X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Flinux-os.c;h=d0aa95a478f21036006707708b1db12758783b9e;hb=e4542bc034db18cf98f005b2dac53a6d7d5c7260;hp=3e68e12ca9694dfe8f61e8f9b99b5dfebc7de0fa;hpb=c8322df812da6eb4ef1ae51735b224b2ad0f1503;p=sbcl.git diff --git a/src/runtime/linux-os.c b/src/runtime/linux-os.c index 3e68e12..d0aa95a 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" @@ -30,7 +31,9 @@ #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 @@ -40,255 +43,323 @@ #include #include #include +#include #include "validate.h" +#include "thread.h" +#include "gc.h" +#if defined LISP_FEATURE_GENCGC +#include "gencgc-internal.h" +#endif + +#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 + size_t os_vm_page_size; -#if defined GENCGC -#include "gencgc.h" +#ifdef LISP_FEATURE_SB_THREAD +#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) +#define FUTEX_FD (2) +#define FUTEX_REQUEUE (3) + +#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); +} + +int +futex_wait(int *lock_word, int oldval) +{ + int t= sys_futex(lock_word,FUTEX_WAIT,oldval, 0); + return t; +} + +int +futex_wake(int *lock_word, int n) +{ + return sys_futex(lock_word,FUTEX_WAKE,n,0); +} #endif + -void os_init(void) +int linux_sparc_siginfo_bug = 0; +int linux_no_threads_p = 0; + +#ifdef LISP_FEATURE_SB_THREAD +int +isnptl (void) { - /* 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); - } - } + 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 +void +os_init(char *argv[], char *envp[]) +{ + /* Conduct various version checks: do we have enough mmap(), is + * this a sparc running 2.2, can we do threads? */ +#ifdef LISP_FEATURE_SB_THREAD + int *futex=0; +#endif + struct utsname name; + int major_version; + int minor_version; + int patch_version; + char *p; + uname(&name); + p=name.release; + major_version = atoi(p); + p=strchr(p,'.')+1; + minor_version = atoi(p); + p=strchr(p,'.')+1; + patch_version = atoi(p); + if (major_version<2) { + lose("linux kernel version too old: major version=%d (can't run in version < 2.0.0)", + 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 + futex_wait(futex,-1); + if(errno==ENOSYS) { + lose("This version of SBCL is compiled with threading support, but your kernel is too old to support this.\n\ +Please use a more recent kernel or a version of SBCL without threading support.\n"); + } + if(! isnptl()) { + lose("This version of SBCL only works correctly with the NPTL threading library. Please use a newer glibc, use an older SBCL, or stop using LD_ASSUME_KERNEL"); + } +#endif 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 */ + + /* 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. + */ +#ifdef LISP_FEATURE_X86 + if ((major_version == 2 + /* Some old kernels will apparently lose unsupported personality flags + * on exec() */ + && ((minor_version == 6 && patch_version >= 11) + || (minor_version > 6))) + || 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 (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) { + runtime[i] = '\0'; + execve(runtime, argv, envp); + } + } + /* 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: Couldn't re-execute SBCL with the proper personality flags (maybe /proc isn't mounted?). Trying to continue anyway.\n"); + } + } #endif } -/* various os_context_*_addr accessors moved to {x86,alpha}-linux-os.c - * -dan 20010125 - */ -/* 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_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. */ -/* Return 0 for success. */ -static int -do_mmap(os_vm_address_t *addr, os_vm_size_t len, int flags) -{ - /* We *must* have the memory where we expect 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, - "/retryable error in allocating memory from the OS\n" - "(addr=0x%lx, len=0x%lx, flags=0x%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; - } - return 0; -} +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"); } - 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"); + 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 +static void sigsegv_handler(int signal, siginfo_t *info, void* void_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_context_t *context = arch_os_get_context(&void_context); + os_vm_address_t addr = arch_get_bad_addr(signal, info, context); -#else +#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) -static void -sigsegv_handler(int signal, siginfo_t *info, void* void_context) -{ - os_context_t *context = (os_context_t*)void_context; - os_vm_address_t addr; + (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 i386 - interrupt_handle_now(signal,contextstruct); +#ifdef LISP_FEATURE_GENCGC + if (!gencgc_handle_wp_violation(addr)) #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); - } + if (!interrupt_maybe_gc(signal, info, context)) #endif -} + if (!handle_guard_page_triggered(context, addr)) +#ifdef LISP_FEATURE_C_STACK_IS_CONTROL_STACK + arrange_return_to_lisp_function(context, SymbolFunction(MEMORY_FAULT_ERROR)); +#else + interrupt_handle_now(signal, info, context); #endif - - +} void os_install_interrupt_handlers(void) { - interrupt_install_low_level_handler(SIGSEGV, sigsegv_handler); + undoably_install_low_level_interrupt_handler(SIG_MEMORY_FAULT, + sigsegv_handler); +#ifdef LISP_FEATURE_SB_THREAD + undoably_install_low_level_interrupt_handler(SIG_INTERRUPT_THREAD, + interrupt_thread_handler); + undoably_install_low_level_interrupt_handler(SIG_STOP_FOR_GC, + sig_stop_for_gc_handler); +#endif } -