X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Finterrupt.c;h=294fc64abab1d3689527c27a21a7071eb2c630dd;hb=58513220fcf87e161a5e0d3fbadd76c45f27d584;hp=20dfab2592023c1a34044702a791dd23e83313c2;hpb=08e9f76dffbb455dcbeb42009209493874cd0830;p=sbcl.git diff --git a/src/runtime/interrupt.c b/src/runtime/interrupt.c index 20dfab2..294fc64 100644 --- a/src/runtime/interrupt.c +++ b/src/runtime/interrupt.c @@ -59,26 +59,223 @@ #include "globals.h" #include "lispregs.h" #include "validate.h" +#include "interr.h" #include "gc.h" #include "alloc.h" #include "dynbind.h" -#include "interr.h" +#include "pseudo-atomic.h" #include "genesis/fdefn.h" #include "genesis/simple-fun.h" #include "genesis/cons.h" -static void run_deferred_handler(struct interrupt_data *data, void *v_context); +/* When we catch an internal error, should we pass it back to Lisp to + * be handled in a high-level way? (Early in cold init, the answer is + * 'no', because Lisp is still too brain-dead to handle anything. + * After sufficient initialization has been completed, the answer + * becomes 'yes'.) */ +boolean internal_errors_enabled = 0; + +#ifndef LISP_FEATURE_WIN32 +static +void (*interrupt_low_level_handlers[NSIG]) (int, siginfo_t*, os_context_t*); +#endif +union interrupt_handler interrupt_handlers[NSIG]; + +/* Under Linux on some architectures, we appear to have to restore the + * FPU control word from the context, as after the signal is delivered + * we appear to have a null FPU control word. */ +#if defined(RESTORE_FP_CONTROL_FROM_CONTEXT) +#define RESTORE_FP_CONTROL_WORD(context,void_context) \ + os_context_t *context = arch_os_get_context(&void_context); \ + os_restore_fp_control(context); +#else +#define RESTORE_FP_CONTROL_WORD(context,void_context) \ + os_context_t *context = arch_os_get_context(&void_context); +#endif + +/* Foreign code may want to start some threads on its own. + * Non-targetted, truly asynchronous signals can be delivered to + * basically any thread, but invoking Lisp handlers in such foregign + * threads is really bad, so let's resignal it. + * + * This should at least bring attention to the problem, but it cannot + * work for SIGSEGV and similar. It is good enough for timers, and + * maybe all deferrables. */ + +#ifdef LISP_FEATURE_SB_THREAD +static void +add_handled_signals(sigset_t *sigset) +{ + int i; + for(i = 1; i < NSIG; i++) { + if (!(ARE_SAME_HANDLER(interrupt_low_level_handlers[i], SIG_DFL)) || + !(ARE_SAME_HANDLER(interrupt_handlers[i].c, SIG_DFL))) { + sigaddset(sigset, i); + } + } +} + +void block_signals(sigset_t *what, sigset_t *where, sigset_t *old); +#endif + +static boolean +maybe_resignal_to_lisp_thread(int signal, os_context_t *context) +{ +#ifdef LISP_FEATURE_SB_THREAD + if (!pthread_getspecific(lisp_thread)) { + if (!(sigismember(&deferrable_sigset,signal))) { + corruption_warning_and_maybe_lose + ("Received signal %d in non-lisp thread %lu, resignalling to a lisp thread.", + signal, + pthread_self()); + } + { + sigset_t sigset; + sigemptyset(&sigset); + add_handled_signals(&sigset); + block_signals(&sigset, 0, 0); + block_signals(&sigset, os_context_sigmask_addr(context), 0); + kill(getpid(), signal); + } + return 1; + } else +#endif + return 0; +} + +/* These are to be used in signal handlers. Currently all handlers are + * called from one of: + * + * interrupt_handle_now_handler + * maybe_now_maybe_later + * unblock_me_trampoline + * low_level_handle_now_handler + * low_level_maybe_now_maybe_later + * low_level_unblock_me_trampoline + * + * This gives us a single point of control (or six) over errno, fp + * control word, and fixing up signal context on sparc. + * + * The SPARC/Linux platform doesn't quite do signals the way we want + * them done. The third argument in the handler isn't filled in by the + * kernel properly, so we fix it up ourselves in the + * arch_os_get_context(..) function. -- CSR, 2002-07-23 + */ +#define SAVE_ERRNO(signal,context,void_context) \ + { \ + int _saved_errno = errno; \ + RESTORE_FP_CONTROL_WORD(context,void_context); \ + if (!maybe_resignal_to_lisp_thread(signal, context)) \ + { + +#define RESTORE_ERRNO \ + } \ + errno = _saved_errno; \ + } + +static void run_deferred_handler(struct interrupt_data *data, + os_context_t *context); #ifndef LISP_FEATURE_WIN32 static void store_signal_data_for_later (struct interrupt_data *data, void *handler, int signal, siginfo_t *info, os_context_t *context); + + +/* Generic signal related utilities. */ + +void +get_current_sigmask(sigset_t *sigset) +{ + /* Get the current sigmask, by blocking the empty set. */ + thread_sigmask(SIG_BLOCK, 0, sigset); +} + +void +block_signals(sigset_t *what, sigset_t *where, sigset_t *old) +{ + if (where) { + int i; + if (old) + sigcopyset(old, where); + for(i = 1; i < NSIG; i++) { + if (sigismember(what, i)) + sigaddset(where, i); + } + } else { + thread_sigmask(SIG_BLOCK, what, old); + } +} + +void +unblock_signals(sigset_t *what, sigset_t *where, sigset_t *old) +{ + if (where) { + int i; + if (old) + sigcopyset(old, where); + for(i = 1; i < NSIG; i++) { + if (sigismember(what, i)) + sigdelset(where, i); + } + } else { + thread_sigmask(SIG_UNBLOCK, what, old); + } +} + +static void +print_sigset(sigset_t *sigset) +{ + int i; + for(i = 1; i < NSIG; i++) { + if (sigismember(sigset, i)) + fprintf(stderr, "Signal %d masked\n", i); + } +} + +/* Return 1 is all signals is sigset2 are masked in sigset, return 0 + * if all re unmasked else die. Passing NULL for sigset is a shorthand + * for the current sigmask. */ +boolean +all_signals_blocked_p(sigset_t *sigset, sigset_t *sigset2, + const char *name) +{ +#if !defined(LISP_FEATURE_WIN32) + int i; + boolean has_blocked = 0, has_unblocked = 0; + sigset_t current; + if (sigset == 0) { + get_current_sigmask(¤t); + sigset = ¤t; + } + for(i = 1; i < NSIG; i++) { + if (sigismember(sigset2, i)) { + if (sigismember(sigset, i)) + has_blocked = 1; + else + has_unblocked = 1; + } + } + if (has_blocked && has_unblocked) { + print_sigset(sigset); + lose("some %s signals blocked, some unblocked\n", name); + } + if (has_blocked) + return 1; + else + return 0; +#endif +} + + +/* Deferrables, blockables, gc signals. */ void sigaddset_deferrable(sigset_t *s) { sigaddset(s, SIGHUP); sigaddset(s, SIGINT); + sigaddset(s, SIGTERM); sigaddset(s, SIGQUIT); sigaddset(s, SIGPIPE); sigaddset(s, SIGALRM); @@ -86,146 +283,318 @@ sigaddset_deferrable(sigset_t *s) sigaddset(s, SIGTSTP); sigaddset(s, SIGCHLD); sigaddset(s, SIGIO); +#ifndef LISP_FEATURE_HPUX sigaddset(s, SIGXCPU); sigaddset(s, SIGXFSZ); +#endif sigaddset(s, SIGVTALRM); sigaddset(s, SIGPROF); sigaddset(s, SIGWINCH); +} -#if !((defined(LISP_FEATURE_DARWIN) || defined(LISP_FEATURE_FREEBSD)) && defined(LISP_FEATURE_SB_THREAD)) - sigaddset(s, SIGUSR1); - sigaddset(s, SIGUSR2); -#endif +void +sigaddset_blockable(sigset_t *sigset) +{ + sigaddset_deferrable(sigset); + sigaddset_gc(sigset); +} +void +sigaddset_gc(sigset_t *sigset) +{ #ifdef LISP_FEATURE_SB_THREAD - sigaddset(s, SIG_INTERRUPT_THREAD); + sigaddset(sigset,SIG_STOP_FOR_GC); #endif } +/* initialized in interrupt_init */ +sigset_t deferrable_sigset; +sigset_t blockable_sigset; +sigset_t gc_sigset; + +#endif + +#if !defined(LISP_FEATURE_WIN32) +boolean +deferrables_blocked_p(sigset_t *sigset) +{ + return all_signals_blocked_p(sigset, &deferrable_sigset, "deferrable"); +} +#endif + void -sigaddset_blockable(sigset_t *s) +check_deferrables_unblocked_or_lose(sigset_t *sigset) { - sigaddset_deferrable(s); -#ifdef LISP_FEATURE_SB_THREAD -#ifdef SIG_RESUME_FROM_GC - sigaddset(s, SIG_RESUME_FROM_GC); +#if !defined(LISP_FEATURE_WIN32) + if (deferrables_blocked_p(sigset)) + lose("deferrables blocked\n"); #endif - sigaddset(s, SIG_STOP_FOR_GC); +} + +void +check_deferrables_blocked_or_lose(sigset_t *sigset) +{ +#if !defined(LISP_FEATURE_WIN32) + if (!deferrables_blocked_p(sigset)) + lose("deferrables unblocked\n"); #endif } -/* initialized in interrupt_init */ -static sigset_t deferrable_sigset; -static sigset_t blockable_sigset; +#if !defined(LISP_FEATURE_WIN32) +boolean +blockables_blocked_p(sigset_t *sigset) +{ + return all_signals_blocked_p(sigset, &blockable_sigset, "blockable"); +} #endif void -check_blockables_blocked_or_lose(void) +check_blockables_unblocked_or_lose(sigset_t *sigset) { #if !defined(LISP_FEATURE_WIN32) - /* Get the current sigmask, by blocking the empty set. */ - sigset_t empty,current; - int i; - sigemptyset(&empty); - thread_sigmask(SIG_BLOCK, &empty, ¤t); - for(i = 1; i < NSIG; i++) { - if (sigismember(&blockable_sigset, i) && !sigismember(¤t, i)) - lose("blockable signal %d not blocked\n",i); - } + if (blockables_blocked_p(sigset)) + lose("blockables blocked\n"); #endif } void -check_gc_signals_unblocked_or_lose(void) +check_blockables_blocked_or_lose(sigset_t *sigset) { -#ifdef LISP_FEATURE_SB_THREAD -# if !defined(LISP_FEATURE_WIN32) - /* Get the current sigmask, by blocking the empty set. */ - sigset_t empty,current; - sigemptyset(&empty); - thread_sigmask(SIG_BLOCK, &empty, ¤t); - if (sigismember(¤t, SIG_STOP_FOR_GC)) - lose("SIG_STOP_FOR_GC blocked in thread %p at a bad place\n", - arch_os_get_current_thread()); -# if defined(SIG_RESUME_FROM_GC) - if (sigismember(¤t, SIG_RESUME_FROM_GC)) - lose("SIG_RESUME_FROM_GC blocked in thread %p at a bad place\n", - arch_os_get_current_thread()); -# endif -# endif +#if !defined(LISP_FEATURE_WIN32) + if (!blockables_blocked_p(sigset)) + lose("blockables unblocked\n"); #endif } +#if !defined(LISP_FEATURE_WIN32) +boolean +gc_signals_blocked_p(sigset_t *sigset) +{ + return all_signals_blocked_p(sigset, &gc_sigset, "gc"); +} +#endif + void -unblock_gc_signals(void) +check_gc_signals_unblocked_or_lose(sigset_t *sigset) { -#ifdef LISP_FEATURE_SB_THREAD - sigset_t new; - sigemptyset(&new); -#if defined(SIG_RESUME_FROM_GC) - sigaddset(&new,SIG_RESUME_FROM_GC); +#if !defined(LISP_FEATURE_WIN32) + if (gc_signals_blocked_p(sigset)) + lose("gc signals blocked\n"); #endif - sigaddset(&new,SIG_STOP_FOR_GC); - thread_sigmask(SIG_UNBLOCK,&new,0); +} + +void +check_gc_signals_blocked_or_lose(sigset_t *sigset) +{ +#if !defined(LISP_FEATURE_WIN32) + if (!gc_signals_blocked_p(sigset)) + lose("gc signals unblocked\n"); #endif } -inline static void -check_interrupts_enabled_or_lose(os_context_t *context) +void +block_deferrable_signals(sigset_t *where, sigset_t *old) { - struct thread *thread=arch_os_get_current_thread(); - if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) - lose("interrupts not enabled\n"); - if ( -#ifdef FOREIGN_FUNCTION_CALL_FLAG - (!foreign_function_call_active) && +#ifndef LISP_FEATURE_WIN32 + block_signals(&deferrable_sigset, where, old); #endif - arch_pseudo_atomic_atomic(context)) - lose ("in pseudo atomic section\n"); } -/* When we catch an internal error, should we pass it back to Lisp to - * be handled in a high-level way? (Early in cold init, the answer is - * 'no', because Lisp is still too brain-dead to handle anything. - * After sufficient initialization has been completed, the answer - * becomes 'yes'.) */ -boolean internal_errors_enabled = 0; +void +block_blockable_signals(sigset_t *where, sigset_t *old) +{ +#ifndef LISP_FEATURE_WIN32 + block_signals(&blockable_sigset, where, old); +#endif +} +void +block_gc_signals(sigset_t *where, sigset_t *old) +{ #ifndef LISP_FEATURE_WIN32 -static void (*interrupt_low_level_handlers[NSIG]) (int, siginfo_t*, void*); + block_signals(&gc_sigset, where, old); #endif -union interrupt_handler interrupt_handlers[NSIG]; +} -/* At the toplevel repl we routinely call this function. The signal - * mask ought to be clear anyway most of the time, but may be non-zero - * if we were interrupted e.g. while waiting for a queue. */ +void +unblock_deferrable_signals(sigset_t *where, sigset_t *old) +{ +#ifndef LISP_FEATURE_WIN32 + if (interrupt_handler_pending_p()) + lose("unblock_deferrable_signals: losing proposition\n"); + check_gc_signals_unblocked_or_lose(where); + unblock_signals(&deferrable_sigset, where, old); +#endif +} void -reset_signal_mask(void) +unblock_blockable_signals(sigset_t *where, sigset_t *old) { #ifndef LISP_FEATURE_WIN32 - sigset_t new; - sigemptyset(&new); - thread_sigmask(SIG_SETMASK,&new,0); + unblock_signals(&blockable_sigset, where, old); #endif } void -block_blockable_signals(void) +unblock_gc_signals(sigset_t *where, sigset_t *old) { #ifndef LISP_FEATURE_WIN32 - thread_sigmask(SIG_BLOCK, &blockable_sigset, 0); + unblock_signals(&gc_sigset, where, old); #endif } void -block_deferrable_signals(void) +unblock_signals_in_context_and_maybe_warn(os_context_t *context) { #ifndef LISP_FEATURE_WIN32 - thread_sigmask(SIG_BLOCK, &deferrable_sigset, 0); + sigset_t *sigset = os_context_sigmask_addr(context); + if (all_signals_blocked_p(sigset, &gc_sigset, "gc")) { + corruption_warning_and_maybe_lose( +"Enabling blocked gc signals to allow returning to Lisp without risking\n\ +gc deadlocks. Since GC signals are only blocked in signal handlers when \n\ +they are not safe to interrupt at all, this is a pretty severe occurrence.\n"); + unblock_gc_signals(sigset, 0); + } + if (!interrupt_handler_pending_p()) { + unblock_deferrable_signals(sigset, 0); + } #endif } + +inline static void +check_interrupts_enabled_or_lose(os_context_t *context) +{ + struct thread *thread=arch_os_get_current_thread(); + if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) + lose("interrupts not enabled\n"); + if (arch_pseudo_atomic_atomic(context)) + lose ("in pseudo atomic section\n"); +} + +/* Save sigset (or the current sigmask if 0) if there is no pending + * handler, because that means that deferabbles are already blocked. + * The purpose is to avoid losing the pending gc signal if a + * deferrable interrupt async unwinds between clearing the pseudo + * atomic and trapping to GC.*/ +void +maybe_save_gc_mask_and_block_deferrables(sigset_t *sigset) +{ +#ifndef LISP_FEATURE_WIN32 + struct thread *thread = arch_os_get_current_thread(); + struct interrupt_data *data = thread->interrupt_data; + sigset_t oldset; + /* Obviously, this function is called when signals may not be + * blocked. Let's make sure we are not interrupted. */ + block_blockable_signals(0, &oldset); +#ifndef LISP_FEATURE_SB_THREAD + /* With threads a SIG_STOP_FOR_GC and a normal GC may also want to + * block. */ + if (data->gc_blocked_deferrables) + lose("gc_blocked_deferrables already true\n"); +#endif + if ((!data->pending_handler) && + (!data->gc_blocked_deferrables)) { + FSHOW_SIGNAL((stderr,"/setting gc_blocked_deferrables\n")); + data->gc_blocked_deferrables = 1; + if (sigset) { + /* This is the sigmask of some context. */ + sigcopyset(&data->pending_mask, sigset); + sigaddset_deferrable(sigset); + thread_sigmask(SIG_SETMASK,&oldset,0); + return; + } else { + /* Operating on the current sigmask. Save oldset and + * unblock gc signals. In the end, this is equivalent to + * blocking the deferrables. */ + sigcopyset(&data->pending_mask, &oldset); + thread_sigmask(SIG_UNBLOCK, &gc_sigset, 0); + return; + } + } + thread_sigmask(SIG_SETMASK,&oldset,0); +#endif +} + +/* Are we leaving WITH-GCING and already running with interrupts + * enabled, without the protection of *GC-INHIBIT* T and there is gc + * (or stop for gc) pending, but we haven't trapped yet? */ +int +in_leaving_without_gcing_race_p(struct thread *thread) +{ + return ((SymbolValue(IN_WITHOUT_GCING,thread) != NIL) && + (SymbolValue(INTERRUPTS_ENABLED,thread) != NIL) && + (SymbolValue(GC_INHIBIT,thread) == NIL) && + ((SymbolValue(GC_PENDING,thread) != NIL) +#if defined(LISP_FEATURE_SB_THREAD) + || (SymbolValue(STOP_FOR_GC_PENDING,thread) != NIL) +#endif + )); +} + +/* Check our baroque invariants. */ +void +check_interrupt_context_or_lose(os_context_t *context) +{ +#ifndef LISP_FEATURE_WIN32 + struct thread *thread = arch_os_get_current_thread(); + struct interrupt_data *data = thread->interrupt_data; + int interrupt_deferred_p = (data->pending_handler != 0); + int interrupt_pending = (SymbolValue(INTERRUPT_PENDING,thread) != NIL); + sigset_t *sigset = os_context_sigmask_addr(context); + /* On PPC pseudo_atomic_interrupted is cleared when coming out of + * handle_allocation_trap. */ +#if defined(LISP_FEATURE_GENCGC) && !defined(LISP_FEATURE_PPC) + int interrupts_enabled = (SymbolValue(INTERRUPTS_ENABLED,thread) != NIL); + int gc_inhibit = (SymbolValue(GC_INHIBIT,thread) != NIL); + int gc_pending = (SymbolValue(GC_PENDING,thread) == T); + int pseudo_atomic_interrupted = get_pseudo_atomic_interrupted(thread); + int in_race_p = in_leaving_without_gcing_race_p(thread); + /* In the time window between leaving the *INTERRUPTS-ENABLED* NIL + * section and trapping, a SIG_STOP_FOR_GC would see the next + * check fail, for this reason sig_stop_for_gc handler does not + * call this function. */ + if (interrupt_deferred_p) { + if (!(!interrupts_enabled || pseudo_atomic_interrupted || in_race_p)) + lose("Stray deferred interrupt.\n"); + } + if (gc_pending) + if (!(pseudo_atomic_interrupted || gc_inhibit || in_race_p)) + lose("GC_PENDING, but why?\n"); +#if defined(LISP_FEATURE_SB_THREAD) + { + int stop_for_gc_pending = + (SymbolValue(STOP_FOR_GC_PENDING,thread) != NIL); + if (stop_for_gc_pending) + if (!(pseudo_atomic_interrupted || gc_inhibit || in_race_p)) + lose("STOP_FOR_GC_PENDING, but why?\n"); + if (pseudo_atomic_interrupted) + if (!(gc_pending || stop_for_gc_pending || interrupt_deferred_p)) + lose("pseudo_atomic_interrupted, but why?\n"); + } +#else + if (pseudo_atomic_interrupted) + if (!(gc_pending || interrupt_deferred_p)) + lose("pseudo_atomic_interrupted, but why?\n"); +#endif +#endif + if (interrupt_pending && !interrupt_deferred_p) + lose("INTERRUPT_PENDING but not pending handler.\n"); + if ((data->gc_blocked_deferrables) && interrupt_pending) + lose("gc_blocked_deferrables and interrupt pending\n."); + if (data->gc_blocked_deferrables) + check_deferrables_blocked_or_lose(sigset); + if (interrupt_pending || interrupt_deferred_p || + data->gc_blocked_deferrables) + check_deferrables_blocked_or_lose(sigset); + else { + check_deferrables_unblocked_or_lose(sigset); + /* If deferrables are unblocked then we are open to signals + * that run lisp code. */ + check_gc_signals_unblocked_or_lose(sigset); + } +#endif +} /* * utility routines used by various signal handlers @@ -293,19 +662,26 @@ fake_foreign_function_call(os_context_t *context) struct thread *thread=arch_os_get_current_thread(); /* context_index incrementing must not be interrupted */ - check_blockables_blocked_or_lose(); + check_blockables_blocked_or_lose(0); /* Get current Lisp state from context. */ #ifdef reg_ALLOC dynamic_space_free_pointer = (lispobj *)(unsigned long) (*os_context_register_addr(context, reg_ALLOC)); - /* fprintf(stderr,"dynamic_space_free_pointer: %p\n", dynamic_space_free_pointer); */ -#if defined(LISP_FEATURE_ALPHA) +/* fprintf(stderr,"dynamic_space_free_pointer: %p\n", */ +/* dynamic_space_free_pointer); */ +#if defined(LISP_FEATURE_ALPHA) || defined(LISP_FEATURE_MIPS) if ((long)dynamic_space_free_pointer & 1) { lose("dead in fake_foreign_function_call, context = %x\n", context); } #endif +/* why doesnt PPC and SPARC do something like this: */ +#if defined(LISP_FEATURE_HPPA) + if ((long)dynamic_space_free_pointer & 4) { + lose("dead in fake_foreign_function_call, context = %x, d_s_f_p = %x\n", context, dynamic_space_free_pointer); + } +#endif #endif #ifdef reg_BSP current_binding_stack_pointer = @@ -342,7 +718,7 @@ undo_fake_foreign_function_call(os_context_t *context) { struct thread *thread=arch_os_get_current_thread(); /* Block all blockable signals. */ - block_blockable_signals(); + block_blockable_signals(0, 0); #ifdef FOREIGN_FUNCTION_CALL_FLAG foreign_function_call_active = 0; @@ -358,7 +734,8 @@ undo_fake_foreign_function_call(os_context_t *context) | (*os_context_register_addr(context, reg_ALLOC) & LOWTAG_MASK); /* - ((unsigned long)(*os_context_register_addr(context, reg_ALLOC)) & ~LOWTAG_MASK) + ((unsigned long)(*os_context_register_addr(context, reg_ALLOC)) + & ~LOWTAG_MASK) | ((unsigned long) dynamic_space_free_pointer & LOWTAG_MASK); */ #endif @@ -382,20 +759,31 @@ interrupt_internal_error(os_context_t *context, boolean continuable) /* Allocate the SAP object while the interrupts are still * disabled. */ + unblock_gc_signals(0, 0); context_sap = alloc_sap(context); #ifndef LISP_FEATURE_WIN32 thread_sigmask(SIG_SETMASK, os_context_sigmask_addr(context), 0); #endif +#if defined(LISP_FEATURE_LINUX) && defined(LISP_FEATURE_MIPS) + /* Workaround for blocked SIGTRAP. */ + { + sigset_t newset; + sigemptyset(&newset); + sigaddset(&newset, SIGTRAP); + thread_sigmask(SIG_UNBLOCK, &newset, 0); + } +#endif + SHOW("in interrupt_internal_error"); -#ifdef QSHOW +#if QSHOW /* Display some rudimentary debugging information about the * error, so that even if the Lisp error handler gets badly * confused, we have a chance to determine what's going on. */ describe_internal_error(context); #endif - funcall2(SymbolFunction(INTERNAL_ERROR), context_sap, + funcall2(StaticSymbolFunction(INTERNAL_ERROR), context_sap, continuable ? T : NIL); undo_fake_foreign_function_call(context); /* blocks signals again */ @@ -403,105 +791,165 @@ interrupt_internal_error(os_context_t *context, boolean continuable) arch_skip_instruction(context); } +boolean +interrupt_handler_pending_p(void) +{ + struct thread *thread = arch_os_get_current_thread(); + struct interrupt_data *data = thread->interrupt_data; + return (data->pending_handler != 0); +} + void interrupt_handle_pending(os_context_t *context) { - /* There are three ways we can get here. First, if an interrupt + /* There are three ways we can get here. First, if an interrupt * occurs within pseudo-atomic, it will be deferred, and we'll - * trap to here at the end of the pseudo-atomic block. Second, if + * trap to here at the end of the pseudo-atomic block. Second, if * the GC (in alloc()) decides that a GC is required, it will set - * *GC-PENDING* and pseudo-atomic-interrupted, and alloc() is - * always called from within pseudo-atomic, and thus we end up - * here again. Third, when calling GC-ON or at the end of a - * WITHOUT-GCING, MAYBE-HANDLE-PENDING-GC will trap to here if - * there is a pending GC. */ + * *GC-PENDING* and pseudo-atomic-interrupted if not *GC-INHIBIT*, + * and alloc() is always called from within pseudo-atomic, and + * thus we end up here again. Third, when calling GC-ON or at the + * end of a WITHOUT-GCING, MAYBE-HANDLE-PENDING-GC will trap to + * here if there is a pending GC. Fourth, ahem, at the end of + * WITHOUT-INTERRUPTS (bar complications with nesting). */ /* Win32 only needs to handle the GC cases (for now?) */ struct thread *thread = arch_os_get_current_thread(); + struct interrupt_data *data = thread->interrupt_data; - FSHOW_SIGNAL((stderr, "/entering interrupt_handle_pending\n")); + if (arch_pseudo_atomic_atomic(context)) { + lose("Handling pending interrupt in pseudo atomic."); + } - check_blockables_blocked_or_lose(); + FSHOW_SIGNAL((stderr, "/entering interrupt_handle_pending\n")); - /* If pseudo_atomic_interrupted is set then the interrupt is going - * to be handled now, ergo it's safe to clear it. */ - arch_clear_pseudo_atomic_interrupted(context); + check_blockables_blocked_or_lose(0); + + /* If GC/SIG_STOP_FOR_GC struck during PA and there was no pending + * handler, then the pending mask was saved and + * gc_blocked_deferrables set. Hence, there can be no pending + * handler and it's safe to restore the pending mask. + * + * Note, that if gc_blocked_deferrables is false we may still have + * to GC. In this case, we are coming out of a WITHOUT-GCING or a + * pseudo atomic was interrupt be a deferrable first. */ + if (data->gc_blocked_deferrables) { + if (data->pending_handler) + lose("GC blocked deferrables but still got a pending handler."); + if (SymbolValue(GC_INHIBIT,thread)!=NIL) + lose("GC blocked deferrables while GC is inhibited."); + /* Restore the saved signal mask from the original signal (the + * one that interrupted us during the critical section) into + * the os_context for the signal we're currently in the + * handler for. This should ensure that when we return from + * the handler the blocked signals are unblocked. */ +#ifndef LISP_FEATURE_WIN32 + sigcopyset(os_context_sigmask_addr(context), &data->pending_mask); +#endif + data->gc_blocked_deferrables = 0; + } if (SymbolValue(GC_INHIBIT,thread)==NIL) { + void *original_pending_handler = data->pending_handler; + #ifdef LISP_FEATURE_SB_THREAD if (SymbolValue(STOP_FOR_GC_PENDING,thread) != NIL) { - /* another thread has already initiated a gc, this attempt - * might as well be cancelled */ - SetSymbolValue(GC_PENDING,NIL,thread); - SetSymbolValue(STOP_FOR_GC_PENDING,NIL,thread); + /* STOP_FOR_GC_PENDING and GC_PENDING are cleared by + * the signal handler if it actually stops us. */ + arch_clear_pseudo_atomic_interrupted(context); sig_stop_for_gc_handler(SIG_STOP_FOR_GC,NULL,context); } else #endif - if (SymbolValue(GC_PENDING,thread) != NIL) { + /* Test for T and not for != NIL since the value :IN-PROGRESS + * is used in SUB-GC as part of the mechanism to supress + * recursive gcs.*/ + if (SymbolValue(GC_PENDING,thread) == T) { + + /* Two reasons for doing this. First, if there is a + * pending handler we don't want to run. Second, we are + * going to clear pseudo atomic interrupted to avoid + * spurious trapping on every allocation in SUB_GC and + * having a pending handler with interrupts enabled and + * without pseudo atomic interrupted breaks an + * invariant. */ + if (data->pending_handler) { + bind_variable(ALLOW_WITH_INTERRUPTS, NIL, thread); + bind_variable(INTERRUPTS_ENABLED, NIL, thread); + } + + arch_clear_pseudo_atomic_interrupted(context); + /* GC_PENDING is cleared in SUB-GC, or if another thread * is doing a gc already we will get a SIG_STOP_FOR_GC and - * that will clear it. */ - maybe_gc(context); + * that will clear it. + * + * If there is a pending handler or gc was triggerred in a + * signal handler then maybe_gc won't run POST_GC and will + * return normally. */ + if (!maybe_gc(context)) + lose("GC not inhibited but maybe_gc did not GC."); + + if (data->pending_handler) { + unbind(thread); + unbind(thread); + } + } else if (SymbolValue(GC_PENDING,thread) != NIL) { + /* It's not NIL or T so GC_PENDING is :IN-PROGRESS. If + * GC-PENDING is not NIL then we cannot trap on pseudo + * atomic due to GC (see if(GC_PENDING) logic in + * cheneygc.c an gengcgc.c), plus there is a outer + * WITHOUT-INTERRUPTS SUB_GC, so how did we end up + * here? */ + lose("Trapping to run pending handler while GC in progress."); } - check_blockables_blocked_or_lose(); + + check_blockables_blocked_or_lose(0); + + /* No GC shall be lost. If SUB_GC triggers another GC then + * that should be handled on the spot. */ + if (SymbolValue(GC_PENDING,thread) != NIL) + lose("GC_PENDING after doing gc."); +#ifdef LISP_FEATURE_SB_THREAD + if (SymbolValue(STOP_FOR_GC_PENDING,thread) != NIL) + lose("STOP_FOR_GC_PENDING after doing gc."); +#endif + /* Check two things. First, that gc does not clobber a handler + * that's already pending. Second, that there is no interrupt + * lossage: if original_pending_handler was NULL then even if + * an interrupt arrived during GC (POST-GC, really) it was + * handled. */ + if (original_pending_handler != data->pending_handler) + lose("pending handler changed in gc: %x -> %d.", + original_pending_handler, data->pending_handler); } #ifndef LISP_FEATURE_WIN32 - /* we may be here only to do the gc stuff, if interrupts are - * enabled run the pending handler */ - if (!((SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) || - ( -#ifdef FOREIGN_FUNCTION_CALL_FLAG - (!foreign_function_call_active) && -#endif - arch_pseudo_atomic_atomic(context)))) { - struct interrupt_data *data = thread->interrupt_data; - - /* There may be no pending handler, because it was only a gc - * that had to be executed or because pseudo atomic triggered - * twice for a single interrupt. For the interested reader, - * that may happen if an interrupt hits after the interrupted - * flag is cleared but before pseduo-atomic is set and a - * pseudo atomic is interrupted in that interrupt. */ - if (data->pending_handler) { - - /* If we're here as the result of a pseudo-atomic as opposed - * to WITHOUT-INTERRUPTS, then INTERRUPT_PENDING is already - * NIL, because maybe_defer_handler sets - * PSEUDO_ATOMIC_INTERRUPTED only if interrupts are enabled.*/ - SetSymbolValue(INTERRUPT_PENDING, NIL,thread); - - /* restore the saved signal mask from the original signal (the - * one that interrupted us during the critical section) into the - * os_context for the signal we're currently in the handler for. - * This should ensure that when we return from the handler the - * blocked signals are unblocked */ - sigcopyset(os_context_sigmask_addr(context), &data->pending_mask); - - sigemptyset(&data->pending_mask); - /* This will break on sparc linux: the deferred handler really wants - * to be called with a void_context */ - run_deferred_handler(data,(void *)context); - } + /* There may be no pending handler, because it was only a gc that + * had to be executed or because Lisp is a bit too eager to call + * DO-PENDING-INTERRUPT. */ + if ((SymbolValue(INTERRUPTS_ENABLED,thread) != NIL) && + (data->pending_handler)) { + /* No matter how we ended up here, clear both + * INTERRUPT_PENDING and pseudo atomic interrupted. It's safe + * because we checked above that there is no GC pending. */ + SetSymbolValue(INTERRUPT_PENDING, NIL, thread); + arch_clear_pseudo_atomic_interrupted(context); + /* Restore the sigmask in the context. */ + sigcopyset(os_context_sigmask_addr(context), &data->pending_mask); + run_deferred_handler(data, context); } #endif +#ifdef LISP_FEATURE_GENCGC + if (get_pseudo_atomic_interrupted(thread)) + lose("pseudo_atomic_interrupted after interrupt_handle_pending\n"); +#endif + /* It is possible that the end of this function was reached + * without never actually doing anything, the tests in Lisp for + * when to call receive-pending-interrupt are not exact. */ + FSHOW_SIGNAL((stderr, "/exiting interrupt_handle_pending\n")); } -/* - * the two main signal handlers: - * interrupt_handle_now(..) - * maybe_now_maybe_later(..) - * - * to which we have added interrupt_handle_now_handler(..). Why? - * Well, mostly because the SPARC/Linux platform doesn't quite do - * signals the way we want them done. The third argument in the - * handler isn't filled in by the kernel properly, so we fix it up - * ourselves in the arch_os_get_context(..) function; however, we only - * want to do this when we first hit the handler, and not when - * interrupt_handle_now(..) is being called from some other handler - * (when the fixup will already have been done). -- CSR, 2002-07-23 - */ void interrupt_handle_now(int signal, siginfo_t *info, os_context_t *context) @@ -511,20 +959,13 @@ interrupt_handle_now(int signal, siginfo_t *info, os_context_t *context) #endif union interrupt_handler handler; - check_blockables_blocked_or_lose(); + check_blockables_blocked_or_lose(0); #ifndef LISP_FEATURE_WIN32 if (sigismember(&deferrable_sigset,signal)) check_interrupts_enabled_or_lose(context); #endif -#if defined(LISP_FEATURE_LINUX) || defined(RESTORE_FP_CONTROL_FROM_CONTEXT) - /* Under Linux on some architectures, we appear to have to restore - the FPU control word from the context, as after the signal is - delivered we appear to have a null FPU control word. */ - os_restore_fp_control(context); -#endif - handler = interrupt_handlers[signal]; if (ARE_SAME_HANDLER(handler.c, SIG_IGN)) { @@ -566,21 +1007,12 @@ interrupt_handle_now(int signal, siginfo_t *info, os_context_t *context) * Yeah, but non-gencgc platforms don't really wrap allocation * in PA. MG - 2005-08-29 */ - lispobj info_sap,context_sap = alloc_sap(context); - info_sap = alloc_sap(info); + lispobj info_sap, context_sap; /* Leave deferrable signals blocked, the handler itself will * allow signals again when it sees fit. */ -#ifdef LISP_FEATURE_SB_THREAD - { - sigset_t unblock; - sigemptyset(&unblock); - sigaddset(&unblock, SIG_STOP_FOR_GC); -#ifdef SIG_RESUME_FROM_GC - sigaddset(&unblock, SIG_RESUME_FROM_GC); -#endif - thread_sigmask(SIG_UNBLOCK, &unblock, 0); - } -#endif + unblock_gc_signals(0, 0); + context_sap = alloc_sap(context); + info_sap = alloc_sap(info); FSHOW_SIGNAL((stderr,"/calling Lisp-level handler\n")); @@ -589,6 +1021,7 @@ interrupt_handle_now(int signal, siginfo_t *info, os_context_t *context) info_sap, context_sap); } else { + /* This cannot happen in sane circumstances. */ FSHOW_SIGNAL((stderr,"/calling C-level handler\n")); @@ -617,15 +1050,18 @@ interrupt_handle_now(int signal, siginfo_t *info, os_context_t *context) * already; we're just doing the Lisp-level processing now that we * put off then */ static void -run_deferred_handler(struct interrupt_data *data, void *v_context) { +run_deferred_handler(struct interrupt_data *data, os_context_t *context) +{ /* The pending_handler may enable interrupts and then another * interrupt may hit, overwrite interrupt_data, so reset the * pending handler before calling it. Trust the handler to finish * with the siginfo before enabling interrupts. */ - void (*pending_handler) (int, siginfo_t*, void*)=data->pending_handler; + void (*pending_handler) (int, siginfo_t*, os_context_t*) = + data->pending_handler; data->pending_handler=0; - (*pending_handler)(data->pending_signal,&(data->pending_info), v_context); + FSHOW_SIGNAL((stderr, "/running deferred handler %p\n", pending_handler)); + (*pending_handler)(data->pending_signal,&(data->pending_info), context); } #ifndef LISP_FEATURE_WIN32 @@ -635,49 +1071,49 @@ maybe_defer_handler(void *handler, struct interrupt_data *data, { struct thread *thread=arch_os_get_current_thread(); - check_blockables_blocked_or_lose(); + check_blockables_blocked_or_lose(0); if (SymbolValue(INTERRUPT_PENDING,thread) != NIL) lose("interrupt already pending\n"); + if (thread->interrupt_data->pending_handler) + lose("there is a pending handler already (PA)\n"); + if (data->gc_blocked_deferrables) + lose("maybe_defer_handler: gc_blocked_deferrables true\n"); + check_interrupt_context_or_lose(context); /* If interrupts are disabled then INTERRUPT_PENDING is set and * not PSEDUO_ATOMIC_INTERRUPTED. This is important for a pseudo * atomic section inside a WITHOUT-INTERRUPTS. + * + * Also, if in_leaving_without_gcing_race_p then + * interrupt_handle_pending is going to be called soon, so + * stashing the signal away is safe. */ - if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) { - store_signal_data_for_later(data,handler,signal,info,context); - SetSymbolValue(INTERRUPT_PENDING, T,thread); + if ((SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) || + in_leaving_without_gcing_race_p(thread)) { FSHOW_SIGNAL((stderr, - "/maybe_defer_handler(%x,%d),thread=%lu: deferred\n", + "/maybe_defer_handler(%x,%d): deferred (RACE=%d)\n", (unsigned int)handler,signal, - (unsigned long)thread->os_thread)); + in_leaving_without_gcing_race_p(thread))); + store_signal_data_for_later(data,handler,signal,info,context); + SetSymbolValue(INTERRUPT_PENDING, T,thread); + check_interrupt_context_or_lose(context); return 1; } - /* a slightly confusing test. arch_pseudo_atomic_atomic() doesn't + /* a slightly confusing test. arch_pseudo_atomic_atomic() doesn't * actually use its argument for anything on x86, so this branch * may succeed even when context is null (gencgc alloc()) */ - if ( -#ifdef FOREIGN_FUNCTION_CALL_FLAG - /* FIXME: this foreign_function_call_active test is dubious at - * best. If a foreign call is made in a pseudo atomic section - * (?) or more likely a pseudo atomic section is in a foreign - * call then an interrupt is executed immediately. Maybe it - * has to do with C code not maintaining pseudo atomic - * properly. MG - 2005-08-10 */ - (!foreign_function_call_active) && -#endif - arch_pseudo_atomic_atomic(context)) { + if (arch_pseudo_atomic_atomic(context)) { + FSHOW_SIGNAL((stderr, + "/maybe_defer_handler(%x,%d): deferred(PA)\n", + (unsigned int)handler,signal)); store_signal_data_for_later(data,handler,signal,info,context); arch_set_pseudo_atomic_interrupted(context); - FSHOW_SIGNAL((stderr, - "/maybe_defer_handler(%x,%d),thread=%lu: deferred(PA)\n", - (unsigned int)handler,signal, - (unsigned long)thread->os_thread)); + check_interrupt_context_or_lose(context); return 1; } FSHOW_SIGNAL((stderr, - "/maybe_defer_handler(%x,%d),thread=%lu: not deferred\n", - (unsigned int)handler,signal, - (unsigned long)thread->os_thread)); + "/maybe_defer_handler(%x,%d): not deferred\n", + (unsigned int)handler,signal)); return 0; } @@ -696,140 +1132,149 @@ store_signal_data_for_later (struct interrupt_data *data, void *handler, if(info) memcpy(&(data->pending_info), info, sizeof(siginfo_t)); - FSHOW_SIGNAL((stderr, "/store_signal_data_for_later: signal: %d\n", signal)); + FSHOW_SIGNAL((stderr, "/store_signal_data_for_later: signal: %d\n", + signal)); + + if(!context) + lose("Null context"); - if(context) { - /* the signal mask in the context (from before we were - * interrupted) is copied to be restored when - * run_deferred_handler happens. Then the usually-blocked - * signals are added to the mask in the context so that we are - * running with blocked signals when the handler returns */ - sigcopyset(&(data->pending_mask),os_context_sigmask_addr(context)); - sigaddset_deferrable(os_context_sigmask_addr(context)); - } + /* the signal mask in the context (from before we were + * interrupted) is copied to be restored when run_deferred_handler + * happens. Then the usually-blocked signals are added to the mask + * in the context so that we are running with blocked signals when + * the handler returns */ + sigcopyset(&(data->pending_mask),os_context_sigmask_addr(context)); + sigaddset_deferrable(os_context_sigmask_addr(context)); } static void maybe_now_maybe_later(int signal, siginfo_t *info, void *void_context) { - os_context_t *context = arch_os_get_context(&void_context); + SAVE_ERRNO(signal,context,void_context); struct thread *thread = arch_os_get_current_thread(); struct interrupt_data *data = thread->interrupt_data; - -#if defined(LISP_FEATURE_LINUX) || defined(RESTORE_FP_CONTROL_FROM_CONTEXT) - os_restore_fp_control(context); -#endif - if(!maybe_defer_handler(interrupt_handle_now,data,signal,info,context)) interrupt_handle_now(signal, info, context); - -#ifdef LISP_FEATURE_DARWIN - DARWIN_FIX_CONTEXT(context); -#endif + RESTORE_ERRNO; } static void -low_level_interrupt_handle_now(int signal, siginfo_t *info, os_context_t *context) +low_level_interrupt_handle_now(int signal, siginfo_t *info, + os_context_t *context) { /* No FP control fixage needed, caller has done that. */ - check_blockables_blocked_or_lose(); + check_blockables_blocked_or_lose(0); check_interrupts_enabled_or_lose(context); - interrupt_low_level_handlers[signal](signal, info, context); + (*interrupt_low_level_handlers[signal])(signal, info, context); /* No Darwin context fixage needed, caller does that. */ } static void low_level_maybe_now_maybe_later(int signal, siginfo_t *info, void *void_context) { - os_context_t *context = arch_os_get_context(&void_context); + SAVE_ERRNO(signal,context,void_context); struct thread *thread = arch_os_get_current_thread(); struct interrupt_data *data = thread->interrupt_data; -#if defined(LISP_FEATURE_LINUX) || defined(RESTORE_FP_CONTROL_FROM_CONTEXT) - os_restore_fp_control(context); -#endif - if(!maybe_defer_handler(low_level_interrupt_handle_now,data, signal,info,context)) low_level_interrupt_handle_now(signal, info, context); - -#ifdef LISP_FEATURE_DARWIN - DARWIN_FIX_CONTEXT(context); -#endif + RESTORE_ERRNO; } #endif #ifdef LISP_FEATURE_SB_THREAD +/* This function must not cons, because that may trigger a GC. */ void -sig_stop_for_gc_handler(int signal, siginfo_t *info, void *void_context) +sig_stop_for_gc_handler(int signal, siginfo_t *info, os_context_t *context) { - os_context_t *context = arch_os_get_context(&void_context); - struct thread *thread=arch_os_get_current_thread(); - sigset_t ss; - if ((arch_pseudo_atomic_atomic(context) || - SymbolValue(GC_INHIBIT,thread) != NIL)) { + /* Test for GC_INHIBIT _first_, else we'd trap on every single + * pseudo atomic until gc is finally allowed. */ + if (SymbolValue(GC_INHIBIT,thread) != NIL) { + FSHOW_SIGNAL((stderr, "sig_stop_for_gc deferred (*GC-INHIBIT*)\n")); SetSymbolValue(STOP_FOR_GC_PENDING,T,thread); - if (SymbolValue(GC_INHIBIT,thread) == NIL) - arch_set_pseudo_atomic_interrupted(context); - FSHOW_SIGNAL((stderr,"thread=%lu sig_stop_for_gc deferred\n", - thread->os_thread)); - } else { - /* need the context stored so it can have registers scavenged */ - fake_foreign_function_call(context); + return; + } else if (arch_pseudo_atomic_atomic(context)) { + FSHOW_SIGNAL((stderr,"sig_stop_for_gc deferred (PA)\n")); + SetSymbolValue(STOP_FOR_GC_PENDING,T,thread); + arch_set_pseudo_atomic_interrupted(context); + maybe_save_gc_mask_and_block_deferrables + (os_context_sigmask_addr(context)); + return; + } - sigfillset(&ss); /* Block everything. */ - thread_sigmask(SIG_BLOCK,&ss,0); + FSHOW_SIGNAL((stderr, "/sig_stop_for_gc_handler\n")); - if(thread->state!=STATE_RUNNING) { - lose("sig_stop_for_gc_handler: wrong thread state: %ld\n", - fixnum_value(thread->state)); - } - thread->state=STATE_SUSPENDED; - FSHOW_SIGNAL((stderr,"thread=%lu suspended\n",thread->os_thread)); + /* Not PA and GC not inhibited -- we can stop now. */ -#if defined(SIG_RESUME_FROM_GC) - sigemptyset(&ss); sigaddset(&ss,SIG_RESUME_FROM_GC); -#else - sigemptyset(&ss); sigaddset(&ss,SIG_STOP_FOR_GC); -#endif + /* need the context stored so it can have registers scavenged */ + fake_foreign_function_call(context); - /* It is possible to get SIGCONT (and probably other - * non-blockable signals) here. */ -#ifdef SIG_RESUME_FROM_GC - { - int sigret; - do { sigwait(&ss, &sigret); } - while (sigret != SIG_RESUME_FROM_GC); - } -#else - while (sigwaitinfo(&ss,0) != SIG_STOP_FOR_GC); -#endif + /* Not pending anymore. */ + SetSymbolValue(GC_PENDING,NIL,thread); + SetSymbolValue(STOP_FOR_GC_PENDING,NIL,thread); + + /* Consider this: in a PA section GC is requested: GC_PENDING, + * pseudo_atomic_interrupted and gc_blocked_deferrables are set, + * deferrables are blocked then pseudo_atomic_atomic is cleared, + * but a SIG_STOP_FOR_GC arrives before trapping to + * interrupt_handle_pending. Here, GC_PENDING is cleared but + * pseudo_atomic_interrupted is not and we go on running with + * pseudo_atomic_interrupted but without a pending interrupt or + * GC. GC_BLOCKED_DEFERRABLES is also left at 1. So let's tidy it + * up. */ + if (thread->interrupt_data->gc_blocked_deferrables) { + FSHOW_SIGNAL((stderr,"cleaning up after gc_blocked_deferrables\n")); + clear_pseudo_atomic_interrupted(thread); + sigcopyset(os_context_sigmask_addr(context), + &thread->interrupt_data->pending_mask); + thread->interrupt_data->gc_blocked_deferrables = 0; + } - FSHOW_SIGNAL((stderr,"thread=%lu resumed\n",thread->os_thread)); - if(thread->state!=STATE_RUNNING) { - lose("sig_stop_for_gc_handler: wrong thread state on wakeup: %ld\n", - fixnum_value(thread->state)); - } + if(thread_state(thread)!=STATE_RUNNING) { + lose("sig_stop_for_gc_handler: wrong thread state: %ld\n", + fixnum_value(thread->state)); + } + + set_thread_state(thread,STATE_SUSPENDED); + FSHOW_SIGNAL((stderr,"suspended\n")); - undo_fake_foreign_function_call(context); + /* While waiting for gc to finish occupy ourselves with zeroing + * the unused portion of the control stack to reduce conservatism. + * On hypothetic platforms with threads and exact gc it is + * actually a must. */ + scrub_control_stack(); + + wait_for_thread_state_change(thread, STATE_SUSPENDED); + FSHOW_SIGNAL((stderr,"resumed\n")); + + if(thread_state(thread)!=STATE_RUNNING) { + lose("sig_stop_for_gc_handler: wrong thread state on wakeup: %ld\n", + fixnum_value(thread_state(thread))); } + + undo_fake_foreign_function_call(context); } + #endif void interrupt_handle_now_handler(int signal, siginfo_t *info, void *void_context) { - os_context_t *context = arch_os_get_context(&void_context); -#if defined(LISP_FEATURE_LINUX) || defined(RESTORE_FP_CONTROL_FROM_CONTEXT) - os_restore_fp_control(context); + SAVE_ERRNO(signal,context,void_context); +#ifndef LISP_FEATURE_WIN32 + if ((signal == SIGILL) || (signal == SIGBUS) +#ifndef LISP_FEATURE_LINUX + || (signal == SIGEMT) #endif - interrupt_handle_now(signal, info, context); -#ifdef LISP_FEATURE_DARWIN - DARWIN_FIX_CONTEXT(context); + ) + corruption_warning_and_maybe_lose("Signal %d recieved", signal); #endif + interrupt_handle_now(signal, info, context); + RESTORE_ERRNO; } /* manipulate the signal context and stack such that when the handler @@ -847,6 +1292,10 @@ extern void call_into_lisp_tramp(void); void arrange_return_to_lisp_function(os_context_t *context, lispobj function) { +#ifndef LISP_FEATURE_WIN32 + check_gc_signals_unblocked_or_lose + (os_context_sigmask_addr(context)); +#endif #if !(defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)) void * fun=native_pointer(function); void *code = &(((struct simple_fun *) fun)->code); @@ -855,8 +1304,14 @@ arrange_return_to_lisp_function(os_context_t *context, lispobj function) /* Build a stack frame showing `interrupted' so that the * user's backtrace makes (as much) sense (as usual) */ - /* FIXME: what about restoring fp state? */ - /* FIXME: what about restoring errno? */ + /* fp state is saved and restored by call_into_lisp */ + /* FIXME: errno is not restored, but since current uses of this + * function only call Lisp code that signals an error, it's not + * much of a problem. In other words, running out of the control + * stack between a syscall and (GET-ERRNO) may clobber errno if + * something fails during signalling or in the handler. But I + * can't see what can go wrong as long as there is no CONTINUE + * like restart on them. */ #ifdef LISP_FEATURE_X86 /* Suppose the existence of some function that saved all * registers, called call_into_lisp, then restored GP registers and @@ -914,8 +1369,10 @@ arrange_return_to_lisp_function(os_context_t *context, lispobj function) *(register_save_area + 7) = *os_context_register_addr(context,reg_EAX); *(register_save_area + 8) = *context_eflags_addr(context); - *os_context_pc_addr(context) = call_into_lisp_tramp; - *os_context_register_addr(context,reg_ECX) = register_save_area; + *os_context_pc_addr(context) = + (os_context_register_t) call_into_lisp_tramp; + *os_context_register_addr(context,reg_ECX) = + (os_context_register_t) register_save_area; #else /* return address for call_into_lisp: */ @@ -1012,24 +1469,10 @@ arrange_return_to_lisp_function(os_context_t *context, lispobj function) *os_context_register_addr(context,reg_CODE) = (os_context_register_t)(fun + FUN_POINTER_LOWTAG); #endif + FSHOW((stderr, "/arranged return to Lisp function (0x%lx)\n", + (long)function)); } -#ifdef LISP_FEATURE_SB_THREAD - -/* FIXME: this function can go away when all lisp handlers are invoked - * via arrange_return_to_lisp_function. */ -void -interrupt_thread_handler(int num, siginfo_t *info, void *v_context) -{ - os_context_t *context = (os_context_t*)arch_os_get_context(&v_context); - - /* let the handler enable interrupts again when it sees fit */ - sigaddset_deferrable(os_context_sigmask_addr(context)); - arrange_return_to_lisp_function(context, SymbolFunction(RUN_INTERRUPTION)); -} - -#endif - /* KLUDGE: Theoretically the approach we use for undefined alien * variables should work for functions as well, but on PPC/Darwin * we get bus error at bogus addresses instead, hence this workaround, @@ -1037,8 +1480,22 @@ interrupt_thread_handler(int num, siginfo_t *info, void *v_context) * functions and variables. */ void -undefined_alien_function() { - funcall0(SymbolFunction(UNDEFINED_ALIEN_FUNCTION_ERROR)); +undefined_alien_function(void) +{ + funcall0(StaticSymbolFunction(UNDEFINED_ALIEN_FUNCTION_ERROR)); +} + +/* Called from the REPL, too. */ +void reset_control_stack_guard_page(void) +{ + struct thread *th=arch_os_get_current_thread(); + if (th->control_stack_guard_page_protected == NIL) { + memset(CONTROL_STACK_GUARD_PAGE(th), 0, os_vm_page_size); + protect_control_stack_guard_page(1, NULL); + protect_control_stack_return_guard_page(0, NULL); + th->control_stack_guard_page_protected = T; + fprintf(stderr, "INFO: Control stack guard page reprotected\n"); + } } boolean @@ -1046,19 +1503,30 @@ handle_guard_page_triggered(os_context_t *context,os_vm_address_t addr) { struct thread *th=arch_os_get_current_thread(); - /* note the os_context hackery here. When the signal handler returns, - * it won't go back to what it was doing ... */ - if(addr >= CONTROL_STACK_GUARD_PAGE(th) && - addr < CONTROL_STACK_GUARD_PAGE(th) + os_vm_page_size) { + if(addr >= CONTROL_STACK_HARD_GUARD_PAGE(th) && + addr < CONTROL_STACK_HARD_GUARD_PAGE(th) + os_vm_page_size) { + lose("Control stack exhausted"); + } + else if(addr >= CONTROL_STACK_GUARD_PAGE(th) && + addr < CONTROL_STACK_GUARD_PAGE(th) + os_vm_page_size) { /* We hit the end of the control stack: disable guard page * protection so the error handler has some headroom, protect the * previous page so that we can catch returns from the guard page * and restore it. */ - protect_control_stack_guard_page(0); - protect_control_stack_return_guard_page(1); + if (th->control_stack_guard_page_protected == NIL) + lose("control_stack_guard_page_protected NIL"); + protect_control_stack_guard_page(0, NULL); + protect_control_stack_return_guard_page(1, NULL); + th->control_stack_guard_page_protected = NIL; + fprintf(stderr, "INFO: Control stack guard page unprotected\n"); +#ifdef LISP_FEATURE_C_STACK_IS_CONTROL_STACK + /* For the unfortunate case, when the control stack is + * exhausted in a signal handler. */ + unblock_signals_in_context_and_maybe_warn(context); +#endif arrange_return_to_lisp_function - (context, SymbolFunction(CONTROL_STACK_EXHAUSTED_ERROR)); + (context, StaticSymbolFunction(CONTROL_STACK_EXHAUSTED_ERROR)); return 1; } else if(addr >= CONTROL_STACK_RETURN_GUARD_PAGE(th) && @@ -1067,14 +1535,63 @@ handle_guard_page_triggered(os_context_t *context,os_vm_address_t addr) * unprotect this one. This works even if we somehow missed * the return-guard-page, and hit it on our way to new * exhaustion instead. */ - protect_control_stack_guard_page(1); - protect_control_stack_return_guard_page(0); + if (th->control_stack_guard_page_protected != NIL) + lose("control_stack_guard_page_protected not NIL"); + reset_control_stack_guard_page(); + return 1; + } + else if(addr >= BINDING_STACK_HARD_GUARD_PAGE(th) && + addr < BINDING_STACK_HARD_GUARD_PAGE(th) + os_vm_page_size) { + lose("Binding stack exhausted"); + } + else if(addr >= BINDING_STACK_GUARD_PAGE(th) && + addr < BINDING_STACK_GUARD_PAGE(th) + os_vm_page_size) { + protect_binding_stack_guard_page(0, NULL); + protect_binding_stack_return_guard_page(1, NULL); + fprintf(stderr, "INFO: Binding stack guard page unprotected\n"); + + /* For the unfortunate case, when the binding stack is + * exhausted in a signal handler. */ + unblock_signals_in_context_and_maybe_warn(context); + arrange_return_to_lisp_function + (context, StaticSymbolFunction(BINDING_STACK_EXHAUSTED_ERROR)); + return 1; + } + else if(addr >= BINDING_STACK_RETURN_GUARD_PAGE(th) && + addr < BINDING_STACK_RETURN_GUARD_PAGE(th) + os_vm_page_size) { + protect_binding_stack_guard_page(1, NULL); + protect_binding_stack_return_guard_page(0, NULL); + fprintf(stderr, "INFO: Binding stack guard page reprotected\n"); + return 1; + } + else if(addr >= ALIEN_STACK_HARD_GUARD_PAGE(th) && + addr < ALIEN_STACK_HARD_GUARD_PAGE(th) + os_vm_page_size) { + lose("Alien stack exhausted"); + } + else if(addr >= ALIEN_STACK_GUARD_PAGE(th) && + addr < ALIEN_STACK_GUARD_PAGE(th) + os_vm_page_size) { + protect_alien_stack_guard_page(0, NULL); + protect_alien_stack_return_guard_page(1, NULL); + fprintf(stderr, "INFO: Alien stack guard page unprotected\n"); + + /* For the unfortunate case, when the alien stack is + * exhausted in a signal handler. */ + unblock_signals_in_context_and_maybe_warn(context); + arrange_return_to_lisp_function + (context, StaticSymbolFunction(ALIEN_STACK_EXHAUSTED_ERROR)); + return 1; + } + else if(addr >= ALIEN_STACK_RETURN_GUARD_PAGE(th) && + addr < ALIEN_STACK_RETURN_GUARD_PAGE(th) + os_vm_page_size) { + protect_alien_stack_guard_page(1, NULL); + protect_alien_stack_return_guard_page(0, NULL); + fprintf(stderr, "INFO: Alien stack guard page reprotected\n"); return 1; } else if (addr >= undefined_alien_address && addr < undefined_alien_address + os_vm_page_size) { arrange_return_to_lisp_function - (context, SymbolFunction(UNDEFINED_ALIEN_VARIABLE_ERROR)); + (context, StaticSymbolFunction(UNDEFINED_ALIEN_VARIABLE_ERROR)); return 1; } else return 0; @@ -1106,10 +1623,9 @@ static volatile int sigaction_nodefer_works = -1; static void sigaction_nodefer_test_handler(int signal, siginfo_t *info, void *void_context) { - sigset_t empty, current; + sigset_t current; int i; - sigemptyset(&empty); - thread_sigmask(SIG_BLOCK, &empty, ¤t); + get_current_sigmask(¤t); /* There should be exactly two blocked signals: the two we added * to sa_mask when setting up the handler. NetBSD doesn't block * the signal we're handling when SA_NODEFER is set; Linux before @@ -1126,7 +1642,7 @@ sigaction_nodefer_test_handler(int signal, siginfo_t *info, void *void_context) } static void -see_if_sigaction_nodefer_works() +see_if_sigaction_nodefer_works(void) { struct sigaction sa, old_sa; @@ -1153,30 +1669,40 @@ see_if_sigaction_nodefer_works() static void unblock_me_trampoline(int signal, siginfo_t *info, void *void_context) { + SAVE_ERRNO(signal,context,void_context); sigset_t unblock; sigemptyset(&unblock); sigaddset(&unblock, signal); thread_sigmask(SIG_UNBLOCK, &unblock, 0); - interrupt_handle_now_handler(signal, info, void_context); + interrupt_handle_now(signal, info, context); + RESTORE_ERRNO; } static void low_level_unblock_me_trampoline(int signal, siginfo_t *info, void *void_context) { + SAVE_ERRNO(signal,context,void_context); sigset_t unblock; sigemptyset(&unblock); sigaddset(&unblock, signal); thread_sigmask(SIG_UNBLOCK, &unblock, 0); - (*interrupt_low_level_handlers[signal])(signal, info, void_context); + (*interrupt_low_level_handlers[signal])(signal, info, context); + RESTORE_ERRNO; +} + +static void +low_level_handle_now_handler(int signal, siginfo_t *info, void *void_context) +{ + SAVE_ERRNO(signal,context,void_context); + (*interrupt_low_level_handlers[signal])(signal, info, context); + RESTORE_ERRNO; } void undoably_install_low_level_interrupt_handler (int signal, - void handler(int, - siginfo_t*, - void*)) + interrupt_handler_t handler) { struct sigaction sa; @@ -1185,33 +1711,20 @@ undoably_install_low_level_interrupt_handler (int signal, } if (ARE_SAME_HANDLER(handler, SIG_DFL)) - sa.sa_sigaction = handler; + sa.sa_sigaction = (void (*)(int, siginfo_t*, void*))handler; else if (sigismember(&deferrable_sigset,signal)) sa.sa_sigaction = low_level_maybe_now_maybe_later; - /* The use of a trampoline appears to break the - arch_os_get_context() workaround for SPARC/Linux. For now, - don't use the trampoline (and so be vulnerable to the problems - that SA_NODEFER is meant to solve. */ -#if !(defined(LISP_FEATURE_SPARC) && defined(LISP_FEATURE_LINUX)) else if (!sigaction_nodefer_works && !sigismember(&blockable_sigset, signal)) sa.sa_sigaction = low_level_unblock_me_trampoline; -#endif else - sa.sa_sigaction = handler; + sa.sa_sigaction = low_level_handle_now_handler; sigcopyset(&sa.sa_mask, &blockable_sigset); sa.sa_flags = SA_SIGINFO | SA_RESTART | (sigaction_nodefer_works ? SA_NODEFER : 0); #ifdef LISP_FEATURE_C_STACK_IS_CONTROL_STACK - if((signal==SIG_MEMORY_FAULT) -#ifdef SIG_MEMORY_FAULT2 - || (signal==SIG_MEMORY_FAULT2) -#endif -#ifdef SIG_INTERRUPT_THREAD - || (signal==SIG_INTERRUPT_THREAD) -#endif - ) + if((signal==SIG_MEMORY_FAULT)) sa.sa_flags |= SA_ONSTACK; #endif @@ -1223,25 +1736,23 @@ undoably_install_low_level_interrupt_handler (int signal, /* This is called from Lisp. */ unsigned long -install_handler(int signal, void handler(int, siginfo_t*, void*)) +install_handler(int signal, void handler(int, siginfo_t*, os_context_t*)) { #ifndef LISP_FEATURE_WIN32 struct sigaction sa; - sigset_t old, new; + sigset_t old; union interrupt_handler oldhandler; FSHOW((stderr, "/entering POSIX install_handler(%d, ..)\n", signal)); - sigemptyset(&new); - sigaddset(&new, signal); - thread_sigmask(SIG_BLOCK, &new, &old); + block_blockable_signals(0, &old); FSHOW((stderr, "/interrupt_low_level_handlers[signal]=%x\n", (unsigned int)interrupt_low_level_handlers[signal])); if (interrupt_low_level_handlers[signal]==0) { if (ARE_SAME_HANDLER(handler, SIG_DFL) || ARE_SAME_HANDLER(handler, SIG_IGN)) - sa.sa_sigaction = handler; + sa.sa_sigaction = (void (*)(int, siginfo_t*, void*))handler; else if (sigismember(&deferrable_sigset, signal)) sa.sa_sigaction = maybe_now_maybe_later; else if (!sigaction_nodefer_works && @@ -1270,8 +1781,16 @@ install_handler(int signal, void handler(int, siginfo_t*, void*)) #endif } +/* This must not go through lisp as it's allowed anytime, even when on + * the altstack. */ +void +sigabrt_handler(int signal, siginfo_t *info, os_context_t *context) +{ + lose("SIGABRT received.\n"); +} + void -interrupt_init() +interrupt_init(void) { #ifndef LISP_FEATURE_WIN32 int i; @@ -1279,8 +1798,10 @@ interrupt_init() see_if_sigaction_nodefer_works(); sigemptyset(&deferrable_sigset); sigemptyset(&blockable_sigset); + sigemptyset(&gc_sigset); sigaddset_deferrable(&deferrable_sigset); sigaddset_blockable(&blockable_sigset); + sigaddset_gc(&gc_sigset); /* Set up high level handler information. */ for (i = 0; i < NSIG; i++) { @@ -1290,9 +1811,9 @@ interrupt_init() * signal(..)-style one-argument handlers, which is OK * because it works to call the 1-argument form where the * 3-argument form is expected.) */ - (void (*)(int, siginfo_t*, void*))SIG_DFL; + (void (*)(int, siginfo_t*, os_context_t*))SIG_DFL; } - + undoably_install_low_level_interrupt_handler(SIGABRT, sigabrt_handler); SHOW("returning from interrupt_init()"); #endif } @@ -1309,30 +1830,47 @@ void lisp_memory_fault_error(os_context_t *context, os_vm_address_t addr) { /* FIXME: This is lossy: if we get another memory fault (eg. from - * another thread) before lisp has read this, we the information. + * another thread) before lisp has read this, we lose the information. * However, since this is mostly informative, we'll live with that for * now -- some address is better then no address in this case. */ current_memory_fault_address = addr; - arrange_return_to_lisp_function(context, SymbolFunction(MEMORY_FAULT_ERROR)); + /* To allow debugging memory faults in signal handlers and such. */ + corruption_warning_and_maybe_lose("Memory fault at %x (pc=%p, sp=%p)", + addr, + *os_context_pc_addr(context), +#ifdef ARCH_HAS_STACK_POINTER + *os_context_sp_addr(context) +#else + 0 +#endif + ); + unblock_signals_in_context_and_maybe_warn(context); +#ifdef LISP_FEATURE_C_STACK_IS_CONTROL_STACK + arrange_return_to_lisp_function(context, + StaticSymbolFunction(MEMORY_FAULT_ERROR)); +#else + funcall0(StaticSymbolFunction(MEMORY_FAULT_ERROR)); +#endif } #endif -void +static void unhandled_trap_error(os_context_t *context) { lispobj context_sap; fake_foreign_function_call(context); + unblock_gc_signals(0, 0); context_sap = alloc_sap(context); #ifndef LISP_FEATURE_WIN32 thread_sigmask(SIG_SETMASK, os_context_sigmask_addr(context), 0); #endif - funcall1(SymbolFunction(UNHANDLED_TRAP_ERROR), context_sap); + funcall1(StaticSymbolFunction(UNHANDLED_TRAP_ERROR), context_sap); lose("UNHANDLED-TRAP-ERROR fell through"); } -/* Common logic far trapping instructions. How we actually handle each - * case is highly architecture dependant, but the overall shape is +/* Common logic for trapping instructions. How we actually handle each + * case is highly architecture dependent, but the overall shape is * this. */ void handle_trap(os_context_t *context, int trap) @@ -1372,4 +1910,3 @@ handle_trap(os_context_t *context, int trap) unhandled_trap_error(context); } } -