive scanning in creative ways. We could have a policy
77 * of putting in the pinned space only objects we know about that have no struct fields
78 * with references and we can easily use a even expensive write barrier for them,
79 * since pointer writes on such objects should be rare.
80 * The best compromise is to just alloc interned strings and System.MonoType in them.
81 * It would be nice to allocate MonoThread in it, too: must check that we properly
82 * use write barriers so we don't have to do any expensive scanning of the whole pinned
83 * chunk list during minor collections. We can avoid it now because we alloc in it only
84 * reference-free objects.
86 struct _SgenPinnedChunk {
89 SgenPinnedAllocator *allocator;
90 int *page_sizes; /* a 0 means the page is still unused */
92 SgenPinnedChunk *free_list_nexts [SGEN_PINNED_FREELIST_NUM_SLOTS];
94 void *data [1]; /* page sizes and free lists are stored here */
97 #define PINNED_FIRST_SLOT_SIZE (sizeof (gpointer) * 4)
98 #define MAX_FREELIST_SIZE 8192
100 /* This is a fixed value used for pinned chunks, not the system pagesize */
101 #define FREELIST_PAGESIZE (16*1024)
103 /* keep each size a multiple of ALLOC_ALIGN */
104 /* on 64 bit systems 8 is likely completely unused. */
105 static const int freelist_sizes [] = {
106 8, 16, 24, 32, 40, 48, 64, 80,
107 96, 128, 160, 192, 224, 256, 320, 384,
108 448, 512, 584, 680, 816, 1024, 1360, 2048,
109 2336, 2728, 3272, 4096, 5456, 8192 };
111 #define LARGE_PINNED_MEM_HEADER_MAGIC 0x7d289f3a
113 typedef struct _LargePinnedMemHeader LargePinnedMemHeader;
114 struct _LargePinnedMemHeader {
120 static long long pinned_chunk_bytes_alloced = 0;
121 static long long large_pinned_bytes_alloced = 0;
123 #ifdef HEAVY_STATISTICS
124 static long long stat_pinned_alloc = 0;
131 report_pinned_chunk (SgenPinnedChunk *chunk, int seq) {
133 int i, free_pages, num_free, free_mem;
135 for (i = 0; i < chunk->num_pages; ++i) {
136 if (!chunk->page_sizes [i])
139 printf ("Pinned chunk %d at %p, size: %d, pages: %d, free: %d\n", seq, chunk, chunk->num_pages * FREELIST_PAGESIZE, chunk->num_pages, free_pages);
140 free_mem = FREELIST_PAGESIZE * free_pages;
141 for (i = 0; i < SGEN_PINNED_FREELIST_NUM_SLOTS; ++i) {
142 if (!chunk->free_list [i])
145 p = chunk->free_list [i];
150 printf ("\tfree list of size %d, %d items\n", freelist_sizes [i], num_free);
151 free_mem += freelist_sizes [i] * num_free;
153 printf ("\tfree memory in chunk: %d\n", free_mem);
160 sgen_report_pinned_mem_usage (SgenPinnedAllocator *alc)
162 SgenPinnedChunk *chunk;
164 for (chunk = alc->chunk_list; chunk; chunk = chunk->block.next)
165 report_pinned_chunk (chunk, i++);
169 * Find the slot number in the freelist for memory chunks that
170 * can contain @size objects.
173 slot_for_size (size_t size)
176 /* do a binary search or lookup table later. */
177 for (slot = 0; slot < SGEN_PINNED_FREELIST_NUM_SLOTS; ++slot) {
178 if (freelist_sizes [slot] >= size)
181 g_assert_not_reached ();
186 * Build a free list for @size memory chunks from the memory area between
187 * start_page and end_page.
190 build_freelist (SgenPinnedAllocator *alc, SgenPinnedChunk *chunk, int slot, int size, char *start_page, char *end_page)
194 /*g_print ("building freelist for slot %d, size %d in %p\n", slot, size, chunk);*/
195 p = (void**)start_page;
196 end = (void**)(end_page - size);
197 g_assert (!chunk->free_list [slot]);
198 chunk->free_list [slot] = p;
199 while ((char*)p + size <= (char*)end) {
201 *p = (void*)((char*)p + size);
205 /*g_print ("%d items created, max: %d\n", count, (end_page - start_page) / size);*/
207 g_assert (!chunk->free_list_nexts [slot]);
208 chunk->free_list_nexts [slot] = alc->free_lists [slot];
209 alc->free_lists [slot] = chunk;
212 static SgenPinnedChunk*
213 alloc_pinned_chunk (SgenPinnedAllocator *alc)
215 SgenPinnedChunk *chunk;
217 int size = SGEN_PINNED_CHUNK_SIZE;
219 chunk = sgen_alloc_os_memory_aligned (size, size, TRUE);
220 chunk->block.role = MEMORY_ROLE_PINNED;
222 sgen_update_heap_boundaries ((mword)chunk, ((mword)chunk + size));
224 pinned_chunk_bytes_alloced += size;
226 /* setup the bookeeping fields */
227 chunk->num_pages = size / FREELIST_PAGESIZE;
228 offset = G_STRUCT_OFFSET (SgenPinnedChunk, data);
229 chunk->page_sizes = (void*)((char*)chunk + offset);
230 offset += sizeof (int) * chunk->num_pages;
231 offset = SGEN_ALIGN_UP (offset);
232 chunk->free_list = (void*)((char*)chunk + offset);
233 offset += sizeof (void*) * SGEN_PINNED_FREELIST_NUM_SLOTS;
234 offset = SGEN_ALIGN_UP (offset);
235 chunk->start_data = (void*)((char*)chunk + offset);
237 /* allocate the first page to the freelist */
238 chunk->page_sizes [0] = PINNED_FIRST_SLOT_SIZE;
239 build_freelist (alc, chunk, slot_for_size (PINNED_FIRST_SLOT_SIZE), PINNED_FIRST_SLOT_SIZE,
240 chunk->start_data, ((char*)chunk + FREELIST_PAGESIZE));
241 sgen_debug_printf (4, "Allocated pinned chunk %p, size: %d\n", chunk, size);
243 chunk->block.next = alc->chunk_list;
244 alc->chunk_list = chunk;
246 chunk->allocator = alc;
251 /* Must be called with an empty freelist for the given slot. */
253 populate_chunk_page (SgenPinnedAllocator *alc, SgenPinnedChunk *chunk, int slot)
255 int size = freelist_sizes [slot];
257 g_assert (!chunk->free_list [slot]);
258 g_assert (!chunk->free_list_nexts [slot]);
259 for (i = 0; i < chunk->num_pages; ++i) {
260 if (chunk->page_sizes [i])
262 chunk->page_sizes [i] = size;
263 build_freelist (alc, chunk, slot, size, (char*)chunk + FREELIST_PAGESIZE * i, (char*)chunk + FREELIST_PAGESIZE * (i + 1));
270 alloc_from_slot (SgenPinnedAllocator *alc, int slot)
272 SgenPinnedChunk *pchunk;
273 size_t size = freelist_sizes [slot];
275 if (alc->delayed_free_lists [slot]) {
278 p = alc->delayed_free_lists [slot];
279 } while (SGEN_CAS_PTR (&alc->delayed_free_lists [slot], *p, p) != p);
285 pchunk = alc->free_lists [slot];
287 void **p = pchunk->free_list [slot];
293 pchunk->free_list [slot] = next;
296 alc->free_lists [slot] = pchunk->free_list_nexts [slot];
297 pchunk->free_list_nexts [slot] = NULL;
304 for (pchunk = alc->chunk_list; pchunk; pchunk = pchunk->block.next) {
305 if (populate_chunk_page (alc, pchunk, slot))
309 pchunk = alloc_pinned_chunk (alc);
310 /* FIXME: handle OOM */
311 if (pchunk->free_list [slot])
313 if (!populate_chunk_page (alc, pchunk, slot))
314 g_assert_not_reached ();
318 /* used for the GC-internal data structures */
320 sgen_alloc_pinned (SgenPinnedAllocator *alc, size_t size)
325 HEAVY_STAT (++stat_pinned_alloc);
327 if (size > freelist_sizes [SGEN_PINNED_FREELIST_NUM_SLOTS - 1]) {
328 LargePinnedMemHeader *mh;
330 size += sizeof (LargePinnedMemHeader);
331 mh = sgen_alloc_os_memory (size, TRUE);
332 mh->magic = LARGE_PINNED_MEM_HEADER_MAGIC;
334 /* FIXME: do a CAS here */
335 large_pinned_bytes_alloced += size;
339 slot = slot_for_size (size);
340 g_assert (size <= freelist_sizes [slot]);
341 res = alloc_from_slot (alc, slot);
347 free_from_slot (SgenPinnedAllocator *alc, void *addr, int slot)
349 SgenPinnedChunk *pchunk = (SgenPinnedChunk*)SGEN_PINNED_CHUNK_FOR_PTR (addr);
353 g_assert (addr >= (void*)pchunk && (char*)addr < (char*)pchunk + pchunk->num_pages * FREELIST_PAGESIZE);
355 next = pchunk->free_list [slot];
357 pchunk->free_list [slot] = p;
360 g_assert (!pchunk->free_list_nexts [slot]);
361 pchunk->free_list_nexts [slot] = alc->free_lists [slot];
362 alc->free_lists [slot] = pchunk;
367 sgen_free_pinned (SgenPinnedAllocator *alc, void *addr, size_t size)
369 LargePinnedMemHeader *mh;
374 if (size <= freelist_sizes [SGEN_PINNED_FREELIST_NUM_SLOTS - 1]) {
375 int slot = slot_for_size (size);
376 free_from_slot (alc, addr, slot);
380 mh = (LargePinnedMemHeader*)((char*)addr - G_STRUCT_OFFSET (LargePinnedMemHeader, data));
381 g_assert (mh->magic == LARGE_PINNED_MEM_HEADER_MAGIC);
382 g_assert (mh->size == size + sizeof (LargePinnedMemHeader));
383 /* FIXME: do a CAS */
384 large_pinned_bytes_alloced -= mh->size;
385 sgen_free_os_memory (mh, mh->size);
389 sgen_init_pinned_allocator (void)
391 g_assert (SGEN_PINNED_FREELIST_NUM_SLOTS == sizeof (freelist_sizes) / sizeof (freelist_sizes [0]));
393 #ifdef HEAVY_STATISTICS
394 mono_counters_register ("Pinned allocs", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_pinned_alloc);
399 sgen_pinned_scan_objects (SgenPinnedAllocator *alc, IterateObjectCallbackFunc callback, void *callback_data)
401 SgenPinnedChunk *chunk;
406 for (chunk = alc->chunk_list; chunk; chunk = chunk->block.next) {
407 end_chunk = (char*)chunk + chunk->num_pages * FREELIST_PAGESIZE;
408 sgen_debug_printf (6, "Scanning pinned chunk %p (range: %p-%p)\n", chunk, chunk->start_data, end_chunk);
409 for (i = 0; i < chunk->num_pages; ++i) {
410 obj_size = chunk->page_sizes [i];
413 p = i? (char*)chunk + i * FREELIST_PAGESIZE: chunk->start_data;
414 endp = i? p + FREELIST_PAGESIZE: (char*)chunk + FREELIST_PAGESIZE;
415 sgen_debug_printf (6, "Page %d (size: %d, range: %p-%p)\n", i, obj_size, p, endp);
416 while (p + obj_size <= endp) {
418 /* if the first word (the vtable) is outside the chunk we have an object */
419 if (*ptr && (*ptr < (void*)chunk || *ptr >= end_chunk))
420 callback ((char*)ptr, obj_size, callback_data);
428 sgen_pinned_update_heap_boundaries (SgenPinnedAllocator *alc)
430 SgenPinnedChunk *chunk;
431 for (chunk = alc->chunk_list; chunk; chunk = chunk->block.next) {
432 char *end_chunk = (char*)chunk + chunk->num_pages * FREELIST_PAGESIZE;
433 sgen_update_heap_boundaries ((mword)chunk, (mword)end_chunk);
438 * the array of pointers from @start to @end contains conservative
439 * pointers to objects inside @chunk: mark each referenced object
443 mark_pinned_from_addresses (SgenPinnedChunk *chunk, void **start, void **end, IterateObjectCallbackFunc callback, void *callback_data)
445 for (; start < end; start++) {
447 int offset = (char*)addr - (char*)chunk;
448 int page = offset / FREELIST_PAGESIZE;
449 int obj_offset = page == 0? offset - ((char*)chunk->start_data - (char*)chunk): offset % FREELIST_PAGESIZE;
450 int slot_size = chunk->page_sizes [page];
452 /* the page is not allocated */
455 /* would be faster if we restrict the sizes to power of two,
456 * but that's a waste of memory: need to measure. it could reduce
457 * fragmentation since there are less pages needed, if for example
458 * someone interns strings of each size we end up with one page per
459 * interned string (still this is just ~40 KB): with more fine-grained sizes
460 * this increases the number of used pages.
463 obj_offset /= slot_size;
464 obj_offset *= slot_size;
465 addr = (char*)chunk->start_data + obj_offset;
467 obj_offset /= slot_size;
468 obj_offset *= slot_size;
469 addr = (char*)chunk + page * FREELIST_PAGESIZE + obj_offset;
472 /* if the vtable is inside the chunk it's on the freelist, so skip */
473 /* FIXME: is it possible that we're pinning objects more than once here? */
474 if (*ptr && (*ptr < (void*)chunk->start_data || *ptr > (void*)((char*)chunk + chunk->num_pages * FREELIST_PAGESIZE)))
475 callback (addr, slot_size, callback_data);
480 sgen_pinned_scan_pinned_objects (SgenPinnedAllocator *alc, IterateObjectCallbackFunc callback, void *callback_data)
482 SgenPinnedChunk *chunk;
484 /* look for pinned addresses for pinned-alloc objects */
485 sgen_debug_printf (6, "Pinning from pinned-alloc objects\n");
486 for (chunk = alc->chunk_list; chunk; chunk = chunk->block.next) {
488 void **pinned = sgen_find_optimized_pin_queue_area (chunk->start_data,
489 (char*)chunk + chunk->num_pages * FREELIST_PAGESIZE, &num_pinned);
491 mark_pinned_from_addresses (chunk, pinned, pinned + num_pinned, callback, callback_data);