X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Flinux-os.c;h=a23bc5ddb306a113a74356164bb616092ea49669;hb=f0cb0cf9c0fe1b6fce5d10dbd34a0b7b249c4ae8;hp=376b08594e3588534ab51b4901fc45fa79778ae6;hpb=dfa55a883f94470267b626dae77ce7e7dfac3df6;p=sbcl.git diff --git a/src/runtime/linux-os.c b/src/runtime/linux-os.c index 376b085..a23bc5d 100644 --- a/src/runtime/linux-os.c +++ b/src/runtime/linux-os.c @@ -23,6 +23,7 @@ #include #include #include +#include "sbcl.h" #include "./signal.h" #include "os.h" #include "arch.h" @@ -30,9 +31,13 @@ #include "interrupt.h" #include "interr.h" #include "lispregs.h" -#include "sbcl.h" +#include "runtime.h" +#include "genesis/static-symbols.h" +#include "genesis/fdefn.h" + #include #include +#include #include #include @@ -40,255 +45,510 @@ #include #include #include +#include #include "validate.h" -size_t os_vm_page_size; +#include "thread.h" +#include "gc.h" +#if defined LISP_FEATURE_GENCGC +#include "gencgc-internal.h" +#else +#include "cheneygc-internal.h" +#endif +#include +#ifdef LISP_FEATURE_SB_WTIMER +# include +#endif -#if defined GENCGC -#include "gencgc.h" +#ifdef LISP_FEATURE_X86 +/* Prototype for personality(2). Done inline here since the header file + * for this isn't available on old versions of glibc. */ +int personality (unsigned long); #endif - -void os_init(void) + +size_t os_vm_page_size; + +#if defined(LISP_FEATURE_SB_THREAD) && defined(LISP_FEATURE_SB_FUTEX) && !defined(LISP_FEATURE_SB_PTHREAD_FUTEX) +#include +#include +#include + +/* values taken from the kernel's linux/futex.h. This header file + doesn't exist in userspace, which is our excuse for not grovelling + them automatically */ +#define FUTEX_WAIT 0 +#define FUTEX_WAKE 1 +/* This is also copied from linux/futex.h so that a binary compiled on + * a not so recent Linux system can still take advantage of private + * futexes when available.*/ +#define FUTEX_WAIT_PRIVATE (0+128) +#define FUTEX_WAKE_PRIVATE (1+128) +#define FUTEX_FD (2) +#define FUTEX_REQUEUE (3) + +/* Not static so that Lisp may query it. */ +boolean futex_private_supported_p; + +static inline int +futex_wait_op() { - /* Early versions of Linux don't support the mmap(..) functionality - * that we need. */ - { - struct utsname name; - int major_version; - uname(&name); - major_version = atoi(name.release); - if (major_version < 2) { - lose("linux major version=%d (can't run in version < 2.0.0)", - major_version); - } + return (futex_private_supported_p ? FUTEX_WAIT_PRIVATE : FUTEX_WAIT); +} + +static inline int +futex_wake_op() +{ + return (futex_private_supported_p ? FUTEX_WAKE_PRIVATE : FUTEX_WAKE); +} + +#define sys_futex sbcl_sys_futex +static inline int sys_futex (void *futex, int op, int val, struct timespec *rel) +{ + return syscall (SYS_futex, futex, op, val, rel); +} + +static void +futex_init() +{ + int x = 0; + sys_futex(&x, FUTEX_WAIT, 1, 0); + if (errno == ENOSYS) + lose("This version of SBCL is compiled with threading support, but your kernel\n" + "is too old to support this. Please use a more recent kernel or\n" + "a version of SBCL without threading support.\n"); + sys_futex(&x, FUTEX_WAIT_PRIVATE, 1, 0); + if (errno == EWOULDBLOCK) { + futex_private_supported_p = 1; + } else { + futex_private_supported_p = 0; + SHOW("No futex private suppport\n"); } +} - os_vm_page_size = getpagesize(); - /* this could just as well be in arch_init, but it's not */ -#ifdef i386 - SET_FPU_CONTROL_WORD(0x1372|4|8|16|32); /* no interrupts */ -#endif +int +futex_wait(int *lock_word, int oldval, long sec, unsigned long usec) +{ + struct timespec timeout; + int t; + + if (sec<0) { + t = sys_futex(lock_word, futex_wait_op(), oldval, 0); + } + else { + timeout.tv_sec = sec; + timeout.tv_nsec = usec * 1000; + t = sys_futex(lock_word, futex_wait_op(), oldval, &timeout); + } + if (t==0) + return 0; + else if (errno==ETIMEDOUT) + return 1; + else if (errno==EINTR) + return 2; + else + /* EWOULDBLOCK and others, need to check the lock */ + return -1; } -/* various os_context_*_addr accessors moved to {x86,alpha}-linux-os.c - * -dan 20010125 +int +futex_wake(int *lock_word, int n) +{ + return sys_futex(lock_word, futex_wake_op(),n,0); +} +#endif + + +int linux_sparc_siginfo_bug = 0; + +/* This variable was in real use for a few months, basically for + * storing autodetected information about whether the Linux + * installation was recent enough to support SBCL threads, and make + * some run-time decisions based on that. But this turned out to be + * unstable, so now we just flat-out refuse to start on the old installations + * when thread support has been compiled in. + * + * Unfortunately, in the meanwhile Slime started depending on this + * variable for deciding which communication style to use. So even + * though this variable looks unused, it shouldn't be deleted until + * it's no longer used in the versions of Slime that people are + * likely to download first. -- JES, 2006-06-07 */ +int linux_no_threads_p = 0; -/* In Debian CMU CL ca. 2.4.9, it was possible to get an infinite - * cascade of errors from do_mmap(..). This variable is a counter to - * prevent that; when it counts down to zero, an error in do_mmap - * causes the low-level monitor to be called. */ -int n_do_mmap_ignorable_errors = 3; +#ifdef LISP_FEATURE_SB_THREAD +int +isnptl (void) +{ + size_t n = confstr (_CS_GNU_LIBPTHREAD_VERSION, NULL, 0); + if (n > 0) { + char *buf = alloca (n); + confstr (_CS_GNU_LIBPTHREAD_VERSION, buf, n); + if (strstr (buf, "NPTL")) { + return 1; + } + } + return 0; +} +#endif -/* Return 0 for success. */ -static int -do_mmap(os_vm_address_t *addr, os_vm_size_t len, int flags) +void +os_init(char *argv[], char *envp[]) { - /* We *must* have the memory where we want it. */ - os_vm_address_t old_addr=*addr; - - *addr = mmap(*addr, len, OS_VM_PROT_ALL, flags, -1, 0); - if (*addr == MAP_FAILED || - ((old_addr != NULL) && (*addr != old_addr))) { - FSHOW((stderr, - "error in allocating memory from the OS\n" - "(addr=%lx, len=%lx, flags=%lx)\n", - (long) addr, - (long) len, - (long) flags)); - if (n_do_mmap_ignorable_errors > 0) { - --n_do_mmap_ignorable_errors; - } else { - lose("too many errors in allocating memory from the OS"); - } - perror("mmap"); - return 1; + /* Conduct various version checks: do we have enough mmap(), is + * this a sparc running 2.2, can we do threads? */ + struct utsname name; + int major_version; + int minor_version; + int patch_version; + char *p; + uname(&name); + + p=name.release; + major_version = atoi(p); + minor_version = patch_version = 0; + p=strchr(p,'.'); + if (p != NULL) { + minor_version = atoi(++p); + p=strchr(p,'.'); + if (p != NULL) + patch_version = atoi(++p); } - return 0; + + if (major_version<2) { + lose("linux kernel version too old: major version=%d (can't run in version < 2.0.0)\n", + major_version); + } + if (!(major_version>2 || minor_version >= 4)) { +#ifdef LISP_FEATURE_SPARC + FSHOW((stderr,"linux kernel %d.%d predates 2.4;\n enabling workarounds for SPARC kernel bugs in signal handling.\n", major_version,minor_version)); + linux_sparc_siginfo_bug = 1; +#endif + } +#ifdef LISP_FEATURE_SB_THREAD +#if defined(LISP_FEATURE_SB_FUTEX) && !defined(LISP_FEATURE_SB_PTHREAD_FUTEX) + futex_init(); +#endif + if(! isnptl()) { + lose("This version of SBCL only works correctly with the NPTL threading\n" + "library. Please use a newer glibc, use an older SBCL, or stop using\n" + "LD_ASSUME_KERNEL\n"); + } +#endif + + /* Don't use getpagesize(), since it's not constant across Linux + * kernel versions on some architectures (for example PPC). FIXME: + * possibly the same should be done on other architectures too. + */ + os_vm_page_size = BACKEND_PAGE_BYTES; + + /* KLUDGE: Disable memory randomization on new Linux kernels + * by setting a personality flag and re-executing. (We need + * to re-execute, since the memory maps that can conflict with + * the SBCL spaces have already been done at this point). + * + * Since randomization is currently implemented only on x86 kernels, + * don't do this trick on other platforms. + */ +#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64) + if ((major_version == 2 + /* Some old kernels will apparently lose unsupported personality flags + * on exec() */ + && ((minor_version == 6 && patch_version >= 11) + || (minor_version > 6) + /* This is what RHEL 3 reports */ + || (minor_version == 4 && patch_version > 20))) + || major_version >= 3) + { + int pers = personality(0xffffffffUL); + /* 0x40000 aka. ADDR_NO_RANDOMIZE */ + if (!(pers & 0x40000)) { + int retval = personality(pers | 0x40000); + /* Allegedly some Linux kernels (the reported case was + * "hardened Linux 2.6.7") won't set the new personality, + * but nor will they return -1 for an error. So as a + * workaround query the new personality... + */ + int newpers = personality(0xffffffffUL); + /* ... and don't re-execute if either the setting resulted + * in an error or if the value didn't change. Otherwise + * this might result in an infinite loop. + */ + + if (!getenv("SBCL_IS_RESTARTING") && + retval != -1 && newpers != pers) { + /* Use /proc/self/exe instead of trying to figure out + * the executable path from PATH and argv[0], since + * that's unreliable. We follow the symlink instead of + * executing the file directly in order to prevent top + * from displaying the name of the process as "exe". */ + char runtime[PATH_MAX+1]; + int i = readlink("/proc/self/exe", runtime, PATH_MAX); + if (i != -1) { + environ = envp; + setenv("SBCL_IS_RESTARTING", "T", 1); + runtime[i] = '\0'; + execv(runtime, argv); + } + } + /* Either changing the personality or execve() failed. Either + * way we might as well continue, and hope that the random + * memory maps are ok this time around. + */ + fprintf(stderr, "WARNING:\ +\nCouldn't re-execute SBCL with proper personality flags (/proc isn't mounted? setuid?)\ +\nTrying to continue anyway.\n"); + } else { + unsetenv("SBCL_IS_RESTARTING"); + } + } +#ifdef LISP_FEATURE_X86 + /* Use SSE detector. Recent versions of Linux enable SSE support + * on SSE capable CPUs. */ + /* FIXME: Are there any old versions that does not support SSE? */ + fast_bzero_pointer = fast_bzero_detect; +#endif +#endif } + +#ifdef LISP_FEATURE_ALPHA +/* The Alpha is a 64 bit CPU. SBCL is a 32 bit application. Due to all + * the places that assume we can get a pointer into a fixnum with no + * information loss, we have to make sure it allocates all its ram in the + * 0-2Gb region. */ + +static void * under_2gb_free_pointer=DYNAMIC_1_SPACE_END; +#endif + os_vm_address_t os_validate(os_vm_address_t addr, os_vm_size_t len) { - if (addr) { - int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED; - os_vm_address_t base_addr = addr; - do { - /* KLUDGE: It looks as though this code allocates memory - * in chunks of size no larger than 'magic', but why? What - * is the significance of 0x1000000 here? Also, can it be - * right that if the first few 'do_mmap' calls succeed, - * then one fails, we leave the memory allocated by the - * first few in place even while we return a code for - * complete failure? -- WHN 19991020 - * - * Peter Van Eynde writes (20000211) - * This was done because the kernel would only check for - * overcommit for every allocation seperately. So if you - * had 16MB of free mem+swap you could allocate 16M. And - * again, and again, etc. - * This in [Linux] 2.X could be bad as they changed the memory - * system. A side effect was/is (I don't really know) that - * programs with a lot of memory mappings run slower. But - * of course for 2.2.2X we now have the NO_RESERVE flag that - * helps... - * - * FIXME: The logic is also flaky w.r.t. failed - * allocations. If we make one or more successful calls to - * do_mmap(..) before one fails, then we've allocated - * memory, and we should ensure that it gets deallocated - * sometime somehow. If this function's response to any - * failed do_mmap(..) is to give up and return NULL (as in - * sbcl-0.6.7), then any failed do_mmap(..) after any - * successful do_mmap(..) causes a memory leak. */ - int magic = 0x1000000; - if (len <= magic) { - if (do_mmap(&addr, len, flags)) { - return NULL; - } - len = 0; - } else { - if (do_mmap(&addr, magic, flags)) { - return NULL; - } - addr += magic; - len = len - magic; - } - } while (len > 0); - return base_addr; - } else { - int flags = MAP_PRIVATE | MAP_ANONYMOUS; - if (do_mmap(&addr, len, flags)) { - return NULL; - } else { - return addr; - } + int flags = MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE; + os_vm_address_t actual; + +#ifdef LISP_FEATURE_ALPHA + if (!addr) { + addr=under_2gb_free_pointer; + } +#endif + actual = mmap(addr, len, OS_VM_PROT_ALL, flags, -1, 0); + if (actual == MAP_FAILED) { + perror("mmap"); + return 0; /* caller should check this */ + } + + if (addr && (addr!=actual)) { + fprintf(stderr, "mmap: wanted %lu bytes at %p, actually mapped at %p\n", + (unsigned long) len, addr, actual); + return 0; } + +#ifdef LISP_FEATURE_ALPHA + + len=(len+(os_vm_page_size-1))&(~(os_vm_page_size-1)); + under_2gb_free_pointer+=len; +#endif + + return actual; } void os_invalidate(os_vm_address_t addr, os_vm_size_t len) { if (munmap(addr,len) == -1) { - perror("munmap"); + perror("munmap"); } } os_vm_address_t os_map(int fd, int offset, os_vm_address_t addr, os_vm_size_t len) { - addr = mmap(addr, len, - OS_VM_PROT_ALL, - MAP_PRIVATE | MAP_FILE | MAP_FIXED, - fd, (off_t) offset); - - if(addr == MAP_FAILED) { - perror("mmap"); - lose("unexpected mmap(..) failure"); + os_vm_address_t actual; + + actual = mmap(addr, len, OS_VM_PROT_ALL, MAP_PRIVATE | MAP_FIXED, + fd, (off_t) offset); + if (actual == MAP_FAILED || (addr && (addr != actual))) { + perror("mmap"); + lose("unexpected mmap(..) failure\n"); } - return addr; + return actual; } void os_protect(os_vm_address_t address, os_vm_size_t length, os_vm_prot_t prot) { if (mprotect(address, length, prot) == -1) { - perror("mprotect"); + if (errno == ENOMEM) { + lose("An mprotect call failed with ENOMEM. This probably means that the maximum amount\n" + "of separate memory mappings was exceeded. To fix the problem, either increase\n" + "the maximum with e.g. 'echo 262144 > /proc/sys/vm/max_map_count' or recompile\n" + "SBCL with a larger value for GENCGC-CARD-BYTES in\n" + "'src/compiler/target/backend-parms.lisp'."); + } else { + perror("mprotect"); + } } } -/* FIXME: Now that FOO_END, rather than FOO_SIZE, is the fundamental - * description of a space, we could probably punt this and just do - * (FOO_START <= x && x < FOO_END) everywhere it's called. */ -static boolean -in_range_p(os_vm_address_t a, lispobj sbeg, size_t slen) -{ - char* beg = (char*)((long)sbeg); - char* end = (char*)((long)sbeg) + slen; - char* adr = (char*)a; - return (adr >= beg && adr < end); -} - boolean is_valid_lisp_addr(os_vm_address_t addr) { - return - in_range_p(addr, READ_ONLY_SPACE_START, READ_ONLY_SPACE_SIZE) || - in_range_p(addr, STATIC_SPACE_START , STATIC_SPACE_SIZE) || - in_range_p(addr, DYNAMIC_SPACE_START , DYNAMIC_SPACE_SIZE) || - in_range_p(addr, CONTROL_STACK_START , CONTROL_STACK_SIZE) || - in_range_p(addr, BINDING_STACK_START , BINDING_STACK_SIZE); + struct thread *th; + size_t ad = (size_t) addr; + + if ((READ_ONLY_SPACE_START <= ad && ad < READ_ONLY_SPACE_END) + || (STATIC_SPACE_START <= ad && ad < STATIC_SPACE_END) +#if defined LISP_FEATURE_GENCGC + || (DYNAMIC_SPACE_START <= ad && ad < DYNAMIC_SPACE_END) +#else + || (DYNAMIC_0_SPACE_START <= ad && ad < DYNAMIC_0_SPACE_END) + || (DYNAMIC_1_SPACE_START <= ad && ad < DYNAMIC_1_SPACE_END) +#endif + ) + return 1; + for_each_thread(th) { + if((size_t)(th->control_stack_start) <= ad + && ad < (size_t)(th->control_stack_end)) + return 1; + if((size_t)(th->binding_stack_start) <= ad + && ad < (size_t)(th->binding_stack_start + BINDING_STACK_SIZE)) + return 1; + } + return 0; } /* * any OS-dependent special low-level handling for signals */ -#if defined GENCGC - /* - * The GENCGC needs to be hooked into whatever signal is raised for + * The GC needs to be hooked into whatever signal is raised for * page fault on this OS. */ -void -sigsegv_handler(int signal, siginfo_t *info, void* void_context) +static void +sigsegv_handler(int signal, siginfo_t *info, os_context_t *context) { - os_context_t *context = (os_context_t*)void_context; - void* fault_addr = (void*)context->uc_mcontext.cr2; - if (!gencgc_handle_wp_violation(fault_addr)) { - interrupt_handle_now(signal, info, void_context); + os_vm_address_t addr = arch_get_bad_addr(signal, info, context); + +#ifdef LISP_FEATURE_ALPHA + /* Alpha stuff: This is the end of a pseudo-atomic section during + which a signal was received. We must deal with the pending + interrupt (see also interrupt.c, ../code/interrupt.lisp) + + (how we got here: when interrupting, we set bit 63 in reg_ALLOC. + At the end of the atomic section we tried to write to reg_ALLOC, + got a SIGSEGV (there's nothing mapped there) so ended up here. */ + if (addr != NULL && + *os_context_register_addr(context, reg_ALLOC) & (1L<<63)) { + *os_context_register_addr(context, reg_ALLOC) -= (1L<<63); + interrupt_handle_pending(context); + return; } -} +#endif +#ifdef LISP_FEATURE_SB_SAFEPOINT + if (!handle_safepoint_violation(context, addr)) +#endif + +#ifdef LISP_FEATURE_GENCGC + if (!gencgc_handle_wp_violation(addr)) #else + if (!cheneygc_handle_wp_violation(context, addr)) +#endif + if (!handle_guard_page_triggered(context, addr)) + lisp_memory_fault_error(context, addr); +} -static void -sigsegv_handler(int signal, siginfo_t *info, void* void_context) +void +os_install_interrupt_handlers(void) { - os_context_t *context = (os_context_t*)void_context; - os_vm_address_t addr; + undoably_install_low_level_interrupt_handler(SIG_MEMORY_FAULT, + sigsegv_handler); -#ifdef i386 - interrupt_handle_now(signal,contextstruct); -#else -#define CONTROL_STACK_TOP (((char*)CONTROL_STACK_START)+CONTROL_STACK_SIZE) - - addr = arch_get_bad_addr(signal,info,context); - - if(addr != NULL && - *os_context_register_addr(context,reg_ALLOC) & (1L<<63)){ - /* This is the end of a pseudo-atomic section during which - * a signal was received. We must deal with the pending interrupt - * (see also interrupt.c, ../code/interrupt.lisp) - */ - - /* (how we got here: when interrupting, we set bit 63 in - * reg_Alloc. At the end of the atomic section we tried to - * write to reg_Alloc, got a SIGSEGV (there's nothing mapped - * there) so ended up here - */ - *os_context_register_addr(context,reg_ALLOC) -= (1L<<63); - interrupt_handle_pending(context); - } else if (addr > CONTROL_STACK_TOP && addr < BINDING_STACK_START) { - fprintf(stderr, "Possible stack overflow at 0x%016lX: CONTROL_STACK_TOP=%lx, BINDING_STACK_START=%lx\n",addr, CONTROL_STACK_TOP,BINDING_STACK_START); - /* try to fix control frame pointer */ - while ( ! (CONTROL_STACK_START <= *current_control_frame_pointer && - *current_control_frame_pointer <= CONTROL_STACK_TOP)) - ((char*)current_control_frame_pointer) -= sizeof(lispobj); - ldb_monitor(); - } else if (!interrupt_maybe_gc(signal, info, context)) { - interrupt_handle_now(signal, info, context); - } + /* OAOOM c.f. sunos-os.c. + * Should we have a reusable function gc_install_interrupt_handlers? */ +#ifdef LISP_FEATURE_SB_THREAD +# ifdef LISP_FEATURE_SB_SAFEPOINT +# ifdef LISP_FEATURE_SB_THRUPTION + undoably_install_low_level_interrupt_handler(SIGPIPE, thruption_handler); +# endif +# else + undoably_install_low_level_interrupt_handler(SIG_STOP_FOR_GC, + sig_stop_for_gc_handler); +# endif #endif } -#endif +char * +os_get_runtime_executable_path(int external) +{ + char path[PATH_MAX + 1]; + int size; + + size = readlink("/proc/self/exe", path, sizeof(path)-1); + if (size < 0) + return NULL; + else + path[size] = '\0'; + + return copied_string(path); +} + +#ifdef LISP_FEATURE_SB_WTIMER +/* + * Waitable timer implementation for the safepoint-based (SIGALRM-free) + * timer facility using timerfd_create(). + */ +int +os_create_wtimer() +{ + int fd = timerfd_create(CLOCK_MONOTONIC, 0); + if (fd == -1) + lose("os_create_wtimer: timerfd_create"); + + /* Cannot count on TFD_CLOEXEC availability, so do it manually: */ + if (fcntl(fd, F_SETFD, FD_CLOEXEC) == -1) + lose("os_create_wtimer: fcntl"); + return fd; +} + +int +os_wait_for_wtimer(int fd) +{ + unsigned char buf[8]; + int n = read(fd, buf, sizeof(buf)); + if (n == -1) { + if (errno == EINTR) + return -1; + lose("os_wtimer_listen failed"); + } + if (n != sizeof(buf)) + lose("os_wtimer_listen read too little"); + return 0; +} void -os_install_interrupt_handlers(void) +os_close_wtimer(int fd) { - interrupt_install_low_level_handler(SIGSEGV, sigsegv_handler); + if (close(fd) == -1) + lose("os_close_wtimer failed"); } +void +os_set_wtimer(int fd, int sec, int nsec) +{ + struct itimerspec spec = { {0,0}, {0,0} }; + spec.it_value.tv_sec = sec; + spec.it_value.tv_nsec = nsec; + if (timerfd_settime(fd, 0, &spec, 0) == -1) + lose("timerfd_settime"); +} + +void +os_cancel_wtimer(int fd) +{ + os_set_wtimer(fd, 0, 0); +} +#endif