X-Git-Url: http://git.sourceforge.jp/view?p=pf3gnuchains%2Fgcc-fork.git;a=blobdiff_plain;f=boehm-gc%2Fos_dep.c;h=3f0f20e834cedda24483ba09fb3d93222b2497af;hp=a7f6d1940054120b3b6ad2ed1ec63f816849549e;hb=fda615954709a34a26d4adfb780c8cebb288cda2;hpb=f8c1da573b1b2b72501630f18fc1452e6b9e9c0c diff --git a/boehm-gc/os_dep.c b/boehm-gc/os_dep.c index a7f6d194005..3f0f20e834c 100644 --- a/boehm-gc/os_dep.c +++ b/boehm-gc/os_dep.c @@ -60,14 +60,18 @@ # include # endif +#if defined(LINUX) || defined(LINUX_STACKBOTTOM) +# include +#endif + /* Blatantly OS dependent routines, except for those that are related */ /* to dynamic loading. */ -# if !defined(THREADS) && !defined(STACKBOTTOM) && defined(HEURISTIC2) +# if defined(HEURISTIC2) || defined(SEARCH_FOR_DATA_START) # define NEED_FIND_LIMIT # endif -# if defined(IRIX_THREADS) || defined(HPUX_THREADS) +# if !defined(STACKBOTTOM) && defined(HEURISTIC2) # define NEED_FIND_LIMIT # endif @@ -75,22 +79,26 @@ # define NEED_FIND_LIMIT # endif -# if (defined(SVR4) || defined(AUX) || defined(DGUX)) && !defined(PCR) +# if (defined(SVR4) || defined(AUX) || defined(DGUX) \ + || (defined(LINUX) && defined(SPARC))) && !defined(PCR) # define NEED_FIND_LIMIT # endif -# if defined(LINUX) && \ - (defined(POWERPC) || defined(SPARC) || defined(ALPHA) || defined(IA64) \ - || defined(MIPS)) -# define NEED_FIND_LIMIT -# endif +#if defined(FREEBSD) && (defined(I386) || defined(X86_64) || defined(powerpc) || defined(__powerpc__)) +# include +# if !defined(PCR) +# define NEED_FIND_LIMIT +# endif +#endif -#ifdef NEED_FIND_LIMIT -# include +#if (defined(NETBSD) || defined(OPENBSD)) && defined(__ELF__) \ + && !defined(NEED_FIND_LIMIT) + /* Used by GC_init_netbsd_elf() below. */ +# define NEED_FIND_LIMIT #endif -#ifdef FREEBSD -# include +#ifdef NEED_FIND_LIMIT +# include #endif #ifdef AMIGA @@ -113,23 +121,35 @@ # include # include /* for locking */ #endif -#ifdef USE_MMAP +#if defined(USE_MMAP) || defined(USE_MUNMAP) +# ifndef USE_MMAP + --> USE_MUNMAP requires USE_MMAP +# endif # include # include # include +# include #endif #ifdef UNIX_LIKE # include +# if defined(SUNOS5SIGS) && !defined(FREEBSD) +# include +# endif + /* Define SETJMP and friends to be the version that restores */ + /* the signal mask. */ +# define SETJMP(env) sigsetjmp(env, 1) +# define LONGJMP(env, val) siglongjmp(env, val) +# define JMP_BUF sigjmp_buf +#else +# define SETJMP(env) setjmp(env) +# define LONGJMP(env, val) longjmp(env, val) +# define JMP_BUF jmp_buf #endif -#ifdef SUNOS5SIGS -# include -# undef setjmp -# undef longjmp -# define setjmp(env) sigsetjmp(env, 1) -# define longjmp(env, val) siglongjmp(env, val) -# define jmp_buf sigjmp_buf +#ifdef DARWIN +/* for get_etext and friends */ +#include #endif #ifdef DJGPP @@ -150,6 +170,142 @@ # define OPT_PROT_EXEC 0 #endif +#if defined(LINUX) && \ + (defined(USE_PROC_FOR_LIBRARIES) || defined(IA64) || !defined(SMALL_CONFIG)) + +/* We need to parse /proc/self/maps, either to find dynamic libraries, */ +/* and/or to find the register backing store base (IA64). Do it once */ +/* here. */ + +#define READ read + +/* Repeatedly perform a read call until the buffer is filled or */ +/* we encounter EOF. */ +ssize_t GC_repeat_read(int fd, char *buf, size_t count) +{ + ssize_t num_read = 0; + ssize_t result; + + while (num_read < count) { + result = READ(fd, buf + num_read, count - num_read); + if (result < 0) return result; + if (result == 0) break; + num_read += result; + } + return num_read; +} + +/* + * Apply fn to a buffer containing the contents of /proc/self/maps. + * Return the result of fn or, if we failed, 0. + * We currently do nothing to /proc/self/maps other than simply read + * it. This code could be simplified if we could determine its size + * ahead of time. + */ + +word GC_apply_to_maps(word (*fn)(char *)) +{ + int f; + int result; + size_t maps_size = 4000; /* Initial guess. */ + static char init_buf[1]; + static char *maps_buf = init_buf; + static size_t maps_buf_sz = 1; + + /* Read /proc/self/maps, growing maps_buf as necessary. */ + /* Note that we may not allocate conventionally, and */ + /* thus can't use stdio. */ + do { + if (maps_size >= maps_buf_sz) { + /* Grow only by powers of 2, since we leak "too small" buffers. */ + while (maps_size >= maps_buf_sz) maps_buf_sz *= 2; + maps_buf = GC_scratch_alloc(maps_buf_sz); + if (maps_buf == 0) return 0; + } + f = open("/proc/self/maps", O_RDONLY); + if (-1 == f) return 0; + maps_size = 0; + do { + result = GC_repeat_read(f, maps_buf, maps_buf_sz-1); + if (result <= 0) return 0; + maps_size += result; + } while (result == maps_buf_sz-1); + close(f); + } while (maps_size >= maps_buf_sz); + maps_buf[maps_size] = '\0'; + + /* Apply fn to result. */ + return fn(maps_buf); +} + +#endif /* Need GC_apply_to_maps */ + +#if defined(LINUX) && (defined(USE_PROC_FOR_LIBRARIES) || defined(IA64)) +// +// GC_parse_map_entry parses an entry from /proc/self/maps so we can +// locate all writable data segments that belong to shared libraries. +// The format of one of these entries and the fields we care about +// is as follows: +// XXXXXXXX-XXXXXXXX r-xp 00000000 30:05 260537 name of mapping...\n +// ^^^^^^^^ ^^^^^^^^ ^^^^ ^^ +// start end prot maj_dev +// +// Note that since about auguat 2003 kernels, the columns no longer have +// fixed offsets on 64-bit kernels. Hence we no longer rely on fixed offsets +// anywhere, which is safer anyway. +// + +/* + * Assign various fields of the first line in buf_ptr to *start, *end, + * *prot_buf and *maj_dev. Only *prot_buf may be set for unwritable maps. + */ +char *GC_parse_map_entry(char *buf_ptr, word *start, word *end, + char *prot_buf, unsigned int *maj_dev) +{ + char *start_start, *end_start, *prot_start, *maj_dev_start; + char *p; + char *endp; + + if (buf_ptr == NULL || *buf_ptr == '\0') { + return NULL; + } + + p = buf_ptr; + while (isspace(*p)) ++p; + start_start = p; + GC_ASSERT(isxdigit(*start_start)); + *start = strtoul(start_start, &endp, 16); p = endp; + GC_ASSERT(*p=='-'); + + ++p; + end_start = p; + GC_ASSERT(isxdigit(*end_start)); + *end = strtoul(end_start, &endp, 16); p = endp; + GC_ASSERT(isspace(*p)); + + while (isspace(*p)) ++p; + prot_start = p; + GC_ASSERT(*prot_start == 'r' || *prot_start == '-'); + memcpy(prot_buf, prot_start, 4); + prot_buf[4] = '\0'; + if (prot_buf[1] == 'w') {/* we can skip the rest if it's not writable. */ + /* Skip past protection field to offset field */ + while (!isspace(*p)) ++p; while (isspace(*p)) ++p; + GC_ASSERT(isxdigit(*p)); + /* Skip past offset field, which we ignore */ + while (!isspace(*p)) ++p; while (isspace(*p)) ++p; + maj_dev_start = p; + GC_ASSERT(isxdigit(*maj_dev_start)); + *maj_dev = strtoul(maj_dev_start, NULL, 16); + } + + while (*p && *p++ != '\n'); + + return p; +} + +#endif /* Need to parse /proc/self/maps. */ + #if defined(SEARCH_FOR_DATA_START) /* The I386 case can be handled without a search. The Alpha case */ /* used to be handled differently as well, but the rules changed */ @@ -157,12 +313,17 @@ /* cover all versions. */ # ifdef LINUX + /* Some Linux distributions arrange to define __data_start. Some */ + /* define data_start as a weak symbol. The latter is technically */ + /* broken, since the user program may define data_start, in which */ + /* case we lose. Nonetheless, we try both, prefering __data_start. */ + /* We assume gcc-compatible pragmas. */ # pragma weak __data_start - extern int __data_start; + extern int __data_start[]; # pragma weak data_start - extern int data_start; + extern int data_start[]; # endif /* LINUX */ - extern int _end; + extern int _end[]; ptr_t GC_data_start; @@ -172,16 +333,16 @@ # ifdef LINUX /* Try the easy approaches first: */ - if (&__data_start != 0) { - GC_data_start = (ptr_t)(&__data_start); + if ((ptr_t)__data_start != 0) { + GC_data_start = (ptr_t)(__data_start); return; } - if (&data_start != 0) { - GC_data_start = (ptr_t)(&data_start); + if ((ptr_t)data_start != 0) { + GC_data_start = (ptr_t)(data_start); return; } # endif /* LINUX */ - GC_data_start = GC_find_limit((ptr_t)(&_end), FALSE); + GC_data_start = GC_find_limit((ptr_t)(_end), FALSE); } #endif @@ -192,7 +353,8 @@ # endif /* ECOS_GC_MEMORY_SIZE */ // setjmp() function, as described in ANSI para 7.6.1.1 -#define setjmp( __env__ ) hal_setjmp( __env__ ) +#undef SETJMP +#define SETJMP( __env__ ) hal_setjmp( __env__ ) // FIXME: This is a simple way of allocating memory which is // compatible with ECOS early releases. Later releases use a more @@ -218,7 +380,7 @@ static void *tiny_sbrk(ptrdiff_t increment) #define sbrk tiny_sbrk # endif /* ECOS */ -#if defined(NETBSD) && defined(__ELF__) +#if (defined(NETBSD) || defined(OPENBSD)) && defined(__ELF__) ptr_t GC_data_start; void GC_init_netbsd_elf() @@ -336,9 +498,10 @@ void GC_enable_signals(void) # if !defined(PCR) && !defined(AMIGA) && !defined(MSWIN32) \ && !defined(MSWINCE) \ - && !defined(MACOS) && !defined(DJGPP) && !defined(DOS4GW) + && !defined(MACOS) && !defined(DJGPP) && !defined(DOS4GW) \ + && !defined(NOSYS) && !defined(ECOS) -# if defined(sigmask) && !defined(UTS4) +# if defined(sigmask) && !defined(UTS4) && !defined(HURD) /* Use the traditional BSD interface */ # define SIGSET_T int # define SIG_DEL(set, signal) (set) &= ~(sigmask(signal)) @@ -527,9 +690,11 @@ ptr_t GC_get_stack_base() typedef void (*handler)(); # endif -# if defined(SUNOS5SIGS) || defined(IRIX5) || defined(OSF1) +# if defined(SUNOS5SIGS) || defined(IRIX5) || defined(OSF1) \ + || defined(HURD) || defined(NETBSD) static struct sigaction old_segv_act; -# if defined(_sigargs) || defined(HPUX) /* !Irix6.x */ +# if defined(IRIX5) || defined(HPUX) \ + || defined(HURD) || defined(NETBSD) static struct sigaction old_bus_act; # endif # else @@ -543,54 +708,55 @@ ptr_t GC_get_stack_base() handler h; # endif { -# ifndef ECOS -# if defined(SUNOS5SIGS) || defined(IRIX5) || defined(OSF1) +# if defined(SUNOS5SIGS) || defined(IRIX5) \ + || defined(OSF1) || defined(HURD) || defined(NETBSD) struct sigaction act; act.sa_handler = h; - act.sa_flags = SA_RESTART | SA_NODEFER; - /* The presence of SA_NODEFER represents yet another gross */ - /* hack. Under Solaris 2.3, siglongjmp doesn't appear to */ - /* interact correctly with -lthread. We hide the confusion */ - /* by making sure that signal handling doesn't affect the */ - /* signal mask. */ +# if 0 /* Was necessary for Solaris 2.3 and very temporary */ + /* NetBSD bugs. */ + act.sa_flags = SA_RESTART | SA_NODEFER; +# else + act.sa_flags = SA_RESTART; +# endif (void) sigemptyset(&act.sa_mask); -# ifdef IRIX_THREADS +# ifdef GC_IRIX_THREADS /* Older versions have a bug related to retrieving and */ /* and setting a handler at the same time. */ (void) sigaction(SIGSEGV, 0, &old_segv_act); (void) sigaction(SIGSEGV, &act, 0); + (void) sigaction(SIGBUS, 0, &old_bus_act); + (void) sigaction(SIGBUS, &act, 0); # else (void) sigaction(SIGSEGV, &act, &old_segv_act); -# if defined(IRIX5) && defined(_sigargs) /* Irix 5.x, not 6.x */ \ - || defined(HPUX) +# if defined(IRIX5) \ + || defined(HPUX) || defined(HURD) || defined(NETBSD) /* Under Irix 5.x or HP/UX, we may get SIGBUS. */ /* Pthreads doesn't exist under Irix 5.x, so we */ /* don't have to worry in the threads case. */ (void) sigaction(SIGBUS, &act, &old_bus_act); # endif -# endif /* IRIX_THREADS */ +# endif /* GC_IRIX_THREADS */ # else old_segv_handler = signal(SIGSEGV, h); # ifdef SIGBUS old_bus_handler = signal(SIGBUS, h); # endif # endif -# endif /* ECOS */ } # endif /* NEED_FIND_LIMIT || UNIX_LIKE */ # ifdef NEED_FIND_LIMIT /* Some tools to implement HEURISTIC2 */ # define MIN_PAGE_SIZE 256 /* Smallest conceivable page size, bytes */ - /* static */ jmp_buf GC_jmp_buf; + /* static */ JMP_BUF GC_jmp_buf; /*ARGSUSED*/ void GC_fault_handler(sig) int sig; { - longjmp(GC_jmp_buf, 1); + LONGJMP(GC_jmp_buf, 1); } void GC_setup_temporary_fault_handler() @@ -600,11 +766,11 @@ ptr_t GC_get_stack_base() void GC_reset_fault_handler() { -# ifndef ECOS -# if defined(SUNOS5SIGS) || defined(IRIX5) || defined(OSF1) +# if defined(SUNOS5SIGS) || defined(IRIX5) \ + || defined(OSF1) || defined(HURD) || defined(NETBSD) (void) sigaction(SIGSEGV, &old_segv_act, 0); -# if defined(IRIX5) && defined(_sigargs) /* Irix 5.x, not 6.x */ \ - || defined(HPUX) +# if defined(IRIX5) \ + || defined(HPUX) || defined(HURD) || defined(NETBSD) (void) sigaction(SIGBUS, &old_bus_act, 0); # endif # else @@ -613,16 +779,15 @@ ptr_t GC_get_stack_base() (void) signal(SIGBUS, old_bus_handler); # endif # endif -# endif /* ECOS */ } /* Return the first nonaddressible location > p (up) or */ - /* the smallest location q s.t. [q,p] is addressible (!up). */ + /* the smallest location q s.t. [q,p) is addressable (!up). */ + /* We assume that p (up) or p-1 (!up) is addressable. */ ptr_t GC_find_limit(p, up) ptr_t p; GC_bool up; { -# ifndef ECOS static VOLATILE ptr_t result; /* Needs to be static, since otherwise it may not be */ /* preserved across the longjmp. Can safely be */ @@ -631,7 +796,7 @@ ptr_t GC_get_stack_base() GC_setup_temporary_fault_handler(); - if (setjmp(GC_jmp_buf) == 0) { + if (SETJMP(GC_jmp_buf) == 0) { result = (ptr_t)(((word)(p)) & ~(MIN_PAGE_SIZE-1)); for (;;) { @@ -648,13 +813,38 @@ ptr_t GC_get_stack_base() result += MIN_PAGE_SIZE; } return(result); -# else /* ECOS */ - abort(); -# endif /* ECOS */ } # endif -# ifndef ECOS +#if defined(ECOS) || defined(NOSYS) + ptr_t GC_get_stack_base() + { + return STACKBOTTOM; + } +#endif + +#ifdef HPUX_STACKBOTTOM + +#include +#include + + ptr_t GC_get_register_stack_base(void) + { + struct pst_vm_status vm_status; + + int i = 0; + while (pstat_getprocvm(&vm_status, sizeof(vm_status), 0, i++) == 1) { + if (vm_status.pst_type == PS_RSESTACK) { + return (ptr_t) vm_status.pst_vaddr; + } + } + + /* old way to get the register stackbottom */ + return (ptr_t)(((word)GC_stackbottom - BACKING_STORE_DISPLACEMENT - 1) + & ~(BACKING_STORE_ALIGNMENT - 1)); + } + +#endif /* HPUX_STACK_BOTTOM */ #ifdef LINUX_STACKBOTTOM @@ -664,23 +854,66 @@ ptr_t GC_get_stack_base() # define STAT_SKIP 27 /* Number of fields preceding startstack */ /* field in /proc/self/stat */ +#ifdef USE_LIBC_PRIVATES # pragma weak __libc_stack_end extern ptr_t __libc_stack_end; +#endif # ifdef IA64 -# pragma weak __libc_ia64_register_backing_store_base - extern ptr_t __libc_ia64_register_backing_store_base; + /* Try to read the backing store base from /proc/self/maps. */ + /* We look for the writable mapping with a 0 major device, */ + /* which is as close to our frame as possible, but below it.*/ + static word backing_store_base_from_maps(char *maps) + { + char prot_buf[5]; + char *buf_ptr = maps; + word start, end; + unsigned int maj_dev; + word current_best = 0; + word dummy; + + for (;;) { + buf_ptr = GC_parse_map_entry(buf_ptr, &start, &end, prot_buf, &maj_dev); + if (buf_ptr == NULL) return current_best; + if (prot_buf[1] == 'w' && maj_dev == 0) { + if (end < (word)(&dummy) && start > current_best) current_best = start; + } + } + return current_best; + } + + static word backing_store_base_from_proc(void) + { + return GC_apply_to_maps(backing_store_base_from_maps); + } + +# ifdef USE_LIBC_PRIVATES +# pragma weak __libc_ia64_register_backing_store_base + extern ptr_t __libc_ia64_register_backing_store_base; +# endif ptr_t GC_get_register_stack_base(void) { - if (0 != &__libc_ia64_register_backing_store_base) { - return __libc_ia64_register_backing_store_base; - } else { - word result = (word)GC_stackbottom - BACKING_STORE_DISPLACEMENT; - result += BACKING_STORE_ALIGNMENT - 1; - result &= ~(BACKING_STORE_ALIGNMENT - 1); - return (ptr_t)result; +# ifdef USE_LIBC_PRIVATES + if (0 != &__libc_ia64_register_backing_store_base + && 0 != __libc_ia64_register_backing_store_base) { + /* Glibc 2.2.4 has a bug such that for dynamically linked */ + /* executables __libc_ia64_register_backing_store_base is */ + /* defined but uninitialized during constructor calls. */ + /* Hence we check for both nonzero address and value. */ + return __libc_ia64_register_backing_store_base; + } +# endif + word result = backing_store_base_from_proc(); + if (0 == result) { + /* Use dumb heuristics. Works only for default configuration. */ + result = (word)GC_stackbottom - BACKING_STORE_DISPLACEMENT; + result += BACKING_STORE_ALIGNMENT - 1; + result &= ~(BACKING_STORE_ALIGNMENT - 1); + /* Verify that it's at least readable. If not, we goofed. */ + GC_noop1(*(word *)result); } + return (ptr_t)result; } # endif @@ -690,11 +923,8 @@ ptr_t GC_get_stack_base() /* using direct I/O system calls in order to avoid calling malloc */ /* in case REDIRECT_MALLOC is defined. */ # define STAT_BUF_SIZE 4096 -# if defined(GC_USE_LD_WRAP) -# define STAT_READ __real_read -# else -# define STAT_READ read -# endif +# define STAT_READ read + /* Should probably call the real read, if read is wrapped. */ char stat_buf[STAT_BUF_SIZE]; int f; char c; @@ -702,9 +932,33 @@ ptr_t GC_get_stack_base() size_t i, buf_offset = 0; /* First try the easy way. This should work for glibc 2.2 */ - if (0 != &__libc_stack_end) { - return __libc_stack_end; + /* This fails in a prelinked ("prelink" command) executable */ + /* since the correct value of __libc_stack_end never */ + /* becomes visible to us. The second test works around */ + /* this. */ +# ifdef USE_LIBC_PRIVATES + if (0 != &__libc_stack_end && 0 != __libc_stack_end ) { +# ifdef IA64 + /* Some versions of glibc set the address 16 bytes too */ + /* low while the initialization code is running. */ + if (((word)__libc_stack_end & 0xfff) + 0x10 < 0x1000) { + return __libc_stack_end + 0x10; + } /* Otherwise it's not safe to add 16 bytes and we fall */ + /* back to using /proc. */ +# else +# ifdef SPARC + /* Older versions of glibc for 64-bit Sparc do not set + * this variable correctly, it gets set to either zero + * or one. + */ + if (__libc_stack_end != (ptr_t) (unsigned long)0x1) + return __libc_stack_end; +# else + return __libc_stack_end; +# endif +# endif } +# endif f = open("/proc/self/stat", O_RDONLY); if (f < 0 || STAT_READ(f, stat_buf, STAT_BUF_SIZE) < 2 * STAT_SKIP) { ABORT("Couldn't read /proc/self/stat"); @@ -740,25 +994,28 @@ ptr_t GC_get_stack_base() ptr_t GC_freebsd_stack_base(void) { - int nm[2] = { CTL_KERN, KERN_USRSTACK}, base, len, r; - - len = sizeof(int); - r = sysctl(nm, 2, &base, &len, NULL, 0); + int nm[2] = {CTL_KERN, KERN_USRSTACK}; + ptr_t base; + size_t len = sizeof(ptr_t); + int r = sysctl(nm, 2, &base, &len, NULL, 0); if (r) ABORT("Error getting stack base"); - return (ptr_t)base; + return base; } #endif /* FREEBSD_STACKBOTTOM */ #if !defined(BEOS) && !defined(AMIGA) && !defined(MSWIN32) \ - && !defined(MSWINCE) && !defined(OS2) + && !defined(MSWINCE) && !defined(OS2) && !defined(NOSYS) && !defined(ECOS) ptr_t GC_get_stack_base() { +# if defined(HEURISTIC1) || defined(HEURISTIC2) || \ + defined(LINUX_STACKBOTTOM) || defined(FREEBSD_STACKBOTTOM) word dummy; ptr_t result; +# endif # define STACKBOTTOM_ALIGNMENT_M1 ((word)STACK_GRAN - 1) @@ -807,9 +1064,8 @@ ptr_t GC_get_stack_base() return(result); # endif /* STACKBOTTOM */ } -# endif /* ECOS */ -# endif /* ! AMIGA, !OS 2, ! MS Windows, !BEOS */ +# endif /* ! AMIGA, !OS 2, ! MS Windows, !BEOS, !NOSYS, !ECOS */ /* * Register static data segment(s) as roots. @@ -917,23 +1173,23 @@ void GC_register_data_segments() /* Unfortunately, we have to handle win32s very differently from NT, */ /* Since VirtualQuery has very different semantics. In particular, */ /* under win32s a VirtualQuery call on an unmapped page returns an */ - /* invalid result. Under GC_register_data_segments is a noop and */ + /* invalid result. Under NT, GC_register_data_segments is a noop and */ /* all real work is done by GC_register_dynamic_libraries. Under */ /* win32s, we cannot find the data segments associated with dll's. */ - /* We rgister the main data segment here. */ - GC_bool GC_win32s = FALSE; /* We're running under win32s. */ - - GC_bool GC_is_win32s() - { - DWORD v = GetVersion(); - - /* Check that this is not NT, and Windows major version <= 3 */ - return ((v & 0x80000000) && (v & 0xff) <= 3); - } + /* We register the main data segment here. */ + GC_bool GC_no_win32_dlls = FALSE; + /* This used to be set for gcc, to avoid dealing with */ + /* the structured exception handling issues. But we now have */ + /* assembly code to do that right. */ + GC_bool GC_wnt = FALSE; + /* This is a Windows NT derivative, i.e. NT, W2K, XP or later. */ void GC_init_win32() { - GC_win32s = GC_is_win32s(); + /* if we're running under win32s, assume that no DLLs will be loaded */ + DWORD v = GetVersion(); + GC_wnt = !(v & 0x80000000); + GC_no_win32_dlls |= ((!GC_wnt) && (v & 0xff) <= 3); } /* Return the smallest address a such that VirtualQuery */ @@ -959,36 +1215,102 @@ void GC_register_data_segments() return(p); } # endif + +# ifndef REDIRECT_MALLOC + /* We maintain a linked list of AllocationBase values that we know */ + /* correspond to malloc heap sections. Currently this is only called */ + /* during a GC. But there is some hope that for long running */ + /* programs we will eventually see most heap sections. */ + + /* In the long run, it would be more reliable to occasionally walk */ + /* the malloc heap with HeapWalk on the default heap. But that */ + /* apparently works only for NT-based Windows. */ + + /* In the long run, a better data structure would also be nice ... */ + struct GC_malloc_heap_list { + void * allocation_base; + struct GC_malloc_heap_list *next; + } *GC_malloc_heap_l = 0; + + /* Is p the base of one of the malloc heap sections we already know */ + /* about? */ + GC_bool GC_is_malloc_heap_base(ptr_t p) + { + struct GC_malloc_heap_list *q = GC_malloc_heap_l; + + while (0 != q) { + if (q -> allocation_base == p) return TRUE; + q = q -> next; + } + return FALSE; + } + + void *GC_get_allocation_base(void *p) + { + MEMORY_BASIC_INFORMATION buf; + DWORD result = VirtualQuery(p, &buf, sizeof(buf)); + if (result != sizeof(buf)) { + ABORT("Weird VirtualQuery result"); + } + return buf.AllocationBase; + } + + size_t GC_max_root_size = 100000; /* Appr. largest root size. */ + + void GC_add_current_malloc_heap() + { + struct GC_malloc_heap_list *new_l = + malloc(sizeof(struct GC_malloc_heap_list)); + void * candidate = GC_get_allocation_base(new_l); + + if (new_l == 0) return; + if (GC_is_malloc_heap_base(candidate)) { + /* Try a little harder to find malloc heap. */ + size_t req_size = 10000; + do { + void *p = malloc(req_size); + if (0 == p) { free(new_l); return; } + candidate = GC_get_allocation_base(p); + free(p); + req_size *= 2; + } while (GC_is_malloc_heap_base(candidate) + && req_size < GC_max_root_size/10 && req_size < 500000); + if (GC_is_malloc_heap_base(candidate)) { + free(new_l); return; + } + } +# ifdef CONDPRINT + if (GC_print_stats) + GC_printf1("Found new system malloc AllocationBase at 0x%lx\n", + candidate); +# endif + new_l -> allocation_base = candidate; + new_l -> next = GC_malloc_heap_l; + GC_malloc_heap_l = new_l; + } +# endif /* REDIRECT_MALLOC */ /* Is p the start of either the malloc heap, or of one of our */ /* heap sections? */ GC_bool GC_is_heap_base (ptr_t p) { - register unsigned i; + unsigned i; # ifndef REDIRECT_MALLOC - static ptr_t malloc_heap_pointer = 0; + static word last_gc_no = -1; - if (0 == malloc_heap_pointer) { - MEMORY_BASIC_INFORMATION buf; - void *pTemp = malloc( 1 ); - register DWORD result = VirtualQuery(pTemp, &buf, sizeof(buf)); - - free( pTemp ); - - - if (result != sizeof(buf)) { - ABORT("Weird VirtualQuery result"); - } - malloc_heap_pointer = (ptr_t)(buf.AllocationBase); + if (last_gc_no != GC_gc_no) { + GC_add_current_malloc_heap(); + last_gc_no = GC_gc_no; } - if (p == malloc_heap_pointer) return(TRUE); + if (GC_root_size > GC_max_root_size) GC_max_root_size = GC_root_size; + if (GC_is_malloc_heap_base(p)) return TRUE; # endif for (i = 0; i < GC_n_heap_bases; i++) { - if (GC_heap_bases[i] == p) return(TRUE); + if (GC_heap_bases[i] == p) return TRUE; } - return(FALSE); + return FALSE ; } # ifdef MSWIN32 @@ -1001,7 +1323,7 @@ void GC_register_data_segments() char * base; char * limit, * new_limit; - if (!GC_win32s) return; + if (!GC_no_win32_dlls) return; p = base = limit = GC_least_described_address(static_root); while (p < GC_sysinfo.lpMaximumApplicationAddress) { result = VirtualQuery(p, &buf, sizeof(buf)); @@ -1038,7 +1360,7 @@ void GC_register_data_segments() # if (defined(SVR4) || defined(AUX) || defined(DGUX) \ || (defined(LINUX) && defined(SPARC))) && !defined(PCR) -char * GC_SysVGetDataStart(max_page_size, etext_addr) +ptr_t GC_SysVGetDataStart(max_page_size, etext_addr) int max_page_size; int * etext_addr; { @@ -1053,7 +1375,7 @@ int * etext_addr; /* max_page_size to &etext if &etext is at a page boundary */ GC_setup_temporary_fault_handler(); - if (setjmp(GC_jmp_buf) == 0) { + if (SETJMP(GC_jmp_buf) == 0) { /* Try writing to the address. */ *result = *result; GC_reset_fault_handler(); @@ -1064,10 +1386,43 @@ int * etext_addr; /* string constants in the text segment, but after etext. */ /* Use plan B. Note that we now know there is a gap between */ /* text and data segments, so plan A bought us something. */ - result = (char *)GC_find_limit((ptr_t)(DATAEND) - MIN_PAGE_SIZE, FALSE); + result = (char *)GC_find_limit((ptr_t)(DATAEND), FALSE); + } + return((ptr_t)result); +} +# endif + +# if defined(FREEBSD) && (defined(I386) || defined(X86_64) || defined(powerpc) || defined(__powerpc__)) && !defined(PCR) +/* Its unclear whether this should be identical to the above, or */ +/* whether it should apply to non-X86 architectures. */ +/* For now we don't assume that there is always an empty page after */ +/* etext. But in some cases there actually seems to be slightly more. */ +/* This also deals with holes between read-only data and writable data. */ +ptr_t GC_FreeBSDGetDataStart(max_page_size, etext_addr) +int max_page_size; +int * etext_addr; +{ + word text_end = ((word)(etext_addr) + sizeof(word) - 1) + & ~(sizeof(word) - 1); + /* etext rounded to word boundary */ + VOLATILE word next_page = (text_end + (word)max_page_size - 1) + & ~((word)max_page_size - 1); + VOLATILE ptr_t result = (ptr_t)text_end; + GC_setup_temporary_fault_handler(); + if (SETJMP(GC_jmp_buf) == 0) { + /* Try reading at the address. */ + /* This should happen before there is another thread. */ + for (; next_page < (word)(DATAEND); next_page += (word)max_page_size) + *(VOLATILE char *)next_page; + GC_reset_fault_handler(); + } else { + GC_reset_fault_handler(); + /* As above, we go to plan B */ + result = GC_find_limit((ptr_t)(DATAEND), FALSE); } - return((char *)result); + return(result); } + # endif @@ -1081,9 +1436,8 @@ int * etext_addr; void GC_register_data_segments() { -# if !defined(PCR) && !defined(SRC_M3) && !defined(NEXT) && !defined(MACOS) \ - && !defined(MACOSX) -# if defined(REDIRECT_MALLOC) && defined(SOLARIS_THREADS) +# if !defined(PCR) && !defined(SRC_M3) && !defined(MACOS) +# if defined(REDIRECT_MALLOC) && defined(GC_SOLARIS_THREADS) /* As of Solaris 2.3, the Solaris threads implementation */ /* allocates the data structure for the initial thread with */ /* sbrk at process startup. It needs to be scanned, so that */ @@ -1094,11 +1448,11 @@ void GC_register_data_segments() GC_add_roots_inner(DATASTART, (char *)sbrk(0), FALSE); # else GC_add_roots_inner(DATASTART, (char *)(DATAEND), FALSE); +# if defined(DATASTART2) + GC_add_roots_inner(DATASTART2, (char *)(DATAEND2), FALSE); +# endif # endif # endif -# if !defined(PCR) && (defined(NEXT) || defined(MACOSX)) - GC_add_roots_inner(DATASTART, (char *) get_end(), FALSE); -# endif # if defined(MACOS) { # if defined(THINK_C) @@ -1159,7 +1513,7 @@ void GC_register_data_segments() # endif -# ifdef RS6000 +# if 0 && defined(RS6000) /* We now use mmap */ /* The compiler seems to generate speculative reads one past the end of */ /* an allocated object. Hence we need to make sure that the page */ /* following the last heap page is also mapped. */ @@ -1190,8 +1544,7 @@ word bytes; #else /* Not RS6000 */ -#if defined(USE_MMAP) -/* Tested only under Linux, IRIX5 and Solaris 2 */ +#if defined(USE_MMAP) || defined(USE_MUNMAP) #ifdef USE_MMAP_FIXED # define GC_MMAP_FLAGS MAP_FIXED | MAP_PRIVATE @@ -1201,6 +1554,23 @@ word bytes; # define GC_MMAP_FLAGS MAP_PRIVATE #endif +#ifdef USE_MMAP_ANON +# define zero_fd -1 +# if defined(MAP_ANONYMOUS) +# define OPT_MAP_ANON MAP_ANONYMOUS +# else +# define OPT_MAP_ANON MAP_ANON +# endif +#else + static int zero_fd; +# define OPT_MAP_ANON 0 +#endif + +#endif /* defined(USE_MMAP) || defined(USE_MUNMAP) */ + +#if defined(USE_MMAP) +/* Tested only under Linux, IRIX5 and Solaris 2 */ + #ifndef HEAP_START # define HEAP_START 0 #endif @@ -1208,18 +1578,22 @@ word bytes; ptr_t GC_unix_get_mem(bytes) word bytes; { - static GC_bool initialized = FALSE; - static int fd; void *result; static ptr_t last_addr = HEAP_START; - if (!initialized) { - fd = open("/dev/zero", O_RDONLY); - initialized = TRUE; - } +# ifndef USE_MMAP_ANON + static GC_bool initialized = FALSE; + + if (!initialized) { + zero_fd = open("/dev/zero", O_RDONLY); + fcntl(zero_fd, F_SETFD, FD_CLOEXEC); + initialized = TRUE; + } +# endif + if (bytes & (GC_page_size -1)) ABORT("Bad GET_MEM arg"); result = mmap(last_addr, bytes, PROT_READ | PROT_WRITE | OPT_PROT_EXEC, - GC_MMAP_FLAGS, fd, 0/* offset */); + GC_MMAP_FLAGS | OPT_MAP_ANON, zero_fd, 0/* offset */); if (result == MAP_FAILED) return(0); last_addr = (ptr_t)result + bytes + GC_page_size - 1; last_addr = (ptr_t)((word)last_addr & ~(GC_page_size - 1)); @@ -1292,8 +1666,14 @@ void * os2_alloc(size_t bytes) SYSTEM_INFO GC_sysinfo; # endif - # ifdef MSWIN32 + +# ifdef USE_GLOBAL_ALLOC +# define GLOBAL_ALLOC_TEST 1 +# else +# define GLOBAL_ALLOC_TEST GC_no_win32_dlls +# endif + word GC_n_heap_bases = 0; ptr_t GC_win32_get_mem(bytes) @@ -1301,14 +1681,22 @@ word bytes; { ptr_t result; - if (GC_win32s) { + if (GLOBAL_ALLOC_TEST) { /* VirtualAlloc doesn't like PAGE_EXECUTE_READWRITE. */ /* There are also unconfirmed rumors of other */ /* problems, so we dodge the issue. */ result = (ptr_t) GlobalAlloc(0, bytes + HBLKSIZE); result = (ptr_t)(((word)result + HBLKSIZE) & ~(HBLKSIZE-1)); } else { - result = (ptr_t) VirtualAlloc(NULL, bytes, + /* VirtualProtect only works on regions returned by a */ + /* single VirtualAlloc call. Thus we allocate one */ + /* extra page, which will prevent merging of blocks */ + /* in separate regions, and eliminate any temptation */ + /* to call VirtualProtect on a range spanning regions. */ + /* This wastes a small amount of memory, and risks */ + /* increased fragmentation. But better alternatives */ + /* would require effort. */ + result = (ptr_t) VirtualAlloc(NULL, bytes + 1, MEM_COMMIT | MEM_RESERVE, PAGE_EXECUTE_READWRITE); } @@ -1322,7 +1710,7 @@ word bytes; void GC_win32_free_heap () { - if (GC_win32s) { + if (GC_no_win32_dlls) { while (GC_n_heap_bases > 0) { GlobalFree (GC_heap_bases[--GC_n_heap_bases]); GC_heap_bases[GC_n_heap_bases] = 0; @@ -1364,6 +1752,10 @@ word bytes; /* Reserve more pages */ word res_bytes = (bytes + GC_sysinfo.dwAllocationGranularity-1) & ~(GC_sysinfo.dwAllocationGranularity-1); + /* If we ever support MPROTECT_VDB here, we will probably need to */ + /* ensure that res_bytes is strictly > bytes, so that VirtualProtect */ + /* never spans regions. It seems to be OK for a VirtualFree argument */ + /* to span regions, so we should be OK for now. */ result = (ptr_t) VirtualAlloc(NULL, res_bytes, MEM_RESERVE | MEM_TOP_DOWN, PAGE_EXECUTE_READWRITE); @@ -1459,7 +1851,15 @@ void GC_unmap(ptr_t start, word bytes) len -= free_len; } # else - if (munmap(start_addr, len) != 0) ABORT("munmap failed"); + /* We immediately remap it to prevent an intervening mmap from */ + /* accidentally grabbing the same address space. */ + { + void * result; + result = mmap(start_addr, len, PROT_NONE, + MAP_PRIVATE | MAP_FIXED | OPT_MAP_ANON, + zero_fd, 0/* offset */); + if (result != (void *)start_addr) ABORT("mmap(...PROT_NONE...) failed"); + } GC_unmapped_bytes += len; # endif } @@ -1467,13 +1867,13 @@ void GC_unmap(ptr_t start, word bytes) void GC_remap(ptr_t start, word bytes) { - static int zero_descr = -1; ptr_t start_addr = GC_unmap_start(start, bytes); ptr_t end_addr = GC_unmap_end(start, bytes); word len = end_addr - start_addr; - ptr_t result; # if defined(MSWIN32) || defined(MSWINCE) + ptr_t result; + if (0 == start_addr) return; while (len != 0) { MEMORY_BASIC_INFORMATION mem_info; @@ -1493,12 +1893,17 @@ void GC_remap(ptr_t start, word bytes) len -= alloc_len; } # else - if (-1 == zero_descr) zero_descr = open("/dev/zero", O_RDWR); + /* It was already remapped with PROT_NONE. */ + int result; + if (0 == start_addr) return; - result = mmap(start_addr, len, PROT_READ | PROT_WRITE | OPT_PROT_EXEC, - MAP_FIXED | MAP_PRIVATE, zero_descr, 0); - if (result != start_addr) { - ABORT("mmap remapping failed"); + result = mprotect(start_addr, len, + PROT_READ | PROT_WRITE | OPT_PROT_EXEC); + if (result != 0) { + GC_err_printf3( + "Mprotect failed at 0x%lx (length %ld) with errno %ld\n", + start_addr, len, errno); + ABORT("Mprotect remapping failed"); } GC_unmapped_bytes -= len; # endif @@ -1644,9 +2049,8 @@ void GC_default_push_other_roots GC_PROTO((void)) # endif /* SRC_M3 */ -# if defined(SOLARIS_THREADS) || defined(WIN32_THREADS) \ - || defined(IRIX_THREADS) || defined(LINUX_THREADS) \ - || defined(HPUX_THREADS) +# if defined(GC_SOLARIS_THREADS) || defined(GC_PTHREADS) || \ + defined(GC_WIN32_THREADS) extern void GC_push_all_stacks(); @@ -1655,11 +2059,11 @@ void GC_default_push_other_roots GC_PROTO((void)) GC_push_all_stacks(); } -# endif /* SOLARIS_THREADS || ... */ +# endif /* GC_SOLARIS_THREADS || GC_PTHREADS */ void (*GC_push_other_roots) GC_PROTO((void)) = GC_default_push_other_roots; -#endif +#endif /* THREADS */ /* * Routines for accessing dirty bits on virtual pages. @@ -1681,7 +2085,6 @@ void (*GC_push_other_roots) GC_PROTO((void)) = GC_default_push_other_roots; * make sure that other system calls are similarly protected * or write only to the stack. */ - GC_bool GC_dirty_maintained = FALSE; # ifdef DEFAULT_VDB @@ -1695,6 +2098,9 @@ GC_bool GC_dirty_maintained = FALSE; /* Initialize virtual dirty bit implementation. */ void GC_dirty_init() { +# ifdef PRINTSTATS + GC_printf0("Initializing DEFAULT_VDB...\n"); +# endif GC_dirty_maintained = TRUE; } @@ -1736,11 +2142,18 @@ word n; { } -/* A call hints that h is about to be written. */ -/* May speed up some dirty bit implementations. */ +/* A call that: */ +/* I) hints that [h, h+nblocks) is about to be written. */ +/* II) guarantees that protection is removed. */ +/* (I) may speed up some dirty bit implementations. */ +/* (II) may be essential if we need to ensure that */ +/* pointer-free system call buffers in the heap are */ +/* not protected. */ /*ARGSUSED*/ -void GC_write_hint(h) +void GC_remove_protection(h, nblocks, is_ptrfree) struct hblk *h; +word nblocks; +GC_bool is_ptrfree; { } @@ -1756,17 +2169,21 @@ struct hblk *h; /* * This implementation maintains dirty bits itself by catching write * faults and keeping track of them. We assume nobody else catches - * SIGBUS or SIGSEGV. We assume no write faults occur in system calls - * except as a result of a read system call. This means clients must - * either ensure that system calls do not touch the heap, or must - * provide their own wrappers analogous to the one for read. + * SIGBUS or SIGSEGV. We assume no write faults occur in system calls. + * This means that clients must ensure that system calls don't write + * to the write-protected heap. Probably the best way to do this is to + * ensure that system calls write at most to POINTERFREE objects in the + * heap, and do even that only if we are on a platform on which those + * are not protected. Another alternative is to wrap system calls + * (see example for read below), but the current implementation holds + * a lock across blocking calls, making it problematic for multithreaded + * applications. * We assume the page size is a multiple of HBLKSIZE. - * This implementation is currently SunOS 4.X and IRIX 5.X specific, though we - * tried to use portable code where easily possible. It is known - * not to work under a number of other systems. + * We prefer them to be the same. We avoid protecting POINTERFREE + * objects only if they are the same. */ -# if !defined(MSWIN32) && !defined(MSWINCE) +# if !defined(MSWIN32) && !defined(MSWINCE) && !defined(DARWIN) # include # include @@ -1785,6 +2202,23 @@ struct hblk *h; # else +# ifdef DARWIN + /* Using vm_protect (mach syscall) over mprotect (BSD syscall) seems to + decrease the likelihood of some of the problems described below. */ + #include + static mach_port_t GC_task_self; + #define PROTECT(addr,len) \ + if(vm_protect(GC_task_self,(vm_address_t)(addr),(vm_size_t)(len), \ + FALSE,VM_PROT_READ) != KERN_SUCCESS) { \ + ABORT("vm_portect failed"); \ + } + #define UNPROTECT(addr,len) \ + if(vm_protect(GC_task_self,(vm_address_t)(addr),(vm_size_t)(len), \ + FALSE,VM_PROT_READ|VM_PROT_WRITE) != KERN_SUCCESS) { \ + ABORT("vm_portect failed"); \ + } +# else + # ifndef MSWINCE # include # endif @@ -1802,19 +2236,22 @@ struct hblk *h; &protect_junk)) { \ ABORT("un-VirtualProtect failed"); \ } - -# endif +# endif /* !DARWIN */ +# endif /* MSWIN32 || MSWINCE || DARWIN */ -#if defined(SUNOS4) || defined(FREEBSD) +#if defined(SUNOS4) || (defined(FREEBSD) && !defined(SUNOS5SIGS)) typedef void (* SIG_PF)(); -#endif -#if defined(SUNOS5SIGS) || defined(OSF1) || defined(LINUX) || defined(MACOSX) +#endif /* SUNOS4 || (FREEBSD && !SUNOS5SIGS) */ + +#if defined(SUNOS5SIGS) || defined(OSF1) || defined(LINUX) \ + || defined(HURD) # ifdef __STDC__ typedef void (* SIG_PF)(int); # else typedef void (* SIG_PF)(); # endif -#endif +#endif /* SUNOS5SIGS || OSF1 || LINUX || HURD */ + #if defined(MSWIN32) typedef LPTOP_LEVEL_EXCEPTION_FILTER SIG_PF; # undef SIG_DFL @@ -1826,33 +2263,42 @@ struct hblk *h; # define SIG_DFL (SIG_PF) (-1) #endif -#if defined(IRIX5) || defined(OSF1) +#if defined(IRIX5) || defined(OSF1) || defined(HURD) typedef void (* REAL_SIG_PF)(int, int, struct sigcontext *); -#endif +#endif /* IRIX5 || OSF1 || HURD */ + #if defined(SUNOS5SIGS) -# ifdef HPUX -# define SIGINFO __siginfo +# if defined(HPUX) || defined(FREEBSD) +# define SIGINFO_T siginfo_t # else -# define SIGINFO siginfo +# define SIGINFO_T struct siginfo # endif # ifdef __STDC__ - typedef void (* REAL_SIG_PF)(int, struct SIGINFO *, void *); + typedef void (* REAL_SIG_PF)(int, SIGINFO_T *, void *); # else typedef void (* REAL_SIG_PF)(); # endif -#endif +#endif /* SUNOS5SIGS */ + #if defined(LINUX) -# include -# if (LINUX_VERSION_CODE >= 0x20100) && !defined(M68K) || defined(ALPHA) || defined(IA64) +# if __GLIBC__ > 2 || __GLIBC__ == 2 && __GLIBC_MINOR__ >= 2 typedef struct sigcontext s_c; -# else - typedef struct sigcontext_struct s_c; -# endif +# else /* glibc < 2.2 */ +# include +# if (LINUX_VERSION_CODE >= 0x20100) && !defined(M68K) || defined(ALPHA) || defined(ARM32) + typedef struct sigcontext s_c; +# else + typedef struct sigcontext_struct s_c; +# endif +# endif /* glibc < 2.2 */ # if defined(ALPHA) || defined(M68K) typedef void (* REAL_SIG_PF)(int, int, s_c *); # else -# if defined(IA64) || defined(HP_PA) +# if defined(IA64) || defined(HP_PA) || defined(X86_64) typedef void (* REAL_SIG_PF)(int, siginfo_t *, s_c *); + /* FIXME: */ + /* According to SUSV3, the last argument should have type */ + /* void * or ucontext_t * */ # else typedef void (* REAL_SIG_PF)(int, s_c); # endif @@ -1870,139 +2316,14 @@ struct hblk *h; return (char *)faultaddr; } # endif /* !ALPHA */ -# endif - -# if defined(MACOSX) /* Should also test for PowerPC? */ - typedef void (* REAL_SIG_PF)(int, int, struct sigcontext *); - -/* Decodes the machine instruction which was responsible for the sending of the - SIGBUS signal. Sadly this is the only way to find the faulting address because - the signal handler doesn't get it directly from the kernel (although it is - available on the Mach level, but droppped by the BSD personality before it - calls our signal handler...) - This code should be able to deal correctly with all PPCs starting from the - 601 up to and including the G4s (including Velocity Engine). */ -#define EXTRACT_OP1(iw) (((iw) & 0xFC000000) >> 26) -#define EXTRACT_OP2(iw) (((iw) & 0x000007FE) >> 1) -#define EXTRACT_REGA(iw) (((iw) & 0x001F0000) >> 16) -#define EXTRACT_REGB(iw) (((iw) & 0x03E00000) >> 21) -#define EXTRACT_REGC(iw) (((iw) & 0x0000F800) >> 11) -#define EXTRACT_DISP(iw) ((short *) &(iw))[1] - -static char *get_fault_addr(struct sigcontext *scp) -{ - unsigned int instr = *((unsigned int *) scp->sc_ir); - unsigned int * regs = &((unsigned int *) scp->sc_regs)[2]; - int disp = 0, tmp; - unsigned int baseA = 0, baseB = 0; - unsigned int addr, alignmask = 0xFFFFFFFF; - -#ifdef GC_DEBUG_DECODER - GC_err_printf1("Instruction: 0x%lx\n", instr); - GC_err_printf1("Opcode 1: d\n", (int)EXTRACT_OP1(instr)); -#endif - switch(EXTRACT_OP1(instr)) { - case 38: /* stb */ - case 39: /* stbu */ - case 54: /* stfd */ - case 55: /* stfdu */ - case 52: /* stfs */ - case 53: /* stfsu */ - case 44: /* sth */ - case 45: /* sthu */ - case 47: /* stmw */ - case 36: /* stw */ - case 37: /* stwu */ - tmp = EXTRACT_REGA(instr); - if(tmp > 0) - baseA = regs[tmp]; - disp = EXTRACT_DISP(instr); - break; - case 31: -#ifdef GC_DEBUG_DECODER - GC_err_printf1("Opcode 2: %d\n", (int)EXTRACT_OP2(instr)); -#endif - switch(EXTRACT_OP2(instr)) { - case 86: /* dcbf */ - case 54: /* dcbst */ - case 1014: /* dcbz */ - case 247: /* stbux */ - case 215: /* stbx */ - case 759: /* stfdux */ - case 727: /* stfdx */ - case 983: /* stfiwx */ - case 695: /* stfsux */ - case 663: /* stfsx */ - case 918: /* sthbrx */ - case 439: /* sthux */ - case 407: /* sthx */ - case 661: /* stswx */ - case 662: /* stwbrx */ - case 150: /* stwcx. */ - case 183: /* stwux */ - case 151: /* stwx */ - case 135: /* stvebx */ - case 167: /* stvehx */ - case 199: /* stvewx */ - case 231: /* stvx */ - case 487: /* stvxl */ - tmp = EXTRACT_REGA(instr); - if(tmp > 0) - baseA = regs[tmp]; - baseB = regs[EXTRACT_REGC(instr)]; - /* determine Altivec alignment mask */ - switch(EXTRACT_OP2(instr)) { - case 167: /* stvehx */ - alignmask = 0xFFFFFFFE; - break; - case 199: /* stvewx */ - alignmask = 0xFFFFFFFC; - break; - case 231: /* stvx */ - alignmask = 0xFFFFFFF0; - break; - case 487: /* stvxl */ - alignmask = 0xFFFFFFF0; - break; - } - break; - case 725: /* stswi */ - tmp = EXTRACT_REGA(instr); - if(tmp > 0) - baseA = regs[tmp]; - break; - default: /* ignore instruction */ -#ifdef GC_DEBUG_DECODER - GC_err_printf("Ignored by inner handler\n"); -#endif - return NULL; - break; - } - break; - default: /* ignore instruction */ -#ifdef GC_DEBUG_DECODER - GC_err_printf("Ignored by main handler\n"); -#endif - return NULL; - break; - } - - addr = (baseA + baseB) + disp; - addr &= alignmask; -#ifdef GC_DEBUG_DECODER - GC_err_printf1("BaseA: %d\n", baseA); - GC_err_printf1("BaseB: %d\n", baseB); - GC_err_printf1("Disp: %d\n", disp); - GC_err_printf1("Address: %d\n", addr); -#endif - return (char *)addr; -} -#endif /* MACOSX */ +# endif /* LINUX */ +#ifndef DARWIN SIG_PF GC_old_bus_handler; SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ +#endif /* !DARWIN */ -#ifdef THREADS +#if defined(THREADS) /* We need to lock around the bitmap update in the write fault handler */ /* in order to avoid the risk of losing a bit. We do this with a */ /* test-and-set spin lock if we know how to do that. Otherwise we */ @@ -2013,7 +2334,7 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ #ifdef GC_TEST_AND_SET_DEFINED static VOLATILE unsigned int fault_handler_lock = 0; void async_set_pht_entry_from_index(VOLATILE page_hash_table db, int index) { - while (GC_test_and_set(&fault_handler_lock)); + while (GC_test_and_set(&fault_handler_lock)) {} /* Could also revert to set_pht_entry_from_index_safe if initial */ /* GC_test_and_set fails. */ set_pht_entry_from_index(db, index); @@ -2051,7 +2372,8 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ #endif /* !THREADS */ /*ARGSUSED*/ -# if defined (SUNOS4) || defined(FREEBSD) +#if !defined(DARWIN) +# if defined (SUNOS4) || (defined(FREEBSD) && !defined(SUNOS5SIGS)) void GC_write_fault_handler(sig, code, scp, addr) int sig, code; struct sigcontext *scp; @@ -2064,28 +2386,39 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ # endif # ifdef FREEBSD # define SIG_OK (sig == SIGBUS) -# define CODE_OK (code == BUS_PAGE_FAULT) +# define CODE_OK TRUE # endif -# endif -# if defined(IRIX5) || defined(OSF1) +# endif /* SUNOS4 || (FREEBSD && !SUNOS5SIGS) */ + +# if defined(IRIX5) || defined(OSF1) || defined(HURD) # include void GC_write_fault_handler(int sig, int code, struct sigcontext *scp) -# define SIG_OK (sig == SIGSEGV) # ifdef OSF1 +# define SIG_OK (sig == SIGSEGV) # define CODE_OK (code == 2 /* experimentally determined */) # endif # ifdef IRIX5 +# define SIG_OK (sig == SIGSEGV) # define CODE_OK (code == EACCES) # endif -# endif +# ifdef HURD +# define SIG_OK (sig == SIGBUS || sig == SIGSEGV) +# define CODE_OK TRUE +# endif +# endif /* IRIX5 || OSF1 || HURD */ + # if defined(LINUX) # if defined(ALPHA) || defined(M68K) void GC_write_fault_handler(int sig, int code, s_c * sc) # else -# if defined(IA64) || defined(HP_PA) +# if defined(IA64) || defined(HP_PA) || defined(X86_64) void GC_write_fault_handler(int sig, siginfo_t * si, s_c * scp) # else - void GC_write_fault_handler(int sig, s_c sc) +# if defined(ARM32) + void GC_write_fault_handler(int sig, int a2, int a3, int a4, s_c sc) +# else + void GC_write_fault_handler(int sig, s_c sc) +# endif # endif # endif # define SIG_OK (sig == SIGSEGV) @@ -2093,14 +2426,15 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ /* Empirically c.trapno == 14, on IA32, but is that useful? */ /* Should probably consider alignment issues on other */ /* architectures. */ -# endif +# endif /* LINUX */ + # if defined(SUNOS5SIGS) # ifdef __STDC__ - void GC_write_fault_handler(int sig, struct SIGINFO *scp, void * context) + void GC_write_fault_handler(int sig, SIGINFO_T *scp, void * context) # else void GC_write_fault_handler(sig, scp, context) int sig; - struct SIGINFO *scp; + SIGINFO_T *scp; void * context; # endif # ifdef HPUX @@ -2111,16 +2445,15 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ || (scp -> si_code == SEGV_UNKNOWN) \ || (scp -> si_code == BUS_OBJERR) # else -# define SIG_OK (sig == SIGSEGV) -# define CODE_OK (scp -> si_code == SEGV_ACCERR) -# endif -# endif - -# if defined(MACOSX) - void GC_write_fault_handler(int sig, int code, struct sigcontext *scp) -# define SIG_OK (sig == SIGBUS) -# define CODE_OK (code == 0 /* experimentally determined */) -# endif +# ifdef FREEBSD +# define SIG_OK (sig == SIGBUS) +# define CODE_OK (scp -> si_code == BUS_PAGE_FAULT) +# else +# define SIG_OK (sig == SIGSEGV) +# define CODE_OK (scp -> si_code == SEGV_ACCERR) +# endif +# endif +# endif /* SUNOS5SIGS */ # if defined(MSWIN32) || defined(MSWINCE) LONG WINAPI GC_write_fault_handler(struct _EXCEPTION_POINTERS *exc_info) @@ -2128,9 +2461,12 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ STATUS_ACCESS_VIOLATION) # define CODE_OK (exc_info -> ExceptionRecord -> ExceptionInformation[0] == 1) /* Write fault */ -# endif +# endif /* MSWIN32 || MSWINCE */ { register unsigned i; +# if defined(HURD) + char *addr = (char *) code; +# endif # ifdef IRIX5 char * addr = (char *) (size_t) (scp -> sc_badvaddr); # endif @@ -2141,7 +2477,7 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ char * addr = (char *) (scp -> si_addr); # endif # ifdef LINUX -# ifdef I386 +# if defined(I386) char * addr = (char *) (sc.cr2); # else # if defined(M68K) @@ -2176,7 +2512,7 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ # ifdef ALPHA char * addr = get_fault_addr(sc); # else -# if defined(IA64) || defined(HP_PA) +# if defined(IA64) || defined(HP_PA) || defined(X86_64) char * addr = si -> si_addr; /* I believe this is claimed to work on all platforms for */ /* Linux 2.3.47 and later. Hopefully we don't have to */ @@ -2185,16 +2521,21 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ # if defined(POWERPC) char * addr = (char *) (sc.regs->dar); # else - --> architecture not supported +# if defined(ARM32) + char * addr = (char *)sc.fault_address; +# else +# if defined(CRIS) + char * addr = (char *)sc.regs.csraddr; +# else + --> architecture not supported +# endif +# endif # endif # endif # endif # endif # endif # endif -# if defined(MACOSX) - char * addr = get_fault_addr(scp); -# endif # if defined(MSWIN32) || defined(MSWINCE) char * addr = (char *) (exc_info -> ExceptionRecord -> ExceptionInformation[1]); @@ -2218,6 +2559,10 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ in_allocd_block = (HDR(addr) != 0); # endif if (!in_allocd_block) { + /* FIXME - We should make sure that we invoke the */ + /* old handler with the appropriate calling */ + /* sequence, which often depends on SA_SIGINFO. */ + /* Heap blocks now begin and end on page boundaries */ SIG_PF old_handler; @@ -2234,11 +2579,17 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ return(EXCEPTION_CONTINUE_SEARCH); # endif } else { -# if defined (SUNOS4) || defined(FREEBSD) +# if defined (SUNOS4) \ + || (defined(FREEBSD) && !defined(SUNOS5SIGS)) (*old_handler) (sig, code, scp, addr); return; # endif # if defined (SUNOS5SIGS) + /* + * FIXME: For FreeBSD, this code should check if the + * old signal handler used the traditional BSD style and + * if so call it using that style. + */ (*(REAL_SIG_PF)old_handler) (sig, scp, context); return; # endif @@ -2246,7 +2597,7 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ # if defined(ALPHA) || defined(M68K) (*(REAL_SIG_PF)old_handler) (sig, code, sc); # else -# if defined(IA64) || defined(HP_PA) +# if defined(IA64) || defined(HP_PA) || defined(X86_64) (*(REAL_SIG_PF)old_handler) (sig, si, scp); # else (*(REAL_SIG_PF)old_handler) (sig, sc); @@ -2254,25 +2605,33 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ # endif return; # endif -# if defined (IRIX5) || defined(OSF1) +# if defined (IRIX5) || defined(OSF1) || defined(HURD) (*(REAL_SIG_PF)old_handler) (sig, code, scp); return; # endif -# ifdef MACOSX - (*(REAL_SIG_PF)old_handler) (sig, code, scp); -# endif # ifdef MSWIN32 return((*old_handler)(exc_info)); # endif } } + UNPROTECT(h, GC_page_size); + /* We need to make sure that no collection occurs between */ + /* the UNPROTECT and the setting of the dirty bit. Otherwise */ + /* a write by a third thread might go unnoticed. Reversing */ + /* the order is just as bad, since we would end up unprotecting */ + /* a page in a GC cycle during which it's not marked. */ + /* Currently we do this by disabling the thread stopping */ + /* signals while this handler is running. An alternative might */ + /* be to record the fact that we're about to unprotect, or */ + /* have just unprotected a page in the GC's thread structure, */ + /* and then to have the thread stopping code set the dirty */ + /* flag, if necessary. */ for (i = 0; i < divHBLKSZ(GC_page_size); i++) { register int index = PHT_HASH(h+i); async_set_pht_entry_from_index(GC_dirty_pages, index); } - UNPROTECT(h, GC_page_size); -# if defined(OSF1) || defined(LINUX) +# if defined(OSF1) /* These reset the signal handler each time by default. */ signal(SIGSEGV, (SIG_PF) GC_write_fault_handler); # endif @@ -2291,54 +2650,63 @@ SIG_PF GC_old_segv_handler; /* Also old MSWIN32 ACCESS_VIOLATION filter */ ABORT("Unexpected bus error or segmentation fault"); #endif } +#endif /* !DARWIN */ /* * We hold the allocation lock. We expect block h to be written - * shortly. + * shortly. Ensure that all pages containing any part of the n hblks + * starting at h are no longer protected. If is_ptrfree is false, + * also ensure that they will subsequently appear to be dirty. */ -void GC_write_hint(h) +void GC_remove_protection(h, nblocks, is_ptrfree) struct hblk *h; +word nblocks; +GC_bool is_ptrfree; { - register struct hblk * h_trunc; - register unsigned i; - register GC_bool found_clean; + struct hblk * h_trunc; /* Truncated to page boundary */ + struct hblk * h_end; /* Page boundary following block end */ + struct hblk * current; + GC_bool found_clean; if (!GC_dirty_maintained) return; h_trunc = (struct hblk *)((word)h & ~(GC_page_size-1)); + h_end = (struct hblk *)(((word)(h + nblocks) + GC_page_size-1) + & ~(GC_page_size-1)); found_clean = FALSE; - for (i = 0; i < divHBLKSZ(GC_page_size); i++) { - register int index = PHT_HASH(h_trunc+i); + for (current = h_trunc; current < h_end; ++current) { + int index = PHT_HASH(current); - if (!get_pht_entry_from_index(GC_dirty_pages, index)) { - found_clean = TRUE; + if (!is_ptrfree || current < h || current >= h + nblocks) { async_set_pht_entry_from_index(GC_dirty_pages, index); } } - if (found_clean) { - UNPROTECT(h_trunc, GC_page_size); - } + UNPROTECT(h_trunc, (ptr_t)h_end - (ptr_t)h_trunc); } +#if !defined(DARWIN) void GC_dirty_init() { -# if defined(SUNOS5SIGS) || defined(IRIX5) /* || defined(OSF1) */ +# if defined(SUNOS5SIGS) || defined(IRIX5) || defined(LINUX) || \ + defined(OSF1) || defined(HURD) struct sigaction act, oldact; -# ifdef IRIX5 + /* We should probably specify SA_SIGINFO for Linux, and handle */ + /* the different architectures more uniformly. */ +# if defined(IRIX5) || defined(LINUX) && !defined(X86_64) \ + || defined(OSF1) || defined(HURD) act.sa_flags = SA_RESTART; - act.sa_handler = GC_write_fault_handler; + act.sa_handler = (SIG_PF)GC_write_fault_handler; # else act.sa_flags = SA_RESTART | SA_SIGINFO; act.sa_sigaction = GC_write_fault_handler; # endif (void)sigemptyset(&act.sa_mask); +# ifdef SIG_SUSPEND + /* Arrange to postpone SIG_SUSPEND while we're in a write fault */ + /* handler. This effectively makes the handler atomic w.r.t. */ + /* stopping the world for GC. */ + (void)sigaddset(&act.sa_mask, SIG_SUSPEND); +# endif /* SIG_SUSPEND */ # endif -# if defined(MACOSX) - struct sigaction act, oldact; - - act.sa_flags = SA_RESTART; - act.sa_handler = GC_write_fault_handler; - sigemptyset(&act.sa_mask); -# endif # ifdef PRINTSTATS GC_printf0("Inititalizing mprotect virtual dirty bit implementation\n"); # endif @@ -2347,7 +2715,7 @@ void GC_dirty_init() GC_err_printf0("Page size not multiple of HBLKSIZE\n"); ABORT("Page size not multiple of HBLKSIZE"); } -# if defined(SUNOS4) || defined(FREEBSD) +# if defined(SUNOS4) || (defined(FREEBSD) && !defined(SUNOS5SIGS)) GC_old_bus_handler = signal(SIGBUS, GC_write_fault_handler); if (GC_old_bus_handler == SIG_IGN) { GC_err_printf0("Previously ignored bus error!?"); @@ -2359,7 +2727,7 @@ void GC_dirty_init() # endif } # endif -# if defined(OSF1) || defined(SUNOS4) || defined(LINUX) +# if defined(SUNOS4) GC_old_segv_handler = signal(SIGSEGV, (SIG_PF)GC_write_fault_handler); if (GC_old_segv_handler == SIG_IGN) { GC_err_printf0("Previously ignored segmentation violation!?"); @@ -2371,18 +2739,23 @@ void GC_dirty_init() # endif } # endif -# if defined(SUNOS5SIGS) || defined(IRIX5) -# if defined(IRIX_THREADS) +# if (defined(SUNOS5SIGS) && !defined(FREEBSD)) || defined(IRIX5) \ + || defined(LINUX) || defined(OSF1) || defined(HURD) + /* SUNOS5SIGS includes HPUX */ +# if defined(GC_IRIX_THREADS) sigaction(SIGSEGV, 0, &oldact); sigaction(SIGSEGV, &act, 0); -# else - sigaction(SIGSEGV, &act, &oldact); +# else + { + int res = sigaction(SIGSEGV, &act, &oldact); + if (res != 0) ABORT("Sigaction failed"); + } # endif -# if defined(_sigargs) +# if defined(_sigargs) || defined(HURD) || !defined(SA_SIGINFO) /* This is Irix 5.x, not 6.x. Irix 5.x does not have */ /* sa_sigaction. */ GC_old_segv_handler = oldact.sa_handler; -# else /* Irix 6.x or SUNOS5SIGS */ +# else /* Irix 6.x or SUNOS5SIGS or LINUX */ if (oldact.sa_flags & SA_SIGINFO) { GC_old_segv_handler = (SIG_PF)(oldact.sa_sigaction); } else { @@ -2398,8 +2771,9 @@ void GC_dirty_init() GC_err_printf0("Replaced other SIGSEGV handler\n"); # endif } -# endif -# if defined(MACOSX) || defined(HPUX) +# endif /* (SUNOS5SIGS && !FREEBSD) || IRIX5 || LINUX || OSF1 || HURD */ +# if defined(HPUX) || defined(LINUX) || defined(HURD) \ + || (defined(FREEBSD) && defined(SUNOS5SIGS)) sigaction(SIGBUS, &act, &oldact); GC_old_bus_handler = oldact.sa_handler; if (GC_old_bus_handler == SIG_IGN) { @@ -2411,7 +2785,7 @@ void GC_dirty_init() GC_err_printf0("Replaced other SIGBUS handler\n"); # endif } -# endif /* MACOS || HPUX */ +# endif /* HPUX || LINUX || HURD || (FREEBSD && SUNOS5SIGS) */ # if defined(MSWIN32) GC_old_segv_handler = SetUnhandledExceptionFilter(GC_write_fault_handler); if (GC_old_segv_handler != NULL) { @@ -2423,19 +2797,79 @@ void GC_dirty_init() } # endif } +#endif /* !DARWIN */ + +int GC_incremental_protection_needs() +{ + if (GC_page_size == HBLKSIZE) { + return GC_PROTECTS_POINTER_HEAP; + } else { + return GC_PROTECTS_POINTER_HEAP | GC_PROTECTS_PTRFREE_HEAP; + } +} +#define HAVE_INCREMENTAL_PROTECTION_NEEDS +#define IS_PTRFREE(hhdr) ((hhdr)->hb_descr == 0) +#define PAGE_ALIGNED(x) !((word)(x) & (GC_page_size - 1)) void GC_protect_heap() { ptr_t start; word len; + struct hblk * current; + struct hblk * current_start; /* Start of block to be protected. */ + struct hblk * limit; unsigned i; - + GC_bool protect_all = + (0 != (GC_incremental_protection_needs() & GC_PROTECTS_PTRFREE_HEAP)); for (i = 0; i < GC_n_heap_sects; i++) { start = GC_heap_sects[i].hs_start; len = GC_heap_sects[i].hs_bytes; - PROTECT(start, len); + if (protect_all) { + PROTECT(start, len); + } else { + GC_ASSERT(PAGE_ALIGNED(len)) + GC_ASSERT(PAGE_ALIGNED(start)) + current_start = current = (struct hblk *)start; + limit = (struct hblk *)(start + len); + while (current < limit) { + hdr * hhdr; + word nhblks; + GC_bool is_ptrfree; + + GC_ASSERT(PAGE_ALIGNED(current)); + GET_HDR(current, hhdr); + if (IS_FORWARDING_ADDR_OR_NIL(hhdr)) { + /* This can happen only if we're at the beginning of a */ + /* heap segment, and a block spans heap segments. */ + /* We will handle that block as part of the preceding */ + /* segment. */ + GC_ASSERT(current_start == current); + current_start = ++current; + continue; + } + if (HBLK_IS_FREE(hhdr)) { + GC_ASSERT(PAGE_ALIGNED(hhdr -> hb_sz)); + nhblks = divHBLKSZ(hhdr -> hb_sz); + is_ptrfree = TRUE; /* dirty on alloc */ + } else { + nhblks = OBJ_SZ_TO_BLOCKS(hhdr -> hb_sz); + is_ptrfree = IS_PTRFREE(hhdr); + } + if (is_ptrfree) { + if (current_start < current) { + PROTECT(current_start, (ptr_t)current - (ptr_t)current_start); + } + current_start = (current += nhblks); + } else { + current += nhblks; + } + } + if (current_start < current) { + PROTECT(current_start, (ptr_t)current - (ptr_t)current_start); + } + } } } @@ -2492,7 +2926,7 @@ word len; register struct hblk *h; ptr_t obj_start; - if (!GC_incremental) return; + if (!GC_dirty_maintained) return; obj_start = GC_base(addr); if (obj_start == 0) return; if (GC_base(addr + len - 1) != obj_start) { @@ -2510,11 +2944,23 @@ word len; ((ptr_t)end_block - (ptr_t)start_block) + HBLKSIZE); } -#if !defined(MSWIN32) && !defined(MSWINCE) && !defined(LINUX_THREADS) \ - && !defined(GC_USE_LD_WRAP) -/* Replacement for UNIX system call. */ -/* Other calls that write to the heap */ -/* should be handled similarly. */ +#if 0 + +/* We no longer wrap read by default, since that was causing too many */ +/* problems. It is preferred that the client instead avoids writing */ +/* to the write-protected heap with a system call. */ +/* This still serves as sample code if you do want to wrap system calls.*/ + +#if !defined(MSWIN32) && !defined(MSWINCE) && !defined(GC_USE_LD_WRAP) +/* Replacement for UNIX system call. */ +/* Other calls that write to the heap should be handled similarly. */ +/* Note that this doesn't work well for blocking reads: It will hold */ +/* the allocation lock for the entire duration of the call. Multithreaded */ +/* clients should really ensure that it won't block, either by setting */ +/* the descriptor nonblocking, or by calling select or poll first, to */ +/* make sure that input is available. */ +/* Another, preferred alternative is to ensure that system calls never */ +/* write to the protected heap (see above). */ # if defined(__STDC__) && !defined(SUNOS4) # include # include @@ -2534,7 +2980,7 @@ word len; GC_begin_syscall(); GC_unprotect_range(buf, (word)nbyte); -# if defined(IRIX5) || defined(LINUX_THREADS) +# if defined(IRIX5) || defined(GC_LINUX_THREADS) /* Indirect system call may not always be easily available. */ /* We could call _read, but that would interfere with the */ /* libpthread interception of read. */ @@ -2548,17 +2994,21 @@ word len; result = readv(fd, &iov, 1); } # else +# if defined(HURD) + result = __read(fd, buf, nbyte); +# else /* The two zero args at the end of this list are because one IA-64 syscall() implementation actually requires six args to be passed, even though they aren't always used. */ result = syscall(SYS_read, fd, buf, nbyte, 0, 0); +# endif /* !HURD */ # endif GC_end_syscall(); return(result); } -#endif /* !MSWIN32 && !MSWINCE && !LINUX_THREADS */ +#endif /* !MSWIN32 && !MSWINCE && !GC_LINUX_THREADS */ -#ifdef GC_USE_LD_WRAP +#if defined(GC_USE_LD_WRAP) && !defined(THREADS) /* We use the GNU ld call wrapping facility. */ /* This requires that the linker be invoked with "--wrap read". */ /* This can be done by passing -Wl,"--wrap read" to gcc. */ @@ -2580,6 +3030,8 @@ word len; /* actually calls. */ #endif +#endif /* 0 */ + /*ARGSUSED*/ GC_bool GC_page_was_ever_dirty(h) struct hblk *h; @@ -2595,13 +3047,6 @@ word n; { } -# else /* !MPROTECT_VDB */ - -# ifdef GC_USE_LD_WRAP - ssize_t __wrap_read(int fd, void *buf, size_t nbyte) - { return __real_read(fd, buf, nbyte); } -# endif - # endif /* MPROTECT_VDB */ # ifdef PROC_VDB @@ -2625,11 +3070,11 @@ word n; #include #include -#define INITIAL_BUF_SZ 4096 +#define INITIAL_BUF_SZ 16384 word GC_proc_buf_size = INITIAL_BUF_SZ; char *GC_proc_buf; -#ifdef SOLARIS_THREADS +#ifdef GC_SOLARIS_THREADS /* We don't have exact sp values for threads. So we count on */ /* occasionally declaring stack pages to be fresh. Thus we */ /* need a real implementation of GC_is_fresh. We can't clear */ @@ -2680,11 +3125,12 @@ void GC_dirty_init() } GC_proc_fd = syscall(SYS_ioctl, fd, PIOCOPENPD, 0); close(fd); + syscall(SYS_fcntl, GC_proc_fd, F_SETFD, FD_CLOEXEC); if (GC_proc_fd < 0) { ABORT("/proc ioctl failed"); } GC_proc_buf = GC_scratch_alloc(GC_proc_buf_size); -# ifdef SOLARIS_THREADS +# ifdef GC_SOLARIS_THREADS GC_fresh_pages = (struct hblk **) GC_scratch_alloc(MAX_FRESH_PAGES * sizeof (struct hblk *)); if (GC_fresh_pages == 0) { @@ -2697,12 +3143,14 @@ void GC_dirty_init() /* Ignore write hints. They don't help us here. */ /*ARGSUSED*/ -void GC_write_hint(h) +void GC_remove_protection(h, nblocks, is_ptrfree) struct hblk *h; +word nblocks; +GC_bool is_ptrfree; { } -#ifdef SOLARIS_THREADS +#ifdef GC_SOLARIS_THREADS # define READ(fd,buf,nbytes) syscall(SYS_read, fd, buf, nbytes) #else # define READ(fd,buf,nbytes) read(fd, buf, nbytes) @@ -2736,12 +3184,12 @@ int dummy; GC_proc_buf = bufp = new_buf; GC_proc_buf_size = new_size; } - if (syscall(SYS_read, GC_proc_fd, bufp, GC_proc_buf_size) <= 0) { + if (READ(GC_proc_fd, bufp, GC_proc_buf_size) <= 0) { WARN("Insufficient space for /proc read\n", 0); /* Punt: */ memset(GC_grungy_pages, 0xff, sizeof (page_hash_table)); memset(GC_written_pages, 0xff, sizeof(page_hash_table)); -# ifdef SOLARIS_THREADS +# ifdef GC_SOLARIS_THREADS BZERO(GC_fresh_pages, MAX_FRESH_PAGES * sizeof (struct hblk *)); # endif @@ -2771,7 +3219,7 @@ int dummy; register word index = PHT_HASH(h); set_pht_entry_from_index(GC_grungy_pages, index); -# ifdef SOLARIS_THREADS +# ifdef GC_SOLARIS_THREADS { register int slot = FRESH_PAGE_SLOT(h); @@ -2789,7 +3237,7 @@ int dummy; } /* Update GC_written_pages. */ GC_or_pages(GC_written_pages, GC_grungy_pages); -# ifdef SOLARIS_THREADS +# ifdef GC_SOLARIS_THREADS /* Make sure that old stacks are considered completely clean */ /* unless written again. */ GC_old_stacks_are_fresh(); @@ -2805,7 +3253,7 @@ struct hblk *h; register GC_bool result; result = get_pht_entry_from_index(GC_grungy_pages, index); -# ifdef SOLARIS_THREADS +# ifdef GC_SOLARIS_THREADS if (result && PAGE_IS_FRESH(h)) result = FALSE; /* This happens only if page was declared fresh since */ /* the read_dirty call, e.g. because it's in an unused */ @@ -2823,7 +3271,7 @@ struct hblk *h; register GC_bool result; result = get_pht_entry_from_index(GC_written_pages, index); -# ifdef SOLARIS_THREADS +# ifdef GC_SOLARIS_THREADS if (result && PAGE_IS_FRESH(h)) result = FALSE; # endif return(result); @@ -2837,7 +3285,7 @@ word n; register word index; -# ifdef SOLARIS_THREADS +# ifdef GC_SOLARIS_THREADS register word i; if (GC_fresh_pages != 0) { @@ -2906,15 +3354,592 @@ struct hblk *h; } /*ARGSUSED*/ -void GC_write_hint(h) +void GC_remove_protection(h, nblocks, is_ptrfree) struct hblk *h; +word nblocks; +GC_bool is_ptrfree; { - PCR_VD_WriteProtectDisable(h, HBLKSIZE); - PCR_VD_WriteProtectEnable(h, HBLKSIZE); + PCR_VD_WriteProtectDisable(h, nblocks*HBLKSIZE); + PCR_VD_WriteProtectEnable(h, nblocks*HBLKSIZE); } # endif /* PCR_VDB */ +#if defined(MPROTECT_VDB) && defined(DARWIN) +/* The following sources were used as a *reference* for this exception handling + code: + 1. Apple's mach/xnu documentation + 2. Timothy J. Wood's "Mach Exception Handlers 101" post to the + omnigroup's macosx-dev list. + www.omnigroup.com/mailman/archive/macosx-dev/2000-June/014178.html + 3. macosx-nat.c from Apple's GDB source code. +*/ + +/* The bug that caused all this trouble should now be fixed. This should + eventually be removed if all goes well. */ +/* define BROKEN_EXCEPTION_HANDLING */ + +#include +#include +#include +#include +#include +#include + +/* These are not defined in any header, although they are documented */ +extern boolean_t exc_server(mach_msg_header_t *,mach_msg_header_t *); +extern kern_return_t exception_raise( + mach_port_t,mach_port_t,mach_port_t, + exception_type_t,exception_data_t,mach_msg_type_number_t); +extern kern_return_t exception_raise_state( + mach_port_t,mach_port_t,mach_port_t, + exception_type_t,exception_data_t,mach_msg_type_number_t, + thread_state_flavor_t*,thread_state_t,mach_msg_type_number_t, + thread_state_t,mach_msg_type_number_t*); +extern kern_return_t exception_raise_state_identity( + mach_port_t,mach_port_t,mach_port_t, + exception_type_t,exception_data_t,mach_msg_type_number_t, + thread_state_flavor_t*,thread_state_t,mach_msg_type_number_t, + thread_state_t,mach_msg_type_number_t*); + + +#define MAX_EXCEPTION_PORTS 16 + +static struct { + mach_msg_type_number_t count; + exception_mask_t masks[MAX_EXCEPTION_PORTS]; + exception_handler_t ports[MAX_EXCEPTION_PORTS]; + exception_behavior_t behaviors[MAX_EXCEPTION_PORTS]; + thread_state_flavor_t flavors[MAX_EXCEPTION_PORTS]; +} GC_old_exc_ports; + +static struct { + mach_port_t exception; +#if defined(THREADS) + mach_port_t reply; +#endif +} GC_ports; + +typedef struct { + mach_msg_header_t head; +} GC_msg_t; + +typedef enum { + GC_MP_NORMAL, GC_MP_DISCARDING, GC_MP_STOPPED +} GC_mprotect_state_t; + +/* FIXME: 1 and 2 seem to be safe to use in the msgh_id field, + but it isn't documented. Use the source and see if they + should be ok. */ +#define ID_STOP 1 +#define ID_RESUME 2 + +/* These values are only used on the reply port */ +#define ID_ACK 3 + +#if defined(THREADS) + +GC_mprotect_state_t GC_mprotect_state; + +/* The following should ONLY be called when the world is stopped */ +static void GC_mprotect_thread_notify(mach_msg_id_t id) { + struct { + GC_msg_t msg; + mach_msg_trailer_t trailer; + } buf; + mach_msg_return_t r; + /* remote, local */ + buf.msg.head.msgh_bits = + MACH_MSGH_BITS(MACH_MSG_TYPE_MAKE_SEND,0); + buf.msg.head.msgh_size = sizeof(buf.msg); + buf.msg.head.msgh_remote_port = GC_ports.exception; + buf.msg.head.msgh_local_port = MACH_PORT_NULL; + buf.msg.head.msgh_id = id; + + r = mach_msg( + &buf.msg.head, + MACH_SEND_MSG|MACH_RCV_MSG|MACH_RCV_LARGE, + sizeof(buf.msg), + sizeof(buf), + GC_ports.reply, + MACH_MSG_TIMEOUT_NONE, + MACH_PORT_NULL); + if(r != MACH_MSG_SUCCESS) + ABORT("mach_msg failed in GC_mprotect_thread_notify"); + if(buf.msg.head.msgh_id != ID_ACK) + ABORT("invalid ack in GC_mprotect_thread_notify"); +} + +/* Should only be called by the mprotect thread */ +static void GC_mprotect_thread_reply() { + GC_msg_t msg; + mach_msg_return_t r; + /* remote, local */ + msg.head.msgh_bits = + MACH_MSGH_BITS(MACH_MSG_TYPE_MAKE_SEND,0); + msg.head.msgh_size = sizeof(msg); + msg.head.msgh_remote_port = GC_ports.reply; + msg.head.msgh_local_port = MACH_PORT_NULL; + msg.head.msgh_id = ID_ACK; + + r = mach_msg( + &msg.head, + MACH_SEND_MSG, + sizeof(msg), + 0, + MACH_PORT_NULL, + MACH_MSG_TIMEOUT_NONE, + MACH_PORT_NULL); + if(r != MACH_MSG_SUCCESS) + ABORT("mach_msg failed in GC_mprotect_thread_reply"); +} + +void GC_mprotect_stop() { + GC_mprotect_thread_notify(ID_STOP); +} +void GC_mprotect_resume() { + GC_mprotect_thread_notify(ID_RESUME); +} + +#else /* !THREADS */ +/* The compiler should optimize away any GC_mprotect_state computations */ +#define GC_mprotect_state GC_MP_NORMAL +#endif + +static void *GC_mprotect_thread(void *arg) { + mach_msg_return_t r; + /* These two structures contain some private kernel data. We don't need to + access any of it so we don't bother defining a proper struct. The + correct definitions are in the xnu source code. */ + struct { + mach_msg_header_t head; + char data[256]; + } reply; + struct { + mach_msg_header_t head; + mach_msg_body_t msgh_body; + char data[1024]; + } msg; + + mach_msg_id_t id; + + GC_darwin_register_mach_handler_thread(mach_thread_self()); + + for(;;) { + r = mach_msg( + &msg.head, + MACH_RCV_MSG|MACH_RCV_LARGE| + (GC_mprotect_state == GC_MP_DISCARDING ? MACH_RCV_TIMEOUT : 0), + 0, + sizeof(msg), + GC_ports.exception, + GC_mprotect_state == GC_MP_DISCARDING ? 0 : MACH_MSG_TIMEOUT_NONE, + MACH_PORT_NULL); + + id = r == MACH_MSG_SUCCESS ? msg.head.msgh_id : -1; + +#if defined(THREADS) + if(GC_mprotect_state == GC_MP_DISCARDING) { + if(r == MACH_RCV_TIMED_OUT) { + GC_mprotect_state = GC_MP_STOPPED; + GC_mprotect_thread_reply(); + continue; + } + if(r == MACH_MSG_SUCCESS && (id == ID_STOP || id == ID_RESUME)) + ABORT("out of order mprotect thread request"); + } +#endif + + if(r != MACH_MSG_SUCCESS) { + GC_err_printf2("mach_msg failed with %d %s\n", + (int)r,mach_error_string(r)); + ABORT("mach_msg failed"); + } + + switch(id) { +#if defined(THREADS) + case ID_STOP: + if(GC_mprotect_state != GC_MP_NORMAL) + ABORT("Called mprotect_stop when state wasn't normal"); + GC_mprotect_state = GC_MP_DISCARDING; + break; + case ID_RESUME: + if(GC_mprotect_state != GC_MP_STOPPED) + ABORT("Called mprotect_resume when state wasn't stopped"); + GC_mprotect_state = GC_MP_NORMAL; + GC_mprotect_thread_reply(); + break; +#endif /* THREADS */ + default: + /* Handle the message (calls catch_exception_raise) */ + if(!exc_server(&msg.head,&reply.head)) + ABORT("exc_server failed"); + /* Send the reply */ + r = mach_msg( + &reply.head, + MACH_SEND_MSG, + reply.head.msgh_size, + 0, + MACH_PORT_NULL, + MACH_MSG_TIMEOUT_NONE, + MACH_PORT_NULL); + if(r != MACH_MSG_SUCCESS) { + /* This will fail if the thread dies, but the thread shouldn't + die... */ + #ifdef BROKEN_EXCEPTION_HANDLING + GC_err_printf2( + "mach_msg failed with %d %s while sending exc reply\n", + (int)r,mach_error_string(r)); + #else + ABORT("mach_msg failed while sending exception reply"); + #endif + } + } /* switch */ + } /* for(;;) */ + /* NOT REACHED */ + return NULL; +} + +/* All this SIGBUS code shouldn't be necessary. All protection faults should + be going throught the mach exception handler. However, it seems a SIGBUS is + occasionally sent for some unknown reason. Even more odd, it seems to be + meaningless and safe to ignore. */ +#ifdef BROKEN_EXCEPTION_HANDLING + +typedef void (* SIG_PF)(); +static SIG_PF GC_old_bus_handler; + +/* Updates to this aren't atomic, but the SIGBUSs seem pretty rare. + Even if this doesn't get updated property, it isn't really a problem */ +static int GC_sigbus_count; + +static void GC_darwin_sigbus(int num,siginfo_t *sip,void *context) { + if(num != SIGBUS) ABORT("Got a non-sigbus signal in the sigbus handler"); + + /* Ugh... some seem safe to ignore, but too many in a row probably means + trouble. GC_sigbus_count is reset for each mach exception that is + handled */ + if(GC_sigbus_count >= 8) { + ABORT("Got more than 8 SIGBUSs in a row!"); + } else { + GC_sigbus_count++; + GC_err_printf0("GC: WARNING: Ignoring SIGBUS.\n"); + } +} +#endif /* BROKEN_EXCEPTION_HANDLING */ + +void GC_dirty_init() { + kern_return_t r; + mach_port_t me; + pthread_t thread; + pthread_attr_t attr; + exception_mask_t mask; + +# ifdef PRINTSTATS + GC_printf0("Inititalizing mach/darwin mprotect virtual dirty bit " + "implementation\n"); +# endif +# ifdef BROKEN_EXCEPTION_HANDLING + GC_err_printf0("GC: WARNING: Enabling workarounds for various darwin " + "exception handling bugs.\n"); +# endif + GC_dirty_maintained = TRUE; + if (GC_page_size % HBLKSIZE != 0) { + GC_err_printf0("Page size not multiple of HBLKSIZE\n"); + ABORT("Page size not multiple of HBLKSIZE"); + } + + GC_task_self = me = mach_task_self(); + + r = mach_port_allocate(me,MACH_PORT_RIGHT_RECEIVE,&GC_ports.exception); + if(r != KERN_SUCCESS) ABORT("mach_port_allocate failed (exception port)"); + + r = mach_port_insert_right(me,GC_ports.exception,GC_ports.exception, + MACH_MSG_TYPE_MAKE_SEND); + if(r != KERN_SUCCESS) + ABORT("mach_port_insert_right failed (exception port)"); + + #if defined(THREADS) + r = mach_port_allocate(me,MACH_PORT_RIGHT_RECEIVE,&GC_ports.reply); + if(r != KERN_SUCCESS) ABORT("mach_port_allocate failed (reply port)"); + #endif + + /* The exceptions we want to catch */ + mask = EXC_MASK_BAD_ACCESS; + + r = task_get_exception_ports( + me, + mask, + GC_old_exc_ports.masks, + &GC_old_exc_ports.count, + GC_old_exc_ports.ports, + GC_old_exc_ports.behaviors, + GC_old_exc_ports.flavors + ); + if(r != KERN_SUCCESS) ABORT("task_get_exception_ports failed"); + + r = task_set_exception_ports( + me, + mask, + GC_ports.exception, + EXCEPTION_DEFAULT, + GC_MACH_THREAD_STATE + ); + if(r != KERN_SUCCESS) ABORT("task_set_exception_ports failed"); + + if(pthread_attr_init(&attr) != 0) ABORT("pthread_attr_init failed"); + if(pthread_attr_setdetachstate(&attr,PTHREAD_CREATE_DETACHED) != 0) + ABORT("pthread_attr_setdetachedstate failed"); + +# undef pthread_create + /* This will call the real pthread function, not our wrapper */ + if(pthread_create(&thread,&attr,GC_mprotect_thread,NULL) != 0) + ABORT("pthread_create failed"); + pthread_attr_destroy(&attr); + + /* Setup the sigbus handler for ignoring the meaningless SIGBUSs */ + #ifdef BROKEN_EXCEPTION_HANDLING + { + struct sigaction sa, oldsa; + sa.sa_handler = (SIG_PF)GC_darwin_sigbus; + sigemptyset(&sa.sa_mask); + sa.sa_flags = SA_RESTART|SA_SIGINFO; + if(sigaction(SIGBUS,&sa,&oldsa) < 0) ABORT("sigaction"); + GC_old_bus_handler = (SIG_PF)oldsa.sa_handler; + if (GC_old_bus_handler != SIG_DFL) { +# ifdef PRINTSTATS + GC_err_printf0("Replaced other SIGBUS handler\n"); +# endif + } + } + #endif /* BROKEN_EXCEPTION_HANDLING */ +} + +/* The source code for Apple's GDB was used as a reference for the exception + forwarding code. This code is similar to be GDB code only because there is + only one way to do it. */ +static kern_return_t GC_forward_exception( + mach_port_t thread, + mach_port_t task, + exception_type_t exception, + exception_data_t data, + mach_msg_type_number_t data_count +) { + int i; + kern_return_t r; + mach_port_t port; + exception_behavior_t behavior; + thread_state_flavor_t flavor; + + thread_state_t thread_state; + mach_msg_type_number_t thread_state_count = THREAD_STATE_MAX; + + for(i=0;i 0 ? code[0] : -1, + code_count > 1 ? code[1] : -1); + #endif + return FWD(); + } + + r = thread_get_state(thread,flavor, + (natural_t*)&exc_state,&exc_state_count); + if(r != KERN_SUCCESS) { + /* The thread is supposed to be suspended while the exception handler + is called. This shouldn't fail. */ + #ifdef BROKEN_EXCEPTION_HANDLING + GC_err_printf0("thread_get_state failed in " + "catch_exception_raise\n"); + return KERN_SUCCESS; + #else + ABORT("thread_get_state failed in catch_exception_raise"); + #endif + } + + /* This is the address that caused the fault */ +#if defined(POWERPC) + addr = (char*) exc_state. THREAD_FLD(dar); +#elif defined (I386) || defined (X86_64) + addr = (char*) exc_state. THREAD_FLD(faultvaddr); +#else +# error FIXME for non POWERPC/I386 +#endif + + if((HDR(addr)) == 0) { + /* Ugh... just like the SIGBUS problem above, it seems we get a bogus + KERN_PROTECTION_FAILURE every once and a while. We wait till we get + a bunch in a row before doing anything about it. If a "real" fault + ever occurres it'll just keep faulting over and over and we'll hit + the limit pretty quickly. */ + #ifdef BROKEN_EXCEPTION_HANDLING + static char *last_fault; + static int last_fault_count; + + if(addr != last_fault) { + last_fault = addr; + last_fault_count = 0; + } + if(++last_fault_count < 32) { + if(last_fault_count == 1) + GC_err_printf1( + "GC: WARNING: Ignoring KERN_PROTECTION_FAILURE at %p\n", + addr); + return KERN_SUCCESS; + } + + GC_err_printf1("Unexpected KERN_PROTECTION_FAILURE at %p\n",addr); + /* Can't pass it along to the signal handler because that is + ignoring SIGBUS signals. We also shouldn't call ABORT here as + signals don't always work too well from the exception handler. */ + GC_err_printf0("Aborting\n"); + exit(EXIT_FAILURE); + #else /* BROKEN_EXCEPTION_HANDLING */ + /* Pass it along to the next exception handler + (which should call SIGBUS/SIGSEGV) */ + return FWD(); + #endif /* !BROKEN_EXCEPTION_HANDLING */ + } + + #ifdef BROKEN_EXCEPTION_HANDLING + /* Reset the number of consecutive SIGBUSs */ + GC_sigbus_count = 0; + #endif + + if(GC_mprotect_state == GC_MP_NORMAL) { /* common case */ + h = (struct hblk*)((word)addr & ~(GC_page_size-1)); + UNPROTECT(h, GC_page_size); + for (i = 0; i < divHBLKSZ(GC_page_size); i++) { + register int index = PHT_HASH(h+i); + async_set_pht_entry_from_index(GC_dirty_pages, index); + } + } else if(GC_mprotect_state == GC_MP_DISCARDING) { + /* Lie to the thread for now. No sense UNPROTECT()ing the memory + when we're just going to PROTECT() it again later. The thread + will just fault again once it resumes */ + } else { + /* Shouldn't happen, i don't think */ + GC_printf0("KERN_PROTECTION_FAILURE while world is stopped\n"); + return FWD(); + } + return KERN_SUCCESS; +} +#undef FWD + +/* These should never be called, but just in case... */ +kern_return_t catch_exception_raise_state(mach_port_name_t exception_port, + int exception, exception_data_t code, mach_msg_type_number_t codeCnt, + int flavor, thread_state_t old_state, int old_stateCnt, + thread_state_t new_state, int new_stateCnt) +{ + ABORT("catch_exception_raise_state"); + return(KERN_INVALID_ARGUMENT); +} +kern_return_t catch_exception_raise_state_identity( + mach_port_name_t exception_port, mach_port_t thread, mach_port_t task, + int exception, exception_data_t code, mach_msg_type_number_t codeCnt, + int flavor, thread_state_t old_state, int old_stateCnt, + thread_state_t new_state, int new_stateCnt) +{ + ABORT("catch_exception_raise_state_identity"); + return(KERN_INVALID_ARGUMENT); +} + + +#endif /* DARWIN && MPROTECT_VDB */ + +# ifndef HAVE_INCREMENTAL_PROTECTION_NEEDS + int GC_incremental_protection_needs() + { + return GC_PROTECTS_NONE; + } +# endif /* !HAVE_INCREMENTAL_PROTECTION_NEEDS */ + /* * Call stack save code for debugging. * Should probably be in mach_dep.c, but that requires reorganization. @@ -2924,6 +3949,8 @@ struct hblk *h; /* long as the frame pointer is explicitly stored. In the case of gcc, */ /* compiler flags (e.g. -fomit-frame-pointer) determine whether it is. */ #if defined(I386) && defined(LINUX) && defined(SAVE_CALL_CHAIN) +# include + struct frame { struct frame *fr_savfp; long fr_savpc; @@ -2933,6 +3960,8 @@ struct hblk *h; #if defined(SPARC) # if defined(LINUX) +# include + struct frame { long fr_local[8]; long fr_arg[6]; @@ -2951,10 +3980,14 @@ struct hblk *h; # if defined (DRSNX) # include # else -# if defined(OPENBSD) || defined(NETBSD) +# if defined(OPENBSD) # include # else -# include +# if defined(FREEBSD) || defined(NETBSD) +# include +# else +# include +# endif # endif # endif # endif @@ -2964,11 +3997,64 @@ struct hblk *h; # endif #endif /* SPARC */ -#ifdef SAVE_CALL_CHAIN +#ifdef NEED_CALLINFO /* Fill in the pc and argument information for up to NFRAMES of my */ /* callers. Ignore my frame and my callers frame. */ -#if (defined(OPENBSD) || defined(NETBSD)) && defined(SPARC) +#ifdef LINUX +# include +#endif + +#endif /* NEED_CALLINFO */ + +#if defined(GC_HAVE_BUILTIN_BACKTRACE) +# include +#endif + +#ifdef SAVE_CALL_CHAIN + +#if NARGS == 0 && NFRAMES % 2 == 0 /* No padding */ \ + && defined(GC_HAVE_BUILTIN_BACKTRACE) + +#ifdef REDIRECT_MALLOC + /* Deal with possible malloc calls in backtrace by omitting */ + /* the infinitely recursing backtrace. */ +# ifdef THREADS + __thread /* If your compiler doesn't understand this */ + /* you could use something like pthread_getspecific. */ +# endif + GC_in_save_callers = FALSE; +#endif + +void GC_save_callers (info) +struct callinfo info[NFRAMES]; +{ + void * tmp_info[NFRAMES + 1]; + int npcs, i; +# define IGNORE_FRAMES 1 + + /* We retrieve NFRAMES+1 pc values, but discard the first, since it */ + /* points to our own frame. */ +# ifdef REDIRECT_MALLOC + if (GC_in_save_callers) { + info[0].ci_pc = (word)(&GC_save_callers); + for (i = 1; i < NFRAMES; ++i) info[i].ci_pc = 0; + return; + } + GC_in_save_callers = TRUE; +# endif + GC_ASSERT(sizeof(struct callinfo) == sizeof(void *)); + npcs = backtrace((void **)tmp_info, NFRAMES + IGNORE_FRAMES); + BCOPY(tmp_info+IGNORE_FRAMES, info, (npcs - IGNORE_FRAMES) * sizeof(void *)); + for (i = npcs - IGNORE_FRAMES; i < NFRAMES; ++i) info[i].ci_pc = 0; +# ifdef REDIRECT_MALLOC + GC_in_save_callers = FALSE; +# endif +} + +#else /* No builtin backtrace; do it ourselves */ + +#if (defined(OPENBSD) || defined(NETBSD) || defined(FREEBSD)) && defined(SPARC) # define FR_SAVFP fr_fp # define FR_SAVPC fr_pc #else @@ -2993,8 +4079,6 @@ struct callinfo info[NFRAMES]; asm("movl %%ebp,%0" : "=r"(frame)); fp = frame; # else - word GC_save_regs_in_stack(); - frame = (struct frame *) GC_save_regs_in_stack (); fp = (struct frame *)((long) frame -> FR_SAVFP + BIAS); #endif @@ -3005,40 +4089,172 @@ struct callinfo info[NFRAMES]; register int i; info[nframes].ci_pc = fp->FR_SAVPC; - for (i = 0; i < NARGS; i++) { - info[nframes].ci_arg[i] = ~(fp->fr_arg[i]); - } +# if NARGS > 0 + for (i = 0; i < NARGS; i++) { + info[nframes].ci_arg[i] = ~(fp->fr_arg[i]); + } +# endif /* NARGS > 0 */ } if (nframes < NFRAMES) info[nframes].ci_pc = 0; } -#endif /* SAVE_CALL_CHAIN */ +#endif /* No builtin backtrace */ -#if defined(LINUX) && defined(__ELF__) && \ - (!defined(SMALL_CONFIG) || defined(USE_PROC_FOR_LIBRARIES)) -#ifdef GC_USE_LD_WRAP -# define READ __real_read -#else -# define READ read -#endif +#endif /* SAVE_CALL_CHAIN */ +#ifdef NEED_CALLINFO -/* Repeatedly perform a read call until the buffer is filled or */ -/* we encounter EOF. */ -ssize_t GC_repeat_read(int fd, char *buf, size_t count) +/* Print info to stderr. We do NOT hold the allocation lock */ +void GC_print_callers (info) +struct callinfo info[NFRAMES]; { - ssize_t num_read = 0; - ssize_t result; + register int i; + static int reentry_count = 0; + GC_bool stop = FALSE; + + /* FIXME: This should probably use a different lock, so that we */ + /* become callable with or without the allocation lock. */ + LOCK(); + ++reentry_count; + UNLOCK(); - while (num_read < count) { - result = READ(fd, buf + num_read, count - num_read); - if (result < 0) return result; - if (result == 0) break; - num_read += result; +# if NFRAMES == 1 + GC_err_printf0("\tCaller at allocation:\n"); +# else + GC_err_printf0("\tCall chain at allocation:\n"); +# endif + for (i = 0; i < NFRAMES && !stop ; i++) { + if (info[i].ci_pc == 0) break; +# if NARGS > 0 + { + int j; + + GC_err_printf0("\t\targs: "); + for (j = 0; j < NARGS; j++) { + if (j != 0) GC_err_printf0(", "); + GC_err_printf2("%d (0x%X)", ~(info[i].ci_arg[j]), + ~(info[i].ci_arg[j])); + } + GC_err_printf0("\n"); + } +# endif + if (reentry_count > 1) { + /* We were called during an allocation during */ + /* a previous GC_print_callers call; punt. */ + GC_err_printf1("\t\t##PC##= 0x%lx\n", info[i].ci_pc); + continue; + } + { +# ifdef LINUX + FILE *pipe; +# endif +# if defined(GC_HAVE_BUILTIN_BACKTRACE) \ + && !defined(GC_BACKTRACE_SYMBOLS_BROKEN) + char **sym_name = + backtrace_symbols((void **)(&(info[i].ci_pc)), 1); + char *name = sym_name[0]; +# else + char buf[40]; + char *name = buf; + sprintf(buf, "##PC##= 0x%lx", info[i].ci_pc); +# endif +# if defined(LINUX) && !defined(SMALL_CONFIG) + /* Try for a line number. */ + { +# define EXE_SZ 100 + static char exe_name[EXE_SZ]; +# define CMD_SZ 200 + char cmd_buf[CMD_SZ]; +# define RESULT_SZ 200 + static char result_buf[RESULT_SZ]; + size_t result_len; + char *old_preload; +# define PRELOAD_SZ 200 + char preload_buf[PRELOAD_SZ]; + static GC_bool found_exe_name = FALSE; + static GC_bool will_fail = FALSE; + int ret_code; + /* Try to get it via a hairy and expensive scheme. */ + /* First we get the name of the executable: */ + if (will_fail) goto out; + if (!found_exe_name) { + ret_code = readlink("/proc/self/exe", exe_name, EXE_SZ); + if (ret_code < 0 || ret_code >= EXE_SZ + || exe_name[0] != '/') { + will_fail = TRUE; /* Dont try again. */ + goto out; + } + exe_name[ret_code] = '\0'; + found_exe_name = TRUE; + } + /* Then we use popen to start addr2line -e */ + /* There are faster ways to do this, but hopefully this */ + /* isn't time critical. */ + sprintf(cmd_buf, "/usr/bin/addr2line -f -e %s 0x%lx", exe_name, + (unsigned long)info[i].ci_pc); + old_preload = getenv ("LD_PRELOAD"); + if (0 != old_preload) { + if (strlen (old_preload) >= PRELOAD_SZ) { + will_fail = TRUE; + goto out; + } + strcpy (preload_buf, old_preload); + unsetenv ("LD_PRELOAD"); + } + pipe = popen(cmd_buf, "r"); + if (0 != old_preload + && 0 != setenv ("LD_PRELOAD", preload_buf, 0)) { + WARN("Failed to reset LD_PRELOAD\n", 0); + } + if (pipe == NULL + || (result_len = fread(result_buf, 1, RESULT_SZ - 1, pipe)) + == 0) { + if (pipe != NULL) pclose(pipe); + will_fail = TRUE; + goto out; + } + if (result_buf[result_len - 1] == '\n') --result_len; + result_buf[result_len] = 0; + if (result_buf[0] == '?' + || result_buf[result_len-2] == ':' + && result_buf[result_len-1] == '0') { + pclose(pipe); + goto out; + } + /* Get rid of embedded newline, if any. Test for "main" */ + { + char * nl = strchr(result_buf, '\n'); + if (nl != NULL && nl < result_buf + result_len) { + *nl = ':'; + } + if (strncmp(result_buf, "main", nl - result_buf) == 0) { + stop = TRUE; + } + } + if (result_len < RESULT_SZ - 25) { + /* Add in hex address */ + sprintf(result_buf + result_len, " [0x%lx]", + (unsigned long)info[i].ci_pc); + } + name = result_buf; + pclose(pipe); + out:; + } +# endif /* LINUX */ + GC_err_printf1("\t\t%s\n", name); +# if defined(GC_HAVE_BUILTIN_BACKTRACE) \ + && !defined(GC_BACKTRACE_SYMBOLS_BROKEN) + free(sym_name); /* May call GC_free; that's OK */ +# endif + } } - return num_read; + LOCK(); + --reentry_count; + UNLOCK(); } -#endif /* LINUX && ... */ + +#endif /* NEED_CALLINFO */ + #if defined(LINUX) && defined(__ELF__) && !defined(SMALL_CONFIG) @@ -3046,20 +4262,16 @@ ssize_t GC_repeat_read(int fd, char *buf, size_t count) /* Dump /proc/self/maps to GC_stderr, to enable looking up names for addresses in FIND_LEAK output. */ +static word dump_maps(char *maps) +{ + GC_err_write(maps, strlen(maps)); + return 1; +} + void GC_print_address_map() { - int f; - int result; - char maps_temp[32768]; GC_err_printf0("---------- Begin address map ----------\n"); - f = open("/proc/self/maps", O_RDONLY); - if (-1 == f) ABORT("Couldn't open /proc/self/maps"); - do { - result = GC_repeat_read(f, maps_temp, sizeof(maps_temp)); - if (result <= 0) ABORT("Couldn't read /proc/self/maps"); - GC_err_write(maps_temp, result); - } while (result == sizeof(maps_temp)); - + GC_apply_to_maps(dump_maps); GC_err_printf0("---------- End address map ----------\n"); }