/* Copyright 2009-2017 * Kaz Kylheku * Vancouver, Canada * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include #include #include #include #include #include #include #include "config.h" #if HAVE_VALGRIND #include #endif #include "lib.h" #include "stream.h" #include "hash.h" #include "txr.h" #include "eval.h" #include "gc.h" #include "signal.h" #define PROT_STACK_SIZE 1024 #define HEAP_SIZE 16384 #define CHECKOBJ_VEC_SIZE (2 * HEAP_SIZE) #define MUTOBJ_VEC_SIZE (2 * HEAP_SIZE) #define FULL_GC_INTERVAL 40 #define FRESHOBJ_VEC_SIZE (8 * HEAP_SIZE) #define DFL_MALLOC_DELTA_THRESH (64L * 1024 * 1024) typedef struct heap { struct heap *next; obj_t block[HEAP_SIZE]; } heap_t; typedef struct mach_context { struct jmp buf; } mach_context_t; #define save_context(X) jmp_save(&(X).buf) int opt_gc_debug; #if HAVE_VALGRIND int opt_vg_debug; #endif static val *gc_stack_bottom; static val *prot_stack[PROT_STACK_SIZE]; static val **prot_stack_limit = prot_stack + PROT_STACK_SIZE; val **gc_prot_top = prot_stack; static val free_list, *free_tail = &free_list; static heap_t *heap_list; static val heap_min_bound, heap_max_bound; alloc_bytes_t gc_bytes; static alloc_bytes_t prev_malloc_bytes; alloc_bytes_t opt_gc_delta = DFL_MALLOC_DELTA_THRESH; int gc_enabled = 1; static int inprogress; static struct fin_reg { struct fin_reg *next; val obj; val fun; int reachable; } *final_list, **final_tail = &final_list; #if CONFIG_GEN_GC static val checkobj[CHECKOBJ_VEC_SIZE]; static int checkobj_idx; static val mutobj[MUTOBJ_VEC_SIZE]; static int mutobj_idx; static val freshobj[FRESHOBJ_VEC_SIZE]; static int freshobj_idx; int full_gc; #endif #if CONFIG_EXTRA_DEBUGGING val break_obj; #endif val prot1(val *loc) { assert (gc_prot_top < prot_stack_limit); assert (loc != 0); *gc_prot_top++ = loc; return nil; /* for use in macros */ } void protect(val *first, ...) { val *next = first; va_list vl; va_start (vl, first); while (next) { prot1(next); next = va_arg(vl, val *); } va_end (vl); } static void more(void) { heap_t *heap = coerce(heap_t *, chk_malloc_gc_more(sizeof *heap)); obj_t *block = heap->block, *end = heap->block + HEAP_SIZE; if (free_list == 0) free_tail = &heap->block[0].t.next; if (end > heap_max_bound) heap_max_bound = end; if (block < heap_min_bound) heap_min_bound = block; while (block < end) { block->t.next = free_list; block->t.type = convert(type_t, FREE); #if CONFIG_EXTRA_DEBUGGING if (block == break_obj) { #if HAVE_VALGRIND VALGRIND_PRINTF_BACKTRACE("object %p newly added to free list\n", convert(void *, block)); #endif breakpt(); } #endif free_list = block++; } heap->next = heap_list; heap_list = heap; #if HAVE_VALGRIND if (opt_vg_debug) VALGRIND_MAKE_MEM_NOACCESS(&heap->block, sizeof heap->block); #endif } val make_obj(void) { int tries; alloc_bytes_t malloc_delta = malloc_bytes - prev_malloc_bytes; assert (!async_sig_enabled); #if CONFIG_GEN_GC if ((opt_gc_debug || freshobj_idx >= FRESHOBJ_VEC_SIZE || malloc_delta >= opt_gc_delta) && gc_enabled) { gc(); } if (freshobj_idx >= FRESHOBJ_VEC_SIZE) full_gc = 1; #else if ((opt_gc_debug || malloc_delta >= opt_gc_delta) && gc_enabled) { gc(); } #endif for (tries = 0; tries < 3; tries++) { if (free_list) { val ret = free_list; #if HAVE_VALGRIND if (opt_vg_debug) VALGRIND_MAKE_MEM_DEFINED(free_list, sizeof *free_list); #endif free_list = free_list->t.next; assert (ret->t.type & FREE); if (free_list == 0) free_tail = &free_list; #if HAVE_VALGRIND if (opt_vg_debug) VALGRIND_MAKE_MEM_UNDEFINED(ret, sizeof *ret); #endif #if CONFIG_GEN_GC ret->t.gen = 0; if (!full_gc) freshobj[freshobj_idx++] = ret; #endif gc_bytes += sizeof (obj_t); #if CONFIG_EXTRA_DEBUGGING if (ret == break_obj) { #if HAVE_VALGRIND VALGRIND_PRINTF_BACKTRACE("object %p allocated\n", convert(void *, ret)); #endif breakpt(); } #endif return ret; } #if CONFIG_GEN_GC if (!full_gc && freshobj_idx < FRESHOBJ_VEC_SIZE) { more(); continue; } #endif switch (tries) { case 0: if (gc_enabled) { gc(); break; } /* fallthrough */ case 1: more(); break; } } abort(); } static void finalize(val obj) { switch (convert(type_t, obj->t.type)) { case NIL: case CONS: case CHR: case NUM: case LIT: case PKG: case FUN: case LCONS: case ENV: case FLNUM: case RNG: return; case SYM: free(obj->s.slot_cache); obj->s.slot_cache = 0; return; case STR: free(obj->st.str); obj->st.str = 0; return; case LSTR: free(obj->ls.props); obj->ls.props = 0; return; case VEC: free(obj->v.vec-2); obj->v.vec = 0; return; case COBJ: case CPTR: obj->co.ops->destroy(obj); obj->co.handle = 0; return; case BGNUM: mp_clear(mp(obj)); return; case BUF: if (obj->b.size) { free(obj->b.data); obj->b.data = 0; } return; } assert (0 && "corrupt type field"); } void cobj_destroy_stub_op(val obj) { } void cobj_destroy_free_op(val obj) { free(obj->co.handle); } static void mark_obj(val obj) { type_t t; #if 1 tail_call: #define mark_obj_tail(o) do { obj = (o); goto tail_call; } while (0) #else #define mark_obj_tail(o) return mark_obj(o) #endif if (!is_ptr(obj)) return; t = obj->t.type; if ((t & REACHABLE) != 0) return; #if CONFIG_GEN_GC if (!full_gc && obj->t.gen > 0) return; #endif if ((t & FREE) != 0) abort(); #if CONFIG_GEN_GC if (obj->t.gen == -1) obj->t.gen = 0; /* Will be promoted to generation 1 by sweep_one */ #endif obj->t.type = convert(type_t, t | REACHABLE); #if CONFIG_EXTRA_DEBUGGING if (obj == break_obj) { #if HAVE_VALGRIND VALGRIND_PRINTF_BACKTRACE("object %p marked\n", convert(void *, obj)); #endif breakpt(); } #endif switch (t) { case NIL: case CHR: case NUM: case LIT: case BGNUM: case FLNUM: return; case CONS: mark_obj(obj->c.car); mark_obj_tail(obj->c.cdr); case STR: mark_obj(obj->st.len); mark_obj_tail(obj->st.alloc); case SYM: mark_obj(obj->s.name); mark_obj_tail(obj->s.package); case PKG: mark_obj(obj->pk.name); mark_obj(obj->pk.hidhash); mark_obj_tail(obj->pk.symhash); case FUN: mark_obj(obj->f.env); if (obj->f.functype == FINTERP) mark_obj_tail(obj->f.f.interp_fun); return; case VEC: { val alloc_size = obj->v.vec[vec_alloc]; val len = obj->v.vec[vec_length]; cnum i, fp = c_num(len); mark_obj(alloc_size); mark_obj(len); for (i = 0; i < fp; i++) mark_obj(obj->v.vec[i]); } return; case LCONS: mark_obj(obj->lc.func); mark_obj(obj->lc.car); mark_obj_tail(obj->lc.cdr); case LSTR: mark_obj(obj->ls.prefix); mark_obj(obj->ls.props->limit); mark_obj(obj->ls.props->term); mark_obj_tail(obj->ls.list); case COBJ: case CPTR: obj->co.ops->mark(obj); mark_obj_tail(obj->co.cls); case ENV: mark_obj(obj->e.vbindings); mark_obj(obj->e.fbindings); mark_obj_tail(obj->e.up_env); case RNG: mark_obj(obj->rn.from); mark_obj_tail(obj->rn.to); case BUF: mark_obj(obj->b.len); mark_obj_tail(obj->b.size); } assert (0 && "corrupt type field"); } void cobj_mark_op(val obj) { } static int in_heap(val ptr) { heap_t *heap; if (!is_ptr(ptr)) return 0; if (ptr < heap_min_bound || ptr >= heap_max_bound) return 0; for (heap = heap_list; heap != 0; heap = heap->next) { if (ptr >= heap->block && ptr < heap->block + HEAP_SIZE) if ((coerce(char *, ptr) - coerce(char *, heap->block)) % sizeof (obj_t) == 0) return 1; } return 0; } static void mark_obj_maybe(val maybe_obj) { #if HAVE_VALGRIND VALGRIND_MAKE_MEM_DEFINED(&maybe_obj, sizeof maybe_obj); #endif if (in_heap(maybe_obj)) { #if HAVE_VALGRIND if (opt_vg_debug) VALGRIND_MAKE_MEM_DEFINED(maybe_obj, SIZEOF_PTR); #endif type_t t = maybe_obj->t.type; if ((t & FREE) == 0) { mark_obj(maybe_obj); } else { #if HAVE_VALGRIND if (opt_vg_debug) VALGRIND_MAKE_MEM_NOACCESS(maybe_obj, sizeof *maybe_obj); #endif } } } static void mark_mem_region(val *low, val *high) { if (low > high) { val *tmp = high; high = low; low = tmp; } for (; low < high; low++) mark_obj_maybe(*low); } static void mark(mach_context_t *pmc, val *gc_stack_top) { val **rootloc; /* * First, scan the officially registered locations. */ for (rootloc = prot_stack; rootloc != gc_prot_top; rootloc++) mark_obj(**rootloc); #if CONFIG_GEN_GC /* * Mark the additional objects indicated for marking. */ if (!full_gc) { int i; for (i = 0; i < checkobj_idx; i++) mark_obj(checkobj[i]); for (i = 0; i < mutobj_idx; i++) mark_obj(mutobj[i]); } #endif /* * Then the machine context */ mark_mem_region(coerce(val *, pmc), coerce(val *, (pmc + 1))); /* * Finally, the stack. */ mark_mem_region(gc_stack_top, gc_stack_bottom); } static int sweep_one(obj_t *block) { #if HAVE_VALGRIND const int vg_dbg = opt_vg_debug; #else const int vg_dbg = 0; #endif #if CONFIG_EXTRA_DEBUGGING if (block == break_obj) { #if HAVE_VALGRIND VALGRIND_PRINTF_BACKTRACE("object %p swept (type = %x)\n", convert(void *, block), convert(unsigned int, block->t.type)); #endif breakpt(); } #endif #if CONFIG_GEN_GC if (!full_gc && block->t.gen > 0) abort(); #endif if ((block->t.type & (REACHABLE | FREE)) == (REACHABLE | FREE)) abort(); if (block->t.type & REACHABLE) { #if CONFIG_GEN_GC block->t.gen = 1; #endif block->t.type = convert(type_t, block->t.type & ~REACHABLE); return 0; } if (block->t.type & FREE) { #if HAVE_VALGRIND if (vg_dbg) VALGRIND_MAKE_MEM_NOACCESS(block, sizeof *block); #endif return 1; } finalize(block); block->t.type = convert(type_t, block->t.type | FREE); /* If debugging is turned on, we want to catch instances where a reachable object is wrongly freed. This is difficult to do if the object is recycled soon after. So when debugging is on, the free list is FIFO rather than LIFO, which increases our chances that the code which is still using the object will trip on the freed object before it is recycled. */ if (vg_dbg || opt_gc_debug) { #if HAVE_VALGRIND if (vg_dbg && free_tail != &free_list) VALGRIND_MAKE_MEM_DEFINED(free_tail, sizeof *free_tail); #endif *free_tail = block; block->t.next = nil; #if HAVE_VALGRIND if (vg_dbg) { if (free_tail != &free_list) VALGRIND_MAKE_MEM_NOACCESS(free_tail, sizeof *free_tail); VALGRIND_MAKE_MEM_NOACCESS(block, sizeof *block); } #endif free_tail = &block->t.next; } else { block->t.next = free_list; free_list = block; } return 1; } static int_ptr_t sweep(void) { int_ptr_t free_count = 0; heap_t *heap; #if HAVE_VALGRIND const int vg_dbg = opt_vg_debug; #endif #if CONFIG_GEN_GC if (!full_gc) { int i; /* No need to mark block defined via Valgrind API; everything in the freshobj is an allocated node! */ for (i = 0; i < freshobj_idx; i++) free_count += sweep_one(freshobj[i]); /* Generation 1 objects that were indicated for dangerous mutation must have their REACHABLE flag flipped off, and must be returned to gen 1. */ for (i = 0; i < mutobj_idx; i++) sweep_one(mutobj[i]); return free_count; } #endif for (heap = heap_list; heap != 0; heap = heap->next) { obj_t *block, *end; #if HAVE_VALGRIND if (vg_dbg) VALGRIND_MAKE_MEM_DEFINED(&heap->block, sizeof heap->block); #endif for (block = heap->block, end = heap->block + HEAP_SIZE; block < end; block++) { free_count += sweep_one(block); } } return free_count; } static int is_reachable(val obj) { type_t t; #if CONFIG_GEN_GC if (!full_gc && obj->t.gen > 0) return 1; #endif t = obj->t.type; return (t & REACHABLE) != 0; } static void prepare_finals(void) { struct fin_reg *f; if (!final_list) return; for (f = final_list; f; f = f->next) f->reachable = is_reachable(f->obj); for (f = final_list; f; f = f->next) { if (!f->reachable) { #if CONFIG_GEN_GC f->obj->t.gen = 0; #endif mark_obj(f->obj); } mark_obj(f->fun); } } static val call_finalizers_impl(val ctx, int (*should_call)(struct fin_reg *, val)) { struct fin_reg *f, **tail; struct fin_reg *found = 0, **ftail = &found; val ret = nil; if (!final_list) return ret; for (f = final_list, tail = &final_list; f; ) { struct fin_reg *next = f->next; if (should_call(f, ctx)) { *ftail = f; ftail = &f->next; f->next = 0; } else { *tail = f; tail = &f->next; } f = next; } *tail = 0; final_tail = tail; while (found) { struct fin_reg *next = found->next; funcall1(found->fun, found->obj); free(found); found = next; ret = t; } return ret; } static int is_unreachable_final(struct fin_reg *f, val ctx) { (void) ctx; return !f->reachable; } static void call_finals(void) { (void) call_finalizers_impl(nil, is_unreachable_final); } void gc(void) { val gc_stack_top = nil; #if CONFIG_GEN_GC int exhausted = (free_list == 0); int full_gc_next_time = 0; static int gc_counter; #endif int swept; mach_context_t mc; assert (gc_enabled); if (inprogress++) assert(0 && "gc re-entered"); #if CONFIG_GEN_GC if (malloc_bytes - prev_malloc_bytes >= opt_gc_delta) full_gc = 1; #endif save_context(mc); gc_enabled = 0; rcyc_empty(); mark(&mc, &gc_stack_top); hash_process_weak(); prepare_finals(); swept = sweep(); #if CONFIG_GEN_GC #if 0 printf("sweep: freed %d full_gc == %d exhausted == %d\n", (int) swept, full_gc, exhausted); #endif if (++gc_counter >= FULL_GC_INTERVAL || freshobj_idx >= FRESHOBJ_VEC_SIZE) { full_gc_next_time = 1; gc_counter = 0; } if (exhausted && full_gc && swept < 3 * HEAP_SIZE / 4) more(); #else if (swept < 3 * HEAP_SIZE / 4) more(); #endif #if CONFIG_GEN_GC checkobj_idx = 0; mutobj_idx = 0; freshobj_idx = 0; full_gc = full_gc_next_time; #endif call_finals(); gc_enabled = 1; prev_malloc_bytes = malloc_bytes; inprogress--; } int gc_state(int enabled) { int old = gc_enabled; gc_enabled = enabled; return old; } int gc_inprogress(void) { return inprogress; } void gc_init(val *stack_bottom) { gc_stack_bottom = stack_bottom; } void gc_mark(val obj) { mark_obj(obj); } void gc_conservative_mark(val maybe_obj) { mark_obj_maybe(maybe_obj); } void gc_mark_mem(val *low, val *high) { mark_mem_region(low, high); } int gc_is_reachable(val obj) { return is_ptr(obj) ? is_reachable(obj) : 1; } #if CONFIG_GEN_GC val gc_set(loc lo, val obj) { val *ptr = valptr(lo); if (lo.obj && is_ptr(obj) && lo.obj->t.gen == 1 && obj->t.gen == 0 && !full_gc) { if (checkobj_idx < CHECKOBJ_VEC_SIZE) { obj->t.gen = -1; checkobj[checkobj_idx++] = obj; } else if (gc_enabled) { gc(); /* obj can't be in gen 0 because there are no baby objects after gc */ } else { /* We have no space to in checkobj record this backreference, and gc is not available to promote obj to gen 1. We must schedule a full gc. */ full_gc = 1; } } *ptr = obj; return obj; } val gc_mutated(val obj) { /* We care only about mature generation objects that have not already been noted. And if a full gc is coming, don't bother. */ if (full_gc || obj->t.gen <= 0) return obj; /* Store in mutobj array *before* triggering gc, otherwise baby objects referenced by obj could be reclaimed! */ if (mutobj_idx < MUTOBJ_VEC_SIZE) { obj->t.gen = -1; mutobj[mutobj_idx++] = obj; } else if (gc_enabled) { gc(); } else { full_gc = 1; } return obj; } val gc_push(val obj, loc plist) { return gc_set(plist, cons(obj, deref(plist))); } #endif static val gc_set_delta(val delta) { opt_gc_delta = c_num(delta); return nil; } static val gc_wrap(void) { if (gc_enabled) { gc(); return t; } return nil; } val gc_finalize(val obj, val fun, val rev_order_p) { type_check(fun, FUN); rev_order_p = default_null_arg(rev_order_p); if (is_ptr(obj)) { struct fin_reg *f = coerce(struct fin_reg *, chk_malloc(sizeof *f)); f->obj = obj; f->fun = fun; f->reachable = 0; if (rev_order_p) { if (!final_list) final_tail = &f->next; f->next = final_list; final_list = f; } else { f->next = 0; *final_tail = f; final_tail = &f->next; } } return obj; } static int is_matching_final(struct fin_reg *f, val obj) { return f->obj == obj; } val gc_call_finalizers(val obj) { return call_finalizers_impl(obj, is_matching_final); } val valid_object_p(val obj) { if (!is_ptr(obj)) return t; if (!in_heap(obj)) return nil; if (obj->t.type & (REACHABLE | FREE)) return nil; return t; } void gc_late_init(void) { reg_fun(intern(lit("gc"), system_package), func_n0(gc_wrap)); reg_fun(intern(lit("gc-set-delta"), system_package), func_n1(gc_set_delta)); reg_fun(intern(lit("finalize"), user_package), func_n3o(gc_finalize, 2)); reg_fun(intern(lit("call-finalizers"), user_package), func_n1(gc_call_finalizers)); } /* * Useful functions for gdb'ing. */ void unmark(void) { heap_t *heap; for (heap = heap_list; heap != 0; heap = heap->next) { val block, end; for (block = heap->block, end = heap->block + HEAP_SIZE; block < end; block++) { block->t.type = convert(type_t, block->t.type & ~REACHABLE); } } } void gc_cancel(void) { unmark(); #if CONFIG_GEN_GC checkobj_idx = 0; mutobj_idx = 0; freshobj_idx = 0; full_gc = 1; #endif inprogress = 0; } void dheap(heap_t *heap, int start, int end); void dheap(heap_t *heap, int start, int end) { int i; for (i = start; i < end; i++) format(std_output, lit("(~a ~s)\n"), num(i), &heap->block[i], nao); } /* * This function does nothing. * gc_hint(x) just takes the address of local variable x * and passes it to this function. This prevents the compiler * from caching the value across function calls. * This is needed for situations where * - a compiler caches a variable in a register, but not entirely (the variable * has a backing memory location); and * - that location contains a stale old value of the variable, which cannot be * garbage-collected as a result; and * - this causes a problem, like unbounded memory growth. */ void gc_hint_func(val *val) { (void) val; } void gc_report_copies(val *pvar) { val *opvar = pvar; val obj = *pvar++; for (; pvar < gc_stack_bottom; pvar++) { if (*pvar == obj) printf("%p found at %p (offset %d)\n", convert(void *, obj), convert(void *, pvar), convert(int, pvar - opvar)); } } void gc_free_all(void) { { heap_t *iter = heap_list; while (iter) { heap_t *next = iter->next; obj_t *block, *end; #if HAVE_VALGRIND if (opt_vg_debug) VALGRIND_MAKE_MEM_DEFINED(&iter->block, sizeof iter->block); #endif for (block = iter->block, end = iter->block + HEAP_SIZE; block < end; block++) { type_t t = block->t.type; if ((t & FREE) != 0) continue; finalize(block); } free(iter); iter = next; } } { struct fin_reg *iter = final_list; while (iter) { struct fin_reg *next = iter->next; free(iter); iter = next; } } }