X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Flinux-os.c;h=222f0affec3d6d694ae6e7ab15aa2a07ad7d881f;hb=70227794f1eefb567c13ec04f7bd6d3b6794aa29;hp=9d8ee0ac8b29e26749bf016f51650ef19d7d5993;hpb=5f338d314224411587a7cac218ea320bc982f19f;p=sbcl.git diff --git a/src/runtime/linux-os.c b/src/runtime/linux-os.c index 9d8ee0a..222f0af 100644 --- a/src/runtime/linux-os.c +++ b/src/runtime/linux-os.c @@ -1,5 +1,6 @@ /* - * the Linux incarnation of OS-dependent routines + * the Linux incarnation of OS-dependent routines. See also + * $(sbcl_arch)-linux-os.c * * This file (along with os.h) exports an OS-independent interface to * the operating system VM facilities. Surprise surprise, this @@ -19,20 +20,20 @@ * files for more information. */ -/* - * $Header$ - */ - #include #include #include +#include "sbcl.h" #include "./signal.h" #include "os.h" #include "arch.h" #include "globals.h" #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 @@ -42,252 +43,248 @@ #include #include #include +#include -#include "x86-validate.h" -size_t os_vm_page_size; - -#if defined GENCGC -#include "gencgc.h" +#include "validate.h" +#include "thread.h" +#include "gc.h" +#if defined LISP_FEATURE_GENCGC +#include "gencgc-internal.h" #endif - -void os_init(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); - } - } - os_vm_page_size = getpagesize(); +size_t os_vm_page_size; - SET_FPU_CONTROL_WORD(0x1372|4|8|16|32); /* no interrupts */ -} +#ifdef LISP_FEATURE_SB_THREAD +#include +#include +#include -/* KLUDGE: As of kernel 2.2.14 on Red Hat 6.2, there's code in the - * file to define symbolic names for offsets into - * gregs[], but it's conditional on __USE_GNU and not defined, so - * we need to do this nasty absolute index magic number thing - * instead. */ -int * -os_context_register_addr(os_context_t *context, int offset) -{ - switch(offset) { - case 0: return &context->uc_mcontext.gregs[11]; /* EAX */ - case 2: return &context->uc_mcontext.gregs[10]; /* ECX */ - case 4: return &context->uc_mcontext.gregs[9]; /* EDX */ - case 6: return &context->uc_mcontext.gregs[8]; /* EBX */ - case 8: return &context->uc_mcontext.gregs[7]; /* ESP */ - case 10: return &context->uc_mcontext.gregs[6]; /* EBP */ - case 12: return &context->uc_mcontext.gregs[5]; /* ESI */ - case 14: return &context->uc_mcontext.gregs[4]; /* EDI */ - default: return 0; - } -} -int * -os_context_pc_addr(os_context_t *context) +/* 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 &context->uc_mcontext.gregs[14]; + return syscall (SYS_futex, futex, op, val, rel); } -int * -os_context_sp_addr(os_context_t *context) + +int +futex_wait(int *lock_word, int oldval) { - return &context->uc_mcontext.gregs[17]; + int t= sys_futex(lock_word,FUTEX_WAIT,oldval, 0); + return t; } -sigset_t * -os_context_sigmask_addr(os_context_t *context) +int +futex_wake(int *lock_word, int n) { - return &context->uc_sigmask; + return sys_futex(lock_word,FUTEX_WAKE,n,0); } +#endif -/* 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; + +int linux_sparc_siginfo_bug = 0; +int linux_no_threads_p = 0; -/* Return 0 for success. */ -static int -do_mmap(os_vm_address_t *addr, os_vm_size_t len, int flags) +void +os_init(void) { - /* 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? */ +#ifdef LISP_FEATURE_SB_THREAD + int *futex=0; +#endif + struct utsname name; + int major_version; + int minor_version; + char *p; + uname(&name); + p=name.release; + major_version = atoi(p); + p=strchr(p,'.')+1; + minor_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); } - return 0; + 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) linux_no_threads_p = 1; + if(linux_no_threads_p) + fprintf(stderr,"Linux with NPTL support (e.g. kernel 2.6 or newer) required for \nthread-enabled SBCL. Disabling thread support.\n\n"); +#endif + os_vm_page_size = getpagesize(); } + +#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 || (addr && (addr!=actual))) { + perror("mmap"); + return 0; /* caller should check this */ + } + +#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; - return addr; -} + 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"); + } -void -os_flush_icache(os_vm_address_t address, os_vm_size_t length) -{ + 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*)sbeg; - char* end = (char*)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 -void -os_install_interrupt_handlers(void) -{} - -#else +#if defined LISP_FEATURE_GENCGC /* * The GENCGC 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 = arch_os_get_context(&void_context); + void* fault_addr = (void*)info->si_addr; + if (!gencgc_handle_wp_violation(fault_addr)) + if(!handle_guard_page_triggered(context,fault_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 +} + +#else + +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); + +#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 + + if(!interrupt_maybe_gc(signal, info, context)) + if(!handle_guard_page_triggered(context,addr)) + 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 +}