X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fgencgc.c;h=a67748aba81d777df13b3bee71bdf6006adf6ffa;hb=df679ed627975948b1cee190f4d79c397588c43e;hp=aec65cd83de703f311dc9607e6956535a4744f89;hpb=2db3b6b4cb740d5b6512459c223859f747807b09;p=sbcl.git diff --git a/src/runtime/gencgc.c b/src/runtime/gencgc.c index aec65cd..a67748a 100644 --- a/src/runtime/gencgc.c +++ b/src/runtime/gencgc.c @@ -38,11 +38,16 @@ #include "arch.h" #include "gc.h" #include "gc-internal.h" +#include "thread.h" #include "genesis/vector.h" #include "genesis/weak-pointer.h" #include "genesis/simple-fun.h" -#include "genesis/static-symbols.h" -#include "genesis/symbol.h" + +#ifdef LISP_FEATURE_SB_THREAD +#include +#include /* threading is presently linux-only */ +#endif + /* assembly language stub that executes trap_PendingInterrupt */ void do_pending_interrupt(void); @@ -124,7 +129,8 @@ boolean gencgc_zero_check_during_free_heap = 0; /* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */ unsigned long bytes_allocated = 0; -static unsigned long auto_gc_trigger = 0; +extern unsigned long bytes_consed_between_gcs; /* gc-common.c */ +unsigned long auto_gc_trigger = 0; /* the source and destination generations. These are set before a GC starts * scavenging. */ @@ -247,6 +253,16 @@ unsigned int gencgc_oldest_gen_to_gc = NUM_GENERATIONS-1; * integrated with the Lisp code. */ static int last_free_page; +/* This lock is to prevent multiple threads from simultaneously + * allocating new regions which overlap each other. Note that the + * majority of GC is single-threaded, but alloc() may be called from + * >1 thread at a time and must be thread-safe. This lock must be + * seized before all accesses to generations[] or to parts of + * page_table[] that other threads may want to see */ + +static lispobj free_pages_lock=0; + + /* * miscellaneous heap functions */ @@ -490,7 +506,7 @@ gc_alloc_new_region(int nbytes, int unboxed, struct alloc_region *alloc_region) gc_assert((alloc_region->first_page == 0) && (alloc_region->last_page == -1) && (alloc_region->free_pointer == alloc_region->end_addr)); - + get_spinlock(&free_pages_lock,alloc_region); if (unboxed) { first_page = generations[gc_alloc_generation].alloc_unboxed_start_page; @@ -510,20 +526,6 @@ gc_alloc_new_region(int nbytes, int unboxed, struct alloc_region *alloc_region) alloc_region->free_pointer = alloc_region->start_addr; alloc_region->end_addr = alloc_region->start_addr + bytes_found; - if (gencgc_zero_check) { - int *p; - for (p = (int *)alloc_region->start_addr; - p < (int *)alloc_region->end_addr; p++) { - if (*p != 0) { - /* KLUDGE: It would be nice to use %lx and explicit casts - * (long) in code like this, so that it is less likely to - * break randomly when running on a machine with different - * word sizes. -- WHN 19991129 */ - lose("The new region at %x is not zero.", p); - } - } - } - /* Set up the pages. */ /* The first page may have already been in use. */ @@ -559,15 +561,32 @@ gc_alloc_new_region(int nbytes, int unboxed, struct alloc_region *alloc_region) alloc_region->start_addr - page_address(i); page_table[i].allocated |= OPEN_REGION_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*4096)); + (lispobj)(((char *)heap_base) + last_free_page*4096), + 0); + } + free_pages_lock=0; + + /* we can do this after releasing free_pages_lock */ + if (gencgc_zero_check) { + int *p; + for (p = (int *)alloc_region->start_addr; + p < (int *)alloc_region->end_addr; p++) { + if (*p != 0) { + /* KLUDGE: It would be nice to use %lx and explicit casts + * (long) in code like this, so that it is less likely to + * break randomly when running on a machine with different + * word sizes. -- WHN 19991129 */ + lose("The new region at %x is not zero.", p); + } } } +} + /* If the record_new_objects flag is 2 then all new regions created * are recorded. * @@ -691,8 +710,9 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) next_page = first_page+1; - /* Skip if no bytes were allocated. */ + get_spinlock(&free_pages_lock,alloc_region); 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; gc_assert(alloc_region->start_addr == (page_address(first_page) + page_table[first_page].bytes_used)); @@ -794,7 +814,8 @@ gc_alloc_update_page_tables(int unboxed, struct alloc_region *alloc_region) page_table[next_page].allocated = FREE_PAGE; next_page++; } - + free_pages_lock=0; + /* alloc_region is per-thread, we're ok to do this unlocked */ gc_set_region_empty(alloc_region); } @@ -836,6 +857,8 @@ gc_alloc_large(int nbytes, int unboxed, struct alloc_region *alloc_region) index ahead of the current region and bumped up here to save a lot of re-scanning. */ + get_spinlock(&free_pages_lock,alloc_region); + if (unboxed) { first_page = generations[gc_alloc_generation].alloc_large_unboxed_start_page; @@ -932,8 +955,9 @@ gc_alloc_large(int nbytes, int unboxed, struct alloc_region *alloc_region) if (last_page+1 > last_free_page) { last_free_page = last_page+1; SetSymbolValue(ALLOCATION_POINTER, - (lispobj)(((char *)heap_base) + last_free_page*4096)); + (lispobj)(((char *)heap_base) + last_free_page*4096),0); } + free_pages_lock=0; return((void *)(page_address(first_page)+orig_first_page_bytes_used)); } @@ -951,6 +975,7 @@ gc_find_freeish_pages(int *restart_page_ptr, int nbytes, int unboxed, struct all int num_pages; int large = !alloc_region && (nbytes >= large_object_size); + gc_assert(free_pages_lock); /* Search for a contiguous free space of at least nbytes. If it's a large object then align it on a page boundary by searching for a free page. */ @@ -2088,7 +2113,7 @@ static lispobj* search_read_only_space(lispobj *pointer) { lispobj* start = (lispobj*)READ_ONLY_SPACE_START; - lispobj* end = (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER); + lispobj* end = (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0); if ((pointer < start) || (pointer >= end)) return NULL; return (search_space(start, (pointer+2)-start, pointer)); @@ -2098,7 +2123,7 @@ static lispobj * search_static_space(lispobj *pointer) { lispobj* start = (lispobj*)STATIC_SPACE_START; - lispobj* end = (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER); + lispobj* end = (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0); if ((pointer < start) || (pointer >= end)) return NULL; return (search_space(start, (pointer+2)-start, pointer)); @@ -2122,7 +2147,8 @@ search_dynamic_space(lispobj *pointer) /* Is there any possibility that pointer is a valid Lisp object * reference, and/or something else (e.g. subroutine call return - * address) which should prevent us from moving the referred-to thing? */ + * address) which should prevent us from moving the referred-to thing? + * This is called from preserve_pointers() */ static int possibly_valid_dynamic_space_pointer(lispobj *pointer) { @@ -2149,21 +2175,7 @@ possibly_valid_dynamic_space_pointer(lispobj *pointer) /* Check that the object pointed to is consistent with the pointer * low tag. - * - * FIXME: It's not safe to rely on the result from this check - * before an object is initialized. Thus, if we were interrupted - * just as an object had been allocated but not initialized, the - * GC relying on this result could bogusly reclaim the memory. - * However, we can't really afford to do without this check. So - * we should make it safe somehow. - * (1) Perhaps just review the code to make sure - * that WITHOUT-GCING or WITHOUT-INTERRUPTS or some such - * thing is wrapped around critical sections where allocated - * memory type bits haven't been set. - * (2) Perhaps find some other hack to protect against this, e.g. - * recording the result of the last call to allocate-lisp-memory, - * and returning true from this function when *pointer is - * a reference to that result. */ + */ switch (lowtag_of((lispobj)pointer)) { case FUN_POINTER_LOWTAG: /* Start_addr should be the enclosing code object, or a closure @@ -2560,7 +2572,7 @@ preserve_pointer(void *addr) * (or, as a special case which also requires dont_move, a return * address referring to something in a CodeObject). This is * expensive but important, since it vastly reduces the - * probability that random garbage will be bogusly interpreter as + * probability that random garbage will be bogusly interpreted as * a pointer which prevents a page from moving. */ if (!(possibly_valid_dynamic_space_pointer(addr))) return; @@ -3231,7 +3243,7 @@ verify_space(lispobj *start, size_t words) int is_in_dynamic_space = (find_page_index((void*)start) != -1); int is_in_readonly_space = (READ_ONLY_SPACE_START <= (unsigned)start && - (unsigned)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER)); + (unsigned)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)); while (words > 0) { size_t count = 1; @@ -3241,10 +3253,10 @@ verify_space(lispobj *start, size_t words) int page_index = find_page_index((void*)thing); int to_readonly_space = (READ_ONLY_SPACE_START <= thing && - thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER)); + thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)); int to_static_space = (STATIC_SPACE_START <= thing && - thing < SymbolValue(STATIC_SPACE_FREE_POINTER)); + thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0)); /* Does it point to the dynamic space? */ if (page_index != -1) { @@ -3439,18 +3451,20 @@ verify_gc(void) * to grep for all foo_size and rename the appropriate ones to * foo_count. */ int read_only_space_size = - (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER) + (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0) - (lispobj*)READ_ONLY_SPACE_START; int static_space_size = - (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER) + (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0) - (lispobj*)STATIC_SPACE_START; + struct thread *th; + for_each_thread(th) { int binding_stack_size = - (lispobj*)SymbolValue(BINDING_STACK_POINTER) - - (lispobj*)BINDING_STACK_START; - + (lispobj*)SymbolValue(BINDING_STACK_POINTER,th) + - (lispobj*)th->binding_stack_start; + verify_space(th->binding_stack_start, binding_stack_size); + } verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size); verify_space((lispobj*)STATIC_SPACE_START , static_space_size); - verify_space((lispobj*)BINDING_STACK_START , binding_stack_size); } static void @@ -3588,7 +3602,7 @@ garbage_collect_generation(int generation, int raise) unsigned long bytes_freed; unsigned long i; unsigned long static_space_size; - + struct thread *th; gc_assert(generation <= (NUM_GENERATIONS-1)); /* The oldest generation can't be raised. */ @@ -3630,11 +3644,30 @@ garbage_collect_generation(int generation, int raise) * be un-protected anyway before unmapping later. */ unprotect_oldspace(); - /* Scavenge the stack's conservative roots. */ - { + /* Scavenge the stacks' conservative roots. */ + for_each_thread(th) { void **ptr; - for (ptr = (void **)CONTROL_STACK_END - 1; +#ifdef LISP_FEATURE_SB_THREAD + struct user_regs_struct regs; + if(ptrace(PTRACE_GETREGS,th->pid,0,®s)){ + /* probably doesn't exist any more. */ + fprintf(stderr,"child pid %d, %s\n",th->pid,strerror(errno)); + perror("PTRACE_GETREGS"); + } + preserve_pointer(regs.ebx); + preserve_pointer(regs.ecx); + preserve_pointer(regs.edx); + preserve_pointer(regs.esi); + preserve_pointer(regs.edi); + preserve_pointer(regs.ebp); + preserve_pointer(regs.eax); +#endif + for (ptr = th->control_stack_end; +#ifdef LISP_FEATURE_SB_THREAD + ptr > regs.esp; +#else ptr > (void **)&raise; +#endif ptr--) { preserve_pointer(*ptr); } @@ -3656,18 +3689,31 @@ garbage_collect_generation(int generation, int raise) /* Scavenge the Lisp functions of the interrupt handlers, taking * care to avoid SIG_DFL and SIG_IGN. */ + for_each_thread(th) { + struct interrupt_data *data=th->interrupt_data; for (i = 0; i < NSIG; i++) { - union interrupt_handler handler = interrupt_handlers[i]; + union interrupt_handler handler = data->interrupt_handlers[i]; if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) && !ARE_SAME_HANDLER(handler.c, SIG_DFL)) { - scavenge((lispobj *)(interrupt_handlers + i), 1); + scavenge((lispobj *)(data->interrupt_handlers + i), 1); + } + } + } + /* Scavenge the binding stacks. */ + { + struct thread *th; + for_each_thread(th) { + long len= (lispobj *)SymbolValue(BINDING_STACK_POINTER,th) - + th->binding_stack_start; + scavenge((lispobj *) th->binding_stack_start,len); +#ifdef LISP_FEATURE_SB_THREAD + /* do the tls as well */ + len=fixnum_value(SymbolValue(FREE_TLS_INDEX,0)) - + (sizeof (struct thread))/(sizeof (lispobj)); + scavenge((lispobj *) (th+1),len); +#endif } } - - /* Scavenge the binding stack. */ - scavenge((lispobj *) BINDING_STACK_START, - (lispobj *)SymbolValue(BINDING_STACK_POINTER) - - (lispobj *)BINDING_STACK_START); /* The original CMU CL code had scavenge-read-only-space code * controlled by the Lisp-level variable @@ -3690,7 +3736,7 @@ garbage_collect_generation(int generation, int raise) /* Scavenge static space. */ static_space_size = - (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER) - + (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) - (lispobj *)STATIC_SPACE_START; if (gencgc_verbose > 1) { FSHOW((stderr, @@ -3801,7 +3847,7 @@ update_x86_dynamic_space_free_pointer(void) last_free_page = last_page+1; SetSymbolValue(ALLOCATION_POINTER, - (lispobj)(((char *)heap_base) + last_free_page*4096)); + (lispobj)(((char *)heap_base) + last_free_page*4096),0); return 0; /* dummy value: return something ... */ } @@ -3923,7 +3969,10 @@ collect_garbage(unsigned last_gen) gc_alloc_generation = 0; update_x86_dynamic_space_free_pointer(); - + auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs; + if(gencgc_verbose) + fprintf(stderr,"Next gc when %d bytes have been consed\n", + auto_gc_trigger); SHOW("returning from collect_garbage"); } @@ -4005,7 +4054,7 @@ gc_free_heap(void) gc_set_region_empty(&unboxed_region); last_free_page = 0; - SetSymbolValue(ALLOCATION_POINTER, (lispobj)((char *)heap_base)); + SetSymbolValue(ALLOCATION_POINTER, (lispobj)((char *)heap_base),0); if (verify_after_free_heap) { /* Check whether purify has left any bad pointers. */ @@ -4076,7 +4125,7 @@ gencgc_pickup_dynamic(void) { int page = 0; int addr = DYNAMIC_SPACE_START; - int alloc_ptr = SymbolValue(ALLOCATION_POINTER); + int alloc_ptr = SymbolValue(ALLOCATION_POINTER,0); /* Initialize the first region. */ do { @@ -4120,19 +4169,33 @@ extern boolean maybe_gc_pending ; char * alloc(int nbytes) { - struct alloc_region *region= &boxed_region; + struct thread *th=arch_os_get_current_thread(); + struct alloc_region *region= + th ? &(th->alloc_region) : &boxed_region; void *new_obj; void *new_free_pointer; /* Check for alignment allocation problems. */ gc_assert((((unsigned)region->free_pointer & 0x7) == 0) && ((nbytes & 0x7) == 0)); - /* At this point we should either be in pseudo-atomic, or early - * enough in cold initn that interrupts are not yet enabled anyway. - * It would be nice to assert same. - */ - gc_assert(SymbolValue(PSEUDO_ATOMIC_ATOMIC)); - + 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)) { + register u32 fs; + fprintf(stderr, "fatal error in thread 0x%x, pid=%d\n", + th,getpid()); + __asm__("movl %fs,%0" : "=r" (fs) : ); + fprintf(stderr, "fs is %x, th->tls_cookie=%x (should be identical)\n", + debug_get_fs(),th->tls_cookie); + lose("If you see this message before 2003.05.01, mail details to sbcl-devel\n"); + } +#else + gc_assert(SymbolValue(PSEUDO_ATOMIC_ATOMIC,th)); +#endif + /* maybe we can do this quickly ... */ new_free_pointer = region->free_pointer + nbytes; if (new_free_pointer <= region->end_addr) { @@ -4145,33 +4208,16 @@ alloc(int nbytes) * we should GC in the near future */ if (auto_gc_trigger && bytes_allocated > auto_gc_trigger) { - auto_gc_trigger *= 2; /* set things up so that GC happens when we finish the PA * section. */ maybe_gc_pending=1; - SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED, make_fixnum(1)); + SetSymbolValue(PSEUDO_ATOMIC_INTERRUPTED, make_fixnum(1),th); } new_obj = gc_alloc_with_region(nbytes,0,region,0); return (new_obj); } -/* - * noise to manipulate the gc trigger stuff - */ - -void -set_auto_gc_trigger(os_vm_size_t dynamic_usage) -{ - auto_gc_trigger += dynamic_usage; -} - -void -clear_auto_gc_trigger(void) -{ - auto_gc_trigger = 0; -} - /* Find the code object for the given pc, or return NULL on failure. * * FIXME: PC shouldn't be lispobj*, should it? Maybe void*? */ @@ -4232,23 +4278,26 @@ gencgc_handle_wp_violation(void* fault_addr) return 0; } else { - - /* The only acceptable reason for an signal like this from the - * heap is that the generational GC write-protected the page. */ - if (page_table[page_index].write_protected != 1) { - lose("access failure in heap page not marked as write-protected"); + if (page_table[page_index].write_protected) { + /* Unprotect the page. */ + os_protect(page_address(page_index), PAGE_BYTES, OS_VM_PROT_ALL); + page_table[page_index].write_protected_cleared = 1; + page_table[page_index].write_protected = 0; + } else { + /* The only acceptable reason for this signal on a heap + * access is that GENCGC write-protected the page. + * However, if two CPUs hit a wp page near-simultaneously, + * we had better not have the second one lose here if it + * 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"); + + /* Don't worry, we can handle it. */ + return 1; } - - /* Unprotect the page. */ - os_protect(page_address(page_index), 4096, OS_VM_PROT_ALL); - page_table[page_index].write_protected = 0; - page_table[page_index].write_protected_cleared = 1; - - /* Don't worry, we can handle it. */ - return 1; } } - /* This is to be called when we catch a SIGSEGV/SIGBUS, determine that * it's not just a case of the program hitting the write barrier, and * are about to let Lisp deal with it. It's basically just a @@ -4260,6 +4309,9 @@ unhandled_sigmemoryfault() gc_alloc_update_all_page_tables(void) { /* Flush the alloc regions updating the tables. */ + struct thread *th; + for_each_thread(th) + gc_alloc_update_page_tables(0, &th->alloc_region); gc_alloc_update_page_tables(1, &unboxed_region); gc_alloc_update_page_tables(0, &boxed_region); }