/*
- * GENerational Conservative Garbage Collector for SBCL x86
+ * GENerational Conservative Garbage Collector for SBCL
*/
/*
* <ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps>.
*/
+#include <stdlib.h>
#include <stdio.h>
#include <signal.h>
#include <errno.h>
#include "thread.h"
#include "genesis/vector.h"
#include "genesis/weak-pointer.h"
+#include "genesis/fdefn.h"
#include "genesis/simple-fun.h"
#include "save.h"
#include "genesis/hash-table.h"
+#include "genesis/instance.h"
+#include "genesis/layout.h"
+
+#ifdef LUTEX_WIDETAG
+#include "genesis/lutex.h"
+#endif
/* forward declarations */
page_index_t gc_find_freeish_pages(long *restart_page_ptr, long nbytes,
generation_index_t from_space;
generation_index_t new_space;
+/* Set to 1 when in GC */
+boolean gc_active_p = 0;
+
/* should the GC be conservative on stack. If false (only right before
* saving a core), don't scan the stack / mark pages dont_move. */
static boolean conservative_stack = 1;
-/* An array of page structures is statically allocated.
+/* An array of page structures is allocated on gc initialization.
* This helps quickly map between an address its page structure.
- * NUM_PAGES is set from the size of the dynamic space. */
-struct page page_table[NUM_PAGES];
+ * page_table_pages is set from the size of the dynamic space. */
+unsigned page_table_pages;
+struct page *page_table;
/* To map addresses to page structures the address of the first page
* is needed. */
static void *heap_base = NULL;
-#if N_WORD_BITS == 32
- #define SIMPLE_ARRAY_WORD_WIDETAG SIMPLE_ARRAY_UNSIGNED_BYTE_32_WIDETAG
-#elif N_WORD_BITS == 64
- #define SIMPLE_ARRAY_WORD_WIDETAG SIMPLE_ARRAY_UNSIGNED_BYTE_64_WIDETAG
-#endif
-
/* Calculate the start address for the given page number. */
inline void *
page_address(page_index_t page_num)
if (index >= 0) {
index = ((unsigned long)index)/PAGE_BYTES;
- if (index < NUM_PAGES)
+ if (index < page_table_pages)
return (index);
}
* prevent a GC when a large number of new live objects have been
* added, in which case a GC could be a waste of time */
double min_av_mem_age;
+
+ /* A linked list of lutex structures in this generation, used for
+ * implementing lutex finalization. */
+#ifdef LUTEX_WIDETAG
+ struct lutex *lutexes;
+#else
+ void *lutexes;
+#endif
};
/* an array of generation structures. There needs to be one more
long count = 0;
for (i = 0; i < last_free_page; i++)
- if ((page_table[i].allocated != 0)
+ if ((page_table[i].allocated != FREE_PAGE_FLAG)
&& (page_table[i].gen == generation))
count++;
return count;
page_index_t i;
long count = 0;
for (i = 0; i < last_free_page; i++) {
- if ((page_table[i].allocated != 0) && (page_table[i].dont_move != 0)) {
+ if ((page_table[i].allocated != FREE_PAGE_FLAG)
+ && (page_table[i].dont_move != 0)) {
++count;
}
}
page_index_t i;
long result = 0;
for (i = 0; i < last_free_page; i++) {
- if ((page_table[i].allocated != 0) && (page_table[i].gen == gen))
+ if ((page_table[i].allocated != FREE_PAGE_FLAG)
+ && (page_table[i].gen == gen))
result += page_table[i].bytes_used;
}
return result;
/ ((double)generations[gen].bytes_allocated);
}
-void fpu_save(int *); /* defined in x86-assem.S */
-void fpu_restore(int *); /* defined in x86-assem.S */
/* The verbose argument controls how much to print: 0 for normal
* level of detail; 1 for debugging. */
static void
print_generation_stats(int verbose) /* FIXME: should take FILE argument */
{
generation_index_t i, gens;
- int fpu_state[27];
+
+#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
+#define FPU_STATE_SIZE 27
+ int fpu_state[FPU_STATE_SIZE];
+#elif defined(LISP_FEATURE_PPC)
+#define FPU_STATE_SIZE 32
+ long long fpu_state[FPU_STATE_SIZE];
+#endif
/* This code uses the FP instructions which may be set up for Lisp
* so they need to be saved and reset for C. */
/* Print the heap stats. */
fprintf(stderr,
- " Gen Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
+ " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB LUB !move Alloc Waste Trig WP GCs Mem-age\n");
for (i = 0; i < gens; i++) {
page_index_t j;
gc_assert(generations[i].bytes_allocated
== count_generation_bytes_allocated(i));
fprintf(stderr,
- " %1d: %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
+ " %1d: %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %5ld %8ld %5ld %8ld %4ld %3d %7.4f\n",
i,
- boxed_cnt, unboxed_cnt, large_boxed_cnt, large_unboxed_cnt,
+ generations[i].alloc_start_page,
+ generations[i].alloc_unboxed_start_page,
+ generations[i].alloc_large_start_page,
+ generations[i].alloc_large_unboxed_start_page,
+ boxed_cnt,
+ unboxed_cnt,
+ large_boxed_cnt,
+ large_unboxed_cnt,
pinned_cnt,
generations[i].bytes_allocated,
- (count_generation_pages(i)*PAGE_BYTES
- - generations[i].bytes_allocated),
+ (count_generation_pages(i)*PAGE_BYTES - generations[i].bytes_allocated),
generations[i].gc_trigger,
count_write_protect_generation_pages(i),
generations[i].num_gc,
}
\f
+#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
void fast_bzero(void*, size_t); /* in <arch>-assem.S */
+#endif
/* Zero the pages from START to END (inclusive), but use mmap/munmap instead
* if zeroing it ourselves, i.e. in practice give the memory back to the
if (start > end)
return;
+#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
fast_bzero(page_address(start), PAGE_BYTES*(1+end-start));
+#else
+ bzero(page_address(start), PAGE_BYTES*(1+end-start));
+#endif
+
}
/* Zero the pages from START to END (inclusive), except for those
page_table[i].need_to_zero = 1;
}
}
-'
/*
page_index_t last_page;
long bytes_found;
page_index_t i;
+ int ret;
/*
FSHOW((stderr,
gc_assert((alloc_region->first_page == 0)
&& (alloc_region->last_page == -1)
&& (alloc_region->free_pointer == alloc_region->end_addr));
- thread_mutex_lock(&free_pages_lock);
+ ret = thread_mutex_lock(&free_pages_lock);
+ gc_assert(ret == 0);
if (unboxed) {
first_page =
generations[gc_alloc_generation].alloc_unboxed_start_page;
/* Bump up last_free_page. */
if (last_page+1 > last_free_page) {
last_free_page = last_page+1;
- SetSymbolValue(ALLOCATION_POINTER,
- (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),
- 0);
+ /* do we only want to call this on special occasions? like for boxed_region? */
+ set_alloc_pointer((lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES));
}
- thread_mutex_unlock(&free_pages_lock);
+ ret = thread_mutex_unlock(&free_pages_lock);
+ gc_assert(ret == 0);
/* we can do this after releasing free_pages_lock */
if (gencgc_zero_check) {
long orig_first_page_bytes_used;
long region_size;
long byte_cnt;
+ int ret;
first_page = alloc_region->first_page;
next_page = first_page+1;
- thread_mutex_lock(&free_pages_lock);
+ ret = thread_mutex_lock(&free_pages_lock);
+ gc_assert(ret == 0);
if (alloc_region->free_pointer != alloc_region->start_addr) {
/* some bytes were allocated in the region */
orig_first_page_bytes_used = page_table[first_page].bytes_used;
page_table[next_page].allocated = FREE_PAGE_FLAG;
next_page++;
}
- thread_mutex_unlock(&free_pages_lock);
+ ret = thread_mutex_unlock(&free_pages_lock);
+ gc_assert(ret == 0);
+
/* alloc_region is per-thread, we're ok to do this unlocked */
gc_set_region_empty(alloc_region);
}
int more;
long bytes_used;
page_index_t next_page;
+ int ret;
- thread_mutex_lock(&free_pages_lock);
+ ret = thread_mutex_lock(&free_pages_lock);
+ gc_assert(ret == 0);
if (unboxed) {
first_page =
/* Bump up last_free_page */
if (last_page+1 > last_free_page) {
last_free_page = last_page+1;
- SetSymbolValue(ALLOCATION_POINTER,
- (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),0);
+ set_alloc_pointer((lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES));
}
- thread_mutex_unlock(&free_pages_lock);
+ ret = thread_mutex_unlock(&free_pages_lock);
+ gc_assert(ret == 0);
#ifdef READ_PROTECT_FREE_PAGES
os_protect(page_address(first_page),
static page_index_t gencgc_alloc_start_page = -1;
+void
+gc_heap_exhausted_error_or_lose (long available, long requested)
+{
+ /* Write basic information before doing anything else: if we don't
+ * call to lisp this is a must, and even if we do there is always the
+ * danger that we bounce back here before the error has been handled,
+ * or indeed even printed.
+ */
+ fprintf(stderr, "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
+ gc_active_p ? "garbage collection" : "allocation", available, requested);
+ if (gc_active_p || (available == 0)) {
+ /* If we are in GC, or totally out of memory there is no way
+ * to sanely transfer control to the lisp-side of things.
+ */
+ print_generation_stats(1);
+ lose("Heap exhausted, game over.");
+ }
+ else {
+ /* FIXME: assert free_pages_lock held */
+ thread_mutex_unlock(&free_pages_lock);
+ funcall2(SymbolFunction(HEAP_EXHAUSTED_ERROR),
+ make_fixnum(available), make_fixnum(requested));
+ lose("HEAP-EXHAUSTED-ERROR fell through");
+ }
+}
+
page_index_t
gc_find_freeish_pages(page_index_t *restart_page_ptr, long nbytes, int unboxed)
{
do {
first_page = restart_page;
if (large_p)
- while ((first_page < NUM_PAGES)
+ while ((first_page < page_table_pages)
&& (page_table[first_page].allocated != FREE_PAGE_FLAG))
first_page++;
else
- while (first_page < NUM_PAGES) {
+ while (first_page < page_table_pages) {
if(page_table[first_page].allocated == FREE_PAGE_FLAG)
break;
if((page_table[first_page].allocated ==
first_page++;
}
- if (first_page >= NUM_PAGES) {
- fprintf(stderr,
- "Argh! gc_find_free_space failed (first_page), nbytes=%ld.\n",
- nbytes);
- print_generation_stats(1);
- lose("\n");
- }
+ if (first_page >= page_table_pages)
+ gc_heap_exhausted_error_or_lose(0, nbytes);
gc_assert(page_table[first_page].write_protected == 0);
num_pages = 1;
while (((bytes_found < nbytes)
|| (!large_p && (num_pages < 2)))
- && (last_page < (NUM_PAGES-1))
+ && (last_page < (page_table_pages-1))
&& (page_table[last_page+1].allocated == FREE_PAGE_FLAG)) {
last_page++;
num_pages++;
gc_assert(bytes_found == region_size);
restart_page = last_page + 1;
- } while ((restart_page < NUM_PAGES) && (bytes_found < nbytes));
+ } while ((restart_page < page_table_pages) && (bytes_found < nbytes));
/* Check for a failure */
- if ((restart_page >= NUM_PAGES) && (bytes_found < nbytes)) {
- fprintf(stderr,
- "Argh! gc_find_freeish_pages failed (restart_page), nbytes=%ld.\n",
- nbytes);
- print_generation_stats(1);
- lose("\n");
- }
+ if ((restart_page >= page_table_pages) && (bytes_found < nbytes))
+ gc_heap_exhausted_error_or_lose(bytes_found, nbytes);
+
*restart_page_ptr=first_page;
return last_page;
\f
/*
- * vector-like objects
+ * Lutexes. Using the normal finalization machinery for finalizing
+ * lutexes is tricky, since the finalization depends on working lutexes.
+ * So we track the lutexes in the GC and finalize them manually.
*/
+#if defined(LUTEX_WIDETAG)
-/* FIXME: What does this mean? */
-int gencgc_hash = 1;
+/*
+ * Start tracking LUTEX in the GC, by adding it to the linked list of
+ * lutexes in the nursery generation. The caller is responsible for
+ * locking, and GCs must be inhibited until the registration is
+ * complete.
+ */
+void
+gencgc_register_lutex (struct lutex *lutex) {
+ int index = find_page_index(lutex);
+ generation_index_t gen;
+ struct lutex *head;
-static long
-scav_vector(lispobj *where, lispobj object)
-{
- unsigned long kv_length;
- lispobj *kv_vector;
- unsigned long length = 0; /* (0 = dummy to stop GCC warning) */
- struct hash_table *hash_table;
- lispobj empty_symbol;
- unsigned long *index_vector = NULL; /* (NULL = dummy to stop GCC warning) */
- unsigned long *next_vector = NULL; /* (NULL = dummy to stop GCC warning) */
- unsigned long *hash_vector = NULL; /* (NULL = dummy to stop GCC warning) */
- lispobj weak_p_obj;
- unsigned long next_vector_length = 0;
-
- /* FIXME: A comment explaining this would be nice. It looks as
- * though SB-VM:VECTOR-VALID-HASHING-SUBTYPE is set for EQ-based
- * hash tables in the Lisp HASH-TABLE code, and nowhere else. */
- if (HeaderValue(object) != subtype_VectorValidHashing)
- return 1;
+ /* This lutex is in static space, so we don't need to worry about
+ * finalizing it.
+ */
+ if (index == -1)
+ return;
- if (!gencgc_hash) {
- /* This is set for backward compatibility. FIXME: Do we need
- * this any more? */
- *where =
- (subtype_VectorMustRehash<<N_WIDETAG_BITS) | SIMPLE_VECTOR_WIDETAG;
- return 1;
- }
+ gen = page_table[index].gen;
- kv_length = fixnum_value(where[1]);
- kv_vector = where + 2; /* Skip the header and length. */
- /*FSHOW((stderr,"/kv_length = %d\n", kv_length));*/
+ gc_assert(gen >= 0);
+ gc_assert(gen < NUM_GENERATIONS);
- /* Scavenge element 0, which may be a hash-table structure. */
- scavenge(where+2, 1);
- if (!is_lisp_pointer(where[2])) {
- lose("no pointer at %x in hash table\n", where[2]);
- }
- hash_table = (struct hash_table *)native_pointer(where[2]);
- /*FSHOW((stderr,"/hash_table = %x\n", hash_table));*/
- if (widetag_of(hash_table->header) != INSTANCE_HEADER_WIDETAG) {
- lose("hash table not instance (%x at %x)\n",
- hash_table->header,
- hash_table);
- }
+ head = generations[gen].lutexes;
- /* Scavenge element 1, which should be some internal symbol that
- * the hash table code reserves for marking empty slots. */
- scavenge(where+3, 1);
- if (!is_lisp_pointer(where[3])) {
- lose("not empty-hash-table-slot symbol pointer: %x\n", where[3]);
- }
- empty_symbol = where[3];
- /* fprintf(stderr,"* empty_symbol = %x\n", empty_symbol);*/
- if (widetag_of(*(lispobj *)native_pointer(empty_symbol)) !=
- SYMBOL_HEADER_WIDETAG) {
- lose("not a symbol where empty-hash-table-slot symbol expected: %x\n",
- *(lispobj *)native_pointer(empty_symbol));
- }
+ lutex->gen = gen;
+ lutex->next = head;
+ lutex->prev = NULL;
+ if (head)
+ head->prev = lutex;
+ generations[gen].lutexes = lutex;
+}
- /* Scavenge hash table, which will fix the positions of the other
- * needed objects. */
- scavenge((lispobj *)hash_table,
- sizeof(struct hash_table) / sizeof(lispobj));
+/*
+ * Stop tracking LUTEX in the GC by removing it from the appropriate
+ * linked lists. This will only be called during GC, so no locking is
+ * needed.
+ */
+void
+gencgc_unregister_lutex (struct lutex *lutex) {
+ if (lutex->prev) {
+ lutex->prev->next = lutex->next;
+ } else {
+ generations[lutex->gen].lutexes = lutex->next;
+ }
- /* Cross-check the kv_vector. */
- if (where != (lispobj *)native_pointer(hash_table->table)) {
- lose("hash_table table!=this table %x\n", hash_table->table);
+ if (lutex->next) {
+ lutex->next->prev = lutex->prev;
}
- /* WEAK-P */
- weak_p_obj = hash_table->weak_p;
+ lutex->next = NULL;
+ lutex->prev = NULL;
+ lutex->gen = -1;
+}
- /* index vector */
- {
- lispobj index_vector_obj = hash_table->index_vector;
-
- if (is_lisp_pointer(index_vector_obj) &&
- (widetag_of(*(lispobj *)native_pointer(index_vector_obj)) ==
- SIMPLE_ARRAY_WORD_WIDETAG)) {
- index_vector =
- ((unsigned long *)native_pointer(index_vector_obj)) + 2;
- /*FSHOW((stderr, "/index_vector = %x\n",index_vector));*/
- length = fixnum_value(((lispobj *)native_pointer(index_vector_obj))[1]);
- /*FSHOW((stderr, "/length = %d\n", length));*/
- } else {
- lose("invalid index_vector %x\n", index_vector_obj);
- }
+/*
+ * Mark all lutexes in generation GEN as not live.
+ */
+static void
+unmark_lutexes (generation_index_t gen) {
+ struct lutex *lutex = generations[gen].lutexes;
+
+ while (lutex) {
+ lutex->live = 0;
+ lutex = lutex->next;
}
+}
- /* next vector */
- {
- lispobj next_vector_obj = hash_table->next_vector;
-
- if (is_lisp_pointer(next_vector_obj) &&
- (widetag_of(*(lispobj *)native_pointer(next_vector_obj)) ==
- SIMPLE_ARRAY_WORD_WIDETAG)) {
- next_vector = ((unsigned long *)native_pointer(next_vector_obj)) + 2;
- /*FSHOW((stderr, "/next_vector = %x\n", next_vector));*/
- next_vector_length = fixnum_value(((lispobj *)native_pointer(next_vector_obj))[1]);
- /*FSHOW((stderr, "/next_vector_length = %d\n", next_vector_length));*/
- } else {
- lose("invalid next_vector %x\n", next_vector_obj);
+/*
+ * Finalize all lutexes in generation GEN that have not been marked live.
+ */
+static void
+reap_lutexes (generation_index_t gen) {
+ struct lutex *lutex = generations[gen].lutexes;
+
+ while (lutex) {
+ struct lutex *next = lutex->next;
+ if (!lutex->live) {
+ lutex_destroy(lutex);
+ gencgc_unregister_lutex(lutex);
}
+ lutex = next;
}
+}
- /* maybe hash vector */
- {
- lispobj hash_vector_obj = hash_table->hash_vector;
-
- if (is_lisp_pointer(hash_vector_obj) &&
- (widetag_of(*(lispobj *)native_pointer(hash_vector_obj)) ==
- SIMPLE_ARRAY_WORD_WIDETAG)){
- hash_vector =
- ((unsigned long *)native_pointer(hash_vector_obj)) + 2;
- /*FSHOW((stderr, "/hash_vector = %x\n", hash_vector));*/
- gc_assert(fixnum_value(((lispobj *)native_pointer(hash_vector_obj))[1])
- == next_vector_length);
- } else {
- hash_vector = NULL;
- /*FSHOW((stderr, "/no hash_vector: %x\n", hash_vector_obj));*/
- }
+/*
+ * Mark LUTEX as live.
+ */
+static void
+mark_lutex (lispobj tagged_lutex) {
+ struct lutex *lutex = (struct lutex*) native_pointer(tagged_lutex);
+
+ lutex->live = 1;
+}
+
+/*
+ * Move all lutexes in generation FROM to generation TO.
+ */
+static void
+move_lutexes (generation_index_t from, generation_index_t to) {
+ struct lutex *tail = generations[from].lutexes;
+
+ /* Nothing to move */
+ if (!tail)
+ return;
+
+ /* Change the generation of the lutexes in FROM. */
+ while (tail->next) {
+ tail->gen = to;
+ tail = tail->next;
}
+ tail->gen = to;
- /* These lengths could be different as the index_vector can be a
- * different length from the others, a larger index_vector could help
- * reduce collisions. */
- gc_assert(next_vector_length*2 == kv_length);
+ /* Link the last lutex in the FROM list to the start of the TO list */
+ tail->next = generations[to].lutexes;
- /* now all set up.. */
+ /* And vice versa */
+ if (generations[to].lutexes) {
+ generations[to].lutexes->prev = tail;
+ }
- /* Work through the KV vector. */
- {
- long i;
- for (i = 1; i < next_vector_length; i++) {
- lispobj old_key = kv_vector[2*i];
+ /* And update the generations structures to match this */
+ generations[to].lutexes = generations[from].lutexes;
+ generations[from].lutexes = NULL;
+}
-#if N_WORD_BITS == 32
- unsigned long old_index = (old_key & 0x1fffffff)%length;
-#elif N_WORD_BITS == 64
- unsigned long old_index = (old_key & 0x1fffffffffffffff)%length;
-#endif
+static long
+scav_lutex(lispobj *where, lispobj object)
+{
+ mark_lutex((lispobj) where);
- /* Scavenge the key and value. */
- scavenge(&kv_vector[2*i],2);
+ return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
+}
- /* Check whether the key has moved and is EQ based. */
- {
- lispobj new_key = kv_vector[2*i];
-#if N_WORD_BITS == 32
- unsigned long new_index = (new_key & 0x1fffffff)%length;
-#elif N_WORD_BITS == 64
- unsigned long new_index = (new_key & 0x1fffffffffffffff)%length;
-#endif
+static lispobj
+trans_lutex(lispobj object)
+{
+ struct lutex *lutex = native_pointer(object);
+ lispobj copied;
+ size_t words = CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
+ gc_assert(is_lisp_pointer(object));
+ copied = copy_object(object, words);
- if ((old_index != new_index) &&
- ((!hash_vector) ||
- (hash_vector[i] == MAGIC_HASH_VECTOR_VALUE)) &&
- ((new_key != empty_symbol) ||
- (kv_vector[2*i] != empty_symbol))) {
-
- /*FSHOW((stderr,
- "* EQ key %d moved from %x to %x; index %d to %d\n",
- i, old_key, new_key, old_index, new_index));*/
-
- if (index_vector[old_index] != 0) {
- /*FSHOW((stderr, "/P1 %d\n", index_vector[old_index]));*/
-
- /* Unlink the key from the old_index chain. */
- if (index_vector[old_index] == i) {
- /*FSHOW((stderr, "/P2a %d\n", next_vector[i]));*/
- index_vector[old_index] = next_vector[i];
- /* Link it into the needing rehash chain. */
- next_vector[i] = fixnum_value(hash_table->needing_rehash);
- hash_table->needing_rehash = make_fixnum(i);
- /*SHOW("P2");*/
- } else {
- unsigned long prior = index_vector[old_index];
- unsigned long next = next_vector[prior];
-
- /*FSHOW((stderr, "/P3a %d %d\n", prior, next));*/
-
- while (next != 0) {
- /*FSHOW((stderr, "/P3b %d %d\n", prior, next));*/
- if (next == i) {
- /* Unlink it. */
- next_vector[prior] = next_vector[next];
- /* Link it into the needing rehash
- * chain. */
- next_vector[next] =
- fixnum_value(hash_table->needing_rehash);
- hash_table->needing_rehash = make_fixnum(next);
- /*SHOW("/P3");*/
- break;
- }
- prior = next;
- next = next_vector[next];
- }
- }
- }
- }
- }
- }
+ /* Update the links, since the lutex moved in memory. */
+ if (lutex->next) {
+ lutex->next->prev = native_pointer(copied);
}
- return (CEILING(kv_length + 2, 2));
+
+ if (lutex->prev) {
+ lutex->prev->next = native_pointer(copied);
+ } else {
+ generations[lutex->gen].lutexes = native_pointer(copied);
+ }
+
+ return copied;
}
+static long
+size_lutex(lispobj *where)
+{
+ return CEILING(sizeof(struct lutex)/sizeof(lispobj), 2);
+}
+#endif /* LUTEX_WIDETAG */
\f
/*
#endif
case SAP_WIDETAG:
case WEAK_POINTER_WIDETAG:
+#ifdef LUTEX_WIDETAG
+ case LUTEX_WIDETAG:
+#endif
break;
default:
return 1;
}
+#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
+
/* Adjust large bignum and vector objects. This will adjust the
* allocated region if the size has shrunk, and move unboxed objects
* into unboxed pages. The pages are not promoted here, and the
return;
}
+#endif
+
/* Take a possible pointer to a Lisp object and mark its page in the
* page_table so that it will not be relocated during a GC.
*
*
* It is also assumed that the current gc_alloc() region has been
* flushed and the tables updated. */
+
+#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
+
static void
preserve_pointer(void *addr)
{
/* Check that the page is now static. */
gc_assert(page_table[addr_page_index].dont_move != 0);
}
+
+#endif
+
\f
/* If the given page is not write-protected, then scan it for pointers
* to younger generations or the top temp. generation, if no
#define SC_GEN_CK 0
#if SC_GEN_CK
/* Clear the write_protected_cleared flags on all pages. */
- for (i = 0; i < NUM_PAGES; i++)
+ for (i = 0; i < page_table_pages; i++)
page_table[i].write_protected_cleared = 0;
#endif
#if SC_GEN_CK
/* Check that none of the write_protected pages in this generation
* have been written to. */
- for (i = 0; i < NUM_PAGES; i++) {
+ for (i = 0; i < page_table_pages; i++) {
if ((page_table[i].allocation != FREE_PAGE_FLAG)
&& (page_table[i].bytes_used != 0)
&& (page_table[i].gen == generation)
/* Record all new areas now. */
record_new_objects = 2;
+ /* Give a chance to weak hash tables to make other objects live.
+ * FIXME: The algorithm implemented here for weak hash table gcing
+ * is O(W^2+N) as Bruno Haible warns in
+ * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
+ * see "Implementation 2". */
+ scav_weak_hash_tables();
+
/* Flush the current regions updating the tables. */
gc_alloc_update_all_page_tables();
if (gencgc_verbose)
SHOW("new_areas overflow, doing full scavenge");
- /* Don't need to record new areas that get scavenge anyway
- * during scavenge_newspace_generation_one_scan. */
+ /* Don't need to record new areas that get scavenged
+ * anyway during scavenge_newspace_generation_one_scan. */
record_new_objects = 1;
scavenge_newspace_generation_one_scan(generation);
/* Record all new areas now. */
record_new_objects = 2;
+ scav_weak_hash_tables();
+
/* Flush the current regions updating the tables. */
gc_alloc_update_all_page_tables();
scavenge(page_address(page)+offset, size);
}
+ scav_weak_hash_tables();
+
/* Flush the current regions updating the tables. */
gc_alloc_update_all_page_tables();
}
#if SC_NS_GEN_CK
/* Check that none of the write_protected pages in this generation
* have been written to. */
- for (i = 0; i < NUM_PAGES; i++) {
+ for (i = 0; i < page_table_pages; i++) {
if ((page_table[i].allocation != FREE_PAGE_FLAG)
&& (page_table[i].bytes_used != 0)
&& (page_table[i].gen == generation)
}
#endif
-extern long undefined_tramp;
-
static void
verify_space(lispobj *start, size_t words)
{
*/
} else {
/* Verify that it points to another valid space. */
- if (!to_readonly_space && !to_static_space
- && (thing != (unsigned long)&undefined_tramp)) {
+ if (!to_readonly_space && !to_static_space) {
lose("Ptr %x @ %x sees junk.\n", thing, start);
}
}
case SINGLE_FLOAT_WIDETAG:
#endif
case UNBOUND_MARKER_WIDETAG:
- case INSTANCE_HEADER_WIDETAG:
case FDEFN_WIDETAG:
count = 1;
break;
+ case INSTANCE_HEADER_WIDETAG:
+ {
+ lispobj nuntagged;
+ long ntotal = HeaderValue(thing);
+ lispobj layout = ((struct instance *)start)->slots[0];
+ if (!layout) {
+ count = 1;
+ break;
+ }
+ nuntagged = ((struct layout *)native_pointer(layout))->n_untagged_slots;
+ verify_space(start + 1, ntotal - fixnum_value(nuntagged));
+ count = ntotal + 1;
+ break;
+ }
case CODE_HEADER_WIDETAG:
{
lispobj object = *start;
#endif
case SAP_WIDETAG:
case WEAK_POINTER_WIDETAG:
+#ifdef LUTEX_WIDETAG
+ case LUTEX_WIDETAG:
+#endif
count = (sizetab[widetag_of(*start)])(start);
break;
default:
+ FSHOW((stderr,
+ "/Unhandled widetag 0x%x at 0x%x\n",
+ widetag_of(*start), start));
+ fflush(stderr);
gc_abort();
}
}
struct thread *th;
for_each_thread(th) {
long binding_stack_size =
- (lispobj*)SymbolValue(BINDING_STACK_POINTER,th)
+ (lispobj*)get_binding_stack_pointer(th)
- (lispobj*)th->binding_stack_start;
verify_space(th->binding_stack_start, binding_stack_size);
}
}
}
+#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
+
+static void
+scavenge_control_stack()
+{
+ unsigned long control_stack_size;
+
+ /* This is going to be a big problem when we try to port threads
+ * to PPC... CLH */
+ struct thread *th = arch_os_get_current_thread();
+ lispobj *control_stack =
+ (lispobj *)(th->control_stack_start);
+
+ control_stack_size = current_control_stack_pointer - control_stack;
+ scavenge(control_stack, control_stack_size);
+}
+
+/* Scavenging Interrupt Contexts */
+
+static int boxed_registers[] = BOXED_REGISTERS;
+
+static void
+scavenge_interrupt_context(os_context_t * context)
+{
+ int i;
+
+#ifdef reg_LIP
+ unsigned long lip;
+ unsigned long lip_offset;
+ int lip_register_pair;
+#endif
+ unsigned long pc_code_offset;
+
+#ifdef ARCH_HAS_LINK_REGISTER
+ unsigned long lr_code_offset;
+#endif
+#ifdef ARCH_HAS_NPC_REGISTER
+ unsigned long npc_code_offset;
+#endif
+
+#ifdef reg_LIP
+ /* Find the LIP's register pair and calculate it's offset */
+ /* before we scavenge the context. */
+
+ /*
+ * I (RLT) think this is trying to find the boxed register that is
+ * closest to the LIP address, without going past it. Usually, it's
+ * reg_CODE or reg_LRA. But sometimes, nothing can be found.
+ */
+ lip = *os_context_register_addr(context, reg_LIP);
+ lip_offset = 0x7FFFFFFF;
+ lip_register_pair = -1;
+ for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
+ unsigned long reg;
+ long offset;
+ int index;
+
+ index = boxed_registers[i];
+ reg = *os_context_register_addr(context, index);
+ if ((reg & ~((1L<<N_LOWTAG_BITS)-1)) <= lip) {
+ offset = lip - reg;
+ if (offset < lip_offset) {
+ lip_offset = offset;
+ lip_register_pair = index;
+ }
+ }
+ }
+#endif /* reg_LIP */
+
+ /* Compute the PC's offset from the start of the CODE */
+ /* register. */
+ pc_code_offset = *os_context_pc_addr(context) - *os_context_register_addr(context, reg_CODE);
+#ifdef ARCH_HAS_NPC_REGISTER
+ npc_code_offset = *os_context_npc_addr(context) - *os_context_register_addr(context, reg_CODE);
+#endif /* ARCH_HAS_NPC_REGISTER */
+
+#ifdef ARCH_HAS_LINK_REGISTER
+ lr_code_offset =
+ *os_context_lr_addr(context) -
+ *os_context_register_addr(context, reg_CODE);
+#endif
+
+ /* Scanvenge all boxed registers in the context. */
+ for (i = 0; i < (sizeof(boxed_registers) / sizeof(int)); i++) {
+ int index;
+ lispobj foo;
+
+ index = boxed_registers[i];
+ foo = *os_context_register_addr(context, index);
+ scavenge(&foo, 1);
+ *os_context_register_addr(context, index) = foo;
+
+ scavenge((lispobj*) &(*os_context_register_addr(context, index)), 1);
+ }
+
+#ifdef reg_LIP
+ /* Fix the LIP */
+
+ /*
+ * But what happens if lip_register_pair is -1? *os_context_register_addr on Solaris
+ * (see solaris_register_address in solaris-os.c) will return
+ * &context->uc_mcontext.gregs[2]. But gregs[2] is REG_nPC. Is
+ * that what we really want? My guess is that that is not what we
+ * want, so if lip_register_pair is -1, we don't touch reg_LIP at
+ * all. But maybe it doesn't really matter if LIP is trashed?
+ */
+ if (lip_register_pair >= 0) {
+ *os_context_register_addr(context, reg_LIP) =
+ *os_context_register_addr(context, lip_register_pair) + lip_offset;
+ }
+#endif /* reg_LIP */
+
+ /* Fix the PC if it was in from space */
+ if (from_space_p(*os_context_pc_addr(context)))
+ *os_context_pc_addr(context) = *os_context_register_addr(context, reg_CODE) + pc_code_offset;
+
+#ifdef ARCH_HAS_LINK_REGISTER
+ /* Fix the LR ditto; important if we're being called from
+ * an assembly routine that expects to return using blr, otherwise
+ * harmless */
+ if (from_space_p(*os_context_lr_addr(context)))
+ *os_context_lr_addr(context) =
+ *os_context_register_addr(context, reg_CODE) + lr_code_offset;
+#endif
+
+#ifdef ARCH_HAS_NPC_REGISTER
+ if (from_space_p(*os_context_npc_addr(context)))
+ *os_context_npc_addr(context) = *os_context_register_addr(context, reg_CODE) + npc_code_offset;
+#endif /* ARCH_HAS_NPC_REGISTER */
+}
+
+void
+scavenge_interrupt_contexts(void)
+{
+ int i, index;
+ os_context_t *context;
+
+ struct thread *th=arch_os_get_current_thread();
+
+ index = fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,0));
+
+#if defined(DEBUG_PRINT_CONTEXT_INDEX)
+ printf("Number of active contexts: %d\n", index);
+#endif
+
+ for (i = 0; i < index; i++) {
+ context = th->interrupt_contexts[i];
+ scavenge_interrupt_context(context);
+ }
+}
+
+#endif
+
+#if defined(LISP_FEATURE_SB_THREAD)
+static void
+preserve_context_registers (os_context_t *c)
+{
+ void **ptr;
+ /* On Darwin the signal context isn't a contiguous block of memory,
+ * so just preserve_pointering its contents won't be sufficient.
+ */
+#if defined(LISP_FEATURE_DARWIN)
+#if defined LISP_FEATURE_X86
+ preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
+ preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
+ preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
+ preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
+ preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
+ preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
+ preserve_pointer((void*)*os_context_pc_addr(c));
+#else
+ #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
+#endif
+#endif
+ for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
+ preserve_pointer(*ptr);
+ }
+}
+#endif
+
/* Garbage collect a generation. If raise is 0 then the remains of the
* generation are not raised to the next generation. */
static void
/* The oldest generation can't be raised. */
gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));
+ /* Check if weak hash tables were processed in the previous GC. */
+ gc_assert(weak_hash_tables == NULL);
+
/* Initialize the weak pointer list. */
weak_pointers = NULL;
+#ifdef LUTEX_WIDETAG
+ unmark_lutexes(generation);
+#endif
+
/* When a generation is not being raised it is transported to a
* temporary generation (NUM_GENERATIONS), and lowered when
* done. Set up this new generation. There should be no pages
* initiates GC. If you ever call GC from inside an altstack
* handler, you will lose. */
+#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
/* And if we're saving a core, there's no point in being conservative. */
if (conservative_stack) {
for_each_thread(th) {
if (esp1>=(void **)th->control_stack_start &&
esp1<(void **)th->control_stack_end) {
if(esp1<esp) esp=esp1;
- for(ptr = (void **)(c+1); ptr>=(void **)c; ptr--) {
- preserve_pointer(*ptr);
- }
+ preserve_context_registers(c);
}
}
}
#else
esp = (void **)((void *)&raise);
#endif
- for (ptr = (void **)th->control_stack_end; ptr > esp; ptr--) {
+ for (ptr = ((void **)th->control_stack_end)-1; ptr > esp; ptr--) {
preserve_pointer(*ptr);
}
}
}
+#endif
+
#ifdef QSHOW
if (gencgc_verbose > 1) {
long num_dont_move_pages = count_dont_move_pages();
/* Scavenge all the rest of the roots. */
+#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
+ /*
+ * If not x86, we need to scavenge the interrupt context(s) and the
+ * control stack.
+ */
+ scavenge_interrupt_contexts();
+ scavenge_control_stack();
+#endif
+
/* Scavenge the Lisp functions of the interrupt handlers, taking
* care to avoid SIG_DFL and SIG_IGN. */
for (i = 0; i < NSIG; i++) {
{
struct thread *th;
for_each_thread(th) {
- long len= (lispobj *)SymbolValue(BINDING_STACK_POINTER,th) -
+ long len= (lispobj *)get_binding_stack_pointer(th) -
th->binding_stack_start;
scavenge((lispobj *) th->binding_stack_start,len);
#ifdef LISP_FEATURE_SB_THREAD
}
#endif
+ scan_weak_hash_tables();
scan_weak_pointers();
/* Flush the current regions, updating the tables. */
generations[generation].num_gc = 0;
else
++generations[generation].num_gc;
+
+#ifdef LUTEX_WIDETAG
+ reap_lutexes(generation);
+ if (raise)
+ move_lutexes(generation, generation+1);
+#endif
}
/* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
last_free_page = last_page+1;
- SetSymbolValue(ALLOCATION_POINTER,
- (lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES),0);
+ set_alloc_pointer((lispobj)(((char *)heap_base) + last_free_page*PAGE_BYTES));
return 0; /* dummy value: return something ... */
}
last_page++;
}
+ /* There's a mysterious Solaris/x86 problem with using mmap
+ * tricks for memory zeroing. See sbcl-devel thread
+ * "Re: patch: standalone executable redux".
+ */
+#if defined(LISP_FEATURE_SUNOS)
+ zero_pages(first_page, last_page-1);
+#else
zero_pages_with_mmap(first_page, last_page-1);
+#endif
first_page = last_page;
}
FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
+ gc_active_p = 1;
+
if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
FSHOW((stderr,
"/collect_garbage: last_gen = %d, doing a level 0 GC\n",
/* Save the high-water mark before updating last_free_page */
if (last_free_page > high_water_mark)
high_water_mark = last_free_page;
+
update_dynamic_space_free_pointer();
+
auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
if(gencgc_verbose)
fprintf(stderr,"Next gc when %ld bytes have been consed\n",
high_water_mark = 0;
}
+ gc_active_p = 0;
+
SHOW("returning from collect_garbage");
}
if (gencgc_verbose > 1)
SHOW("entering gc_free_heap");
- for (page = 0; page < NUM_PAGES; page++) {
+ for (page = 0; page < page_table_pages; page++) {
/* Skip free pages which should already be zero filled. */
if (page_table[page].allocated != FREE_PAGE_FLAG) {
void *page_start, *addr;
generations[page].gc_trigger = 2000000;
generations[page].num_gc = 0;
generations[page].cum_sum_bytes_allocated = 0;
+ generations[page].lutexes = NULL;
}
if (gencgc_verbose > 1)
gc_set_region_empty(&unboxed_region);
last_free_page = 0;
- SetSymbolValue(ALLOCATION_POINTER, (lispobj)((char *)heap_base),0);
+ set_alloc_pointer((lispobj)((char *)heap_base));
if (verify_after_free_heap) {
/* Check whether purify has left any bad pointers. */
{
page_index_t i;
+ /* Compute the number of pages needed for the dynamic space.
+ * Dynamic space size should be aligned on page size. */
+ page_table_pages = dynamic_space_size/PAGE_BYTES;
+ gc_assert(dynamic_space_size == (size_t) page_table_pages*PAGE_BYTES);
+
+ page_table = calloc(page_table_pages, sizeof(struct page));
+ gc_assert(page_table);
+
gc_init_tables();
- scavtab[SIMPLE_VECTOR_WIDETAG] = scav_vector;
scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;
+#ifdef LUTEX_WIDETAG
+ scavtab[LUTEX_WIDETAG] = scav_lutex;
+ transother[LUTEX_WIDETAG] = trans_lutex;
+ sizetab[LUTEX_WIDETAG] = size_lutex;
+#endif
+
heap_base = (void*)DYNAMIC_SPACE_START;
/* Initialize each page structure. */
- for (i = 0; i < NUM_PAGES; i++) {
+ for (i = 0; i < page_table_pages; i++) {
/* Initialize all pages as free. */
page_table[i].allocated = FREE_PAGE_FLAG;
page_table[i].bytes_used = 0;
generations[i].bytes_consed_between_gc = 2000000;
generations[i].trigger_age = 1;
generations[i].min_av_mem_age = 0.75;
+ generations[i].lutexes = NULL;
}
/* Initialize gc_alloc. */
gencgc_pickup_dynamic(void)
{
page_index_t page = 0;
- long alloc_ptr = SymbolValue(ALLOCATION_POINTER,0);
+ long alloc_ptr = get_alloc_pointer();
lispobj *prev=(lispobj *)page_address(page);
generation_index_t gen = PSEUDO_STATIC_GENERATION;
page++;
} while ((long)page_address(page) < alloc_ptr);
+#ifdef LUTEX_WIDETAG
+ /* Lutexes have been registered in generation 0 by coreparse, and
+ * need to be moved to the right one manually.
+ */
+ move_lutexes(0, PSEUDO_STATIC_GENERATION);
+#endif
+
last_free_page = page;
generations[gen].bytes_allocated = PAGE_BYTES*page;
void *new_obj;
void *new_free_pointer;
gc_assert(nbytes>0);
+
/* Check for alignment allocation problems. */
gc_assert((((unsigned long)region->free_pointer & LOWTAG_MASK) == 0)
&& ((nbytes & LOWTAG_MASK) == 0));
+
#if 0
if(all_threads)
/* there are a few places in the C code that allocate data in the
* heap before Lisp starts. This is before interrupts are enabled,
* so we don't need to check for pseudo-atomic */
#ifdef LISP_FEATURE_SB_THREAD
- if(!SymbolValue(PSEUDO_ATOMIC_ATOMIC,th)) {
+ if(!get_psuedo_atomic_atomic(th)) {
register u32 fs;
fprintf(stderr, "fatal error in thread 0x%x, tid=%ld\n",
th,th->os_thread);
lose("If you see this message before 2004.01.31, mail details to sbcl-devel\n");
}
#else
- gc_assert(SymbolValue(PSEUDO_ATOMIC_ATOMIC,th));
+ gc_assert(get_pseudo_atomic_atomic(th));
#endif
#endif
* we should GC in the near future
*/
if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) {
- gc_assert(fixnum_value(SymbolValue(PSEUDO_ATOMIC_ATOMIC,thread)));
+ gc_assert(get_pseudo_atomic_atomic(thread));
/* Don't flood the system with interrupts if the need to gc is
* already noted. This can happen for example when SUB-GC
* allocates or after a gc triggered in a WITHOUT-GCING. */
* section */
SetSymbolValue(GC_PENDING,T,thread);
if (SymbolValue(GC_INHIBIT,thread) == NIL)
- arch_set_pseudo_atomic_interrupted(0);
+ set_pseudo_atomic_interrupted(thread);
}
}
new_obj = gc_alloc_with_region(nbytes,0,region,0);
* does this test after the first one has already set wp=0
*/
if(page_table[page_index].write_protected_cleared != 1)
- lose("fault in heap page not marked as write-protected\n");
+ lose("fault in heap page %d not marked as write-protected\nboxed_region.first_page: %d, boxed_region.last_page %d\n",
+ page_index, boxed_region.first_page, boxed_region.last_page);
}
/* Don't worry, we can handle it. */
return 1;
* function being set to the value of the static symbol
* SB!VM:RESTART-LISP-FUNCTION */
void
-gc_and_save(char *filename)
+gc_and_save(char *filename, int prepend_runtime)
{
- FILE *file = open_core_for_saving(filename);
- if (!file) {
- perror(filename);
- return;
- }
+ FILE *file;
+ void *runtime_bytes = NULL;
+ size_t runtime_size;
+
+ file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
+ &runtime_size);
+ if (file == NULL)
+ return;
+
conservative_stack = 0;
/* The filename might come from Lisp, and be moved by the now
gencgc_alloc_start_page = -1;
collect_garbage(HIGHEST_NORMAL_GENERATION+1);
+ if (prepend_runtime)
+ save_runtime_to_filehandle(file, runtime_bytes, runtime_size);
+
/* The dumper doesn't know that pages need to be zeroed before use. */
zero_all_free_pages();
- save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0));
+ save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
+ prepend_runtime);
/* Oops. Save still managed to fail. Since we've mangled the stack
* beyond hope, there's not much we can do.
* (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
projects / sbcl.git / blobdiff