X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Finterrupt.c;h=b6c5c0dbf4695ed7052710500f4c1d779c6cf13c;hb=1b32a547d26cf078ba9f2948edeb27ff91e78f49;hp=4497428503cd8e9dd777cdb93990cfeaf90dadbf;hpb=0b5119848b6b8713e473fa669356645747e11dbd;p=sbcl.git diff --git a/src/runtime/interrupt.c b/src/runtime/interrupt.c index 4497428..b6c5c0d 100644 --- a/src/runtime/interrupt.c +++ b/src/runtime/interrupt.c @@ -40,16 +40,18 @@ * * - WHN 20000728, dan 20010128 */ +#include "sbcl.h" #include #include #include #include #include +#ifndef LISP_FEATURE_WIN32 #include +#endif #include -#include "sbcl.h" #include "runtime.h" #include "arch.h" #include "os.h" @@ -57,28 +59,79 @@ #include "globals.h" #include "lispregs.h" #include "validate.h" -#include "monitor.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" +/* 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 +/* 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(context,void_context) \ + { \ + int _saved_errno = errno; \ + RESTORE_FP_CONTROL_WORD(context,void_context); \ + { + +#define RESTORE_ERRNO \ + } \ + errno = _saved_errno; \ + } -void run_deferred_handler(struct interrupt_data *data, void *v_context) ; +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); -boolean interrupt_maybe_gc_int(int signal, siginfo_t *info, void *v_context); -void sigaddset_deferrable(sigset_t *s) +static void +fill_current_sigmask(sigset_t *sigset) +{ + /* Get the current sigmask, by blocking the empty set. */ + sigset_t empty; + sigemptyset(&empty); + thread_sigmask(SIG_BLOCK, &empty, sigset); +} + +void +sigaddset_deferrable(sigset_t *s) { sigaddset(s, SIGHUP); sigaddset(s, SIGINT); + sigaddset(s, SIGTERM); sigaddset(s, SIGQUIT); sigaddset(s, SIGPIPE); sigaddset(s, SIGALRM); @@ -86,54 +139,282 @@ void 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); - sigaddset(s, SIGUSR1); - sigaddset(s, SIGUSR2); +} + +void +sigdelset_deferrable(sigset_t *s) +{ + sigdelset(s, SIGHUP); + sigdelset(s, SIGINT); + sigdelset(s, SIGQUIT); + sigdelset(s, SIGPIPE); + sigdelset(s, SIGALRM); + sigdelset(s, SIGURG); + sigdelset(s, SIGTSTP); + sigdelset(s, SIGCHLD); + sigdelset(s, SIGIO); +#ifndef LISP_FEATURE_HPUX + sigdelset(s, SIGXCPU); + sigdelset(s, SIGXFSZ); +#endif + sigdelset(s, SIGVTALRM); + sigdelset(s, SIGPROF); + sigdelset(s, SIGWINCH); +} + +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 } -void sigaddset_blockable(sigset_t *s) +void +sigdelset_gc(sigset_t *sigset) { - sigaddset_deferrable(s); #ifdef LISP_FEATURE_SB_THREAD - sigaddset(s, SIG_STOP_FOR_GC); + sigdelset(sigset,SIG_STOP_FOR_GC); #endif } /* initialized in interrupt_init */ -static sigset_t deferrable_sigset; -static sigset_t blockable_sigset; +sigset_t deferrable_sigset; +sigset_t blockable_sigset; +sigset_t gc_sigset; +#endif -inline static void check_blockables_blocked_or_lose() +boolean +deferrables_blocked_in_sigset_p(sigset_t *sigset) { - /* Get the current sigmask, by blocking the empty set. */ - sigset_t empty,current; +#if !defined(LISP_FEATURE_WIN32) int i; - sigemptyset(&empty); - thread_sigmask(SIG_BLOCK, &empty, ¤t); - for(i=0;iinterrupt_data; + sigset_t oldset; + /* Obviously, this function is called when signals may not be + * blocked. Let's make sure we are not interrupted. */ + thread_sigmask(SIG_BLOCK, &blockable_sigset, &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); + unblock_gc_signals(); + 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"); + } +#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_in_sigset_or_lose(sigset); + if (interrupt_pending || interrupt_deferred_p) + check_deferrables_blocked_in_sigset_or_lose(sigset); + else { + check_deferrables_unblocked_in_sigset_or_lose(sigset); + /* If deferrables are unblocked then we are open to signals + * that run lisp code. */ + check_gc_signals_unblocked_in_sigset_or_lose(sigset); + } #endif - arch_pseudo_atomic_atomic(context)) - lose ("in pseudo atomic section"); } /* When we catch an internal error, should we pass it back to Lisp to @@ -143,25 +424,80 @@ inline static void check_interrupts_enabled_or_lose(os_context_t *context) * becomes 'yes'.) */ boolean internal_errors_enabled = 0; -static void (*interrupt_low_level_handlers[NSIG]) (int, siginfo_t*, void*); +#ifndef LISP_FEATURE_WIN32 +static +void (*interrupt_low_level_handlers[NSIG]) (int, siginfo_t*, os_context_t*); +#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 +block_blockable_signals(void) +{ +#ifndef LISP_FEATURE_WIN32 + thread_sigmask(SIG_BLOCK, &blockable_sigset, 0); +#endif +} -void reset_signal_mask(void) +void +block_deferrable_signals(void) { - sigset_t new; - sigemptyset(&new); - thread_sigmask(SIG_SETMASK,&new,0); +#ifndef LISP_FEATURE_WIN32 + thread_sigmask(SIG_BLOCK, &deferrable_sigset, 0); +#endif } -void block_blockable_signals(void) +void +unblock_deferrable_signals_in_sigset(sigset_t *sigset) { - sigset_t block; - sigcopyset(&block, &blockable_sigset); - thread_sigmask(SIG_BLOCK, &block, 0); +#ifndef LISP_FEATURE_WIN32 + if (interrupt_handler_pending_p()) + lose("unblock_deferrable_signals_in_sigset: losing proposition\n"); + check_gc_signals_unblocked_in_sigset_or_lose(sigset); + sigdelset_deferrable(sigset); +#endif +} + +void +unblock_deferrable_signals(void) +{ +#ifndef LISP_FEATURE_WIN32 + if (interrupt_handler_pending_p()) + lose("unblock_deferrable_signals: losing proposition\n"); + check_gc_signals_unblocked_or_lose(); + thread_sigmask(SIG_UNBLOCK, &deferrable_sigset, 0); +#endif +} + +void +unblock_gc_signals(void) +{ +#if defined(LISP_FEATURE_SB_THREAD) && !defined(LISP_FEATURE_WIN32) + thread_sigmask(SIG_UNBLOCK,&gc_sigset,0); +#endif +} + +void +unblock_signals_in_context_and_maybe_warn(os_context_t *context) +{ +#ifndef LISP_FEATURE_WIN32 + int i, oops=0; + sigset_t *sigset=os_context_sigmask_addr(context); + for(i = 1; i < NSIG; i++) { + if (sigismember(&gc_sigset, i) && sigismember(sigset, i)) { + if (!oops) { + fprintf(stderr, +"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"); + } + oops=1; + } + } + sigdelset_gc(sigset); + if (!interrupt_handler_pending_p()) { + unblock_deferrable_signals_in_sigset(sigset); + } +#endif } @@ -222,6 +558,8 @@ build_fake_control_stack_frames(struct thread *th,os_context_t *context) #endif } +/* Stores the context for gc to scavange and builds fake stack + * frames. */ void fake_foreign_function_call(os_context_t *context) { @@ -236,9 +574,17 @@ fake_foreign_function_call(os_context_t *context) dynamic_space_free_pointer = (lispobj *)(unsigned long) (*os_context_register_addr(context, reg_ALLOC)); -#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", context); + 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 @@ -256,7 +602,7 @@ fake_foreign_function_call(os_context_t *context) fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,thread)); if (context_index >= MAX_INTERRUPTS) { - lose("maximum interrupt nesting depth (%d) exceeded", MAX_INTERRUPTS); + lose("maximum interrupt nesting depth (%d) exceeded\n", MAX_INTERRUPTS); } bind_variable(FREE_INTERRUPT_CONTEXT_INDEX, @@ -264,14 +610,14 @@ fake_foreign_function_call(os_context_t *context) thread->interrupt_contexts[context_index] = context; - /* no longer in Lisp now */ +#ifdef FOREIGN_FUNCTION_CALL_FLAG foreign_function_call_active = 1; +#endif } /* blocks all blockable signals. If you are calling from a signal handler, * the usual signal mask will be restored from the context when the handler * finishes. Otherwise, be careful */ - void undo_fake_foreign_function_call(os_context_t *context) { @@ -279,8 +625,9 @@ undo_fake_foreign_function_call(os_context_t *context) /* Block all blockable signals. */ block_blockable_signals(); - /* going back into Lisp */ +#ifdef FOREIGN_FUNCTION_CALL_FLAG foreign_function_call_active = 0; +#endif /* Undo dynamic binding of FREE_INTERRUPT_CONTEXT_INDEX */ unbind(thread); @@ -288,15 +635,21 @@ undo_fake_foreign_function_call(os_context_t *context) #ifdef reg_ALLOC /* Put the dynamic space free pointer back into the context. */ *os_context_register_addr(context, reg_ALLOC) = - (unsigned long) dynamic_space_free_pointer; + (unsigned long) dynamic_space_free_pointer + | (*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 } /* a handler for the signal caused by execution of a trap opcode * signalling an internal error */ void -interrupt_internal_error(int signal, siginfo_t *info, os_context_t *context, - boolean continuable) +interrupt_internal_error(os_context_t *context, boolean continuable) { lispobj context_sap; @@ -306,14 +659,27 @@ interrupt_internal_error(int signal, siginfo_t *info, os_context_t *context, describe_internal_error(context); /* There's no good way to recover from an internal error * before the Lisp error handling mechanism is set up. */ - lose("internal error too early in init, can't recover"); + lose("internal error too early in init, can't recover\n"); } /* Allocate the SAP object while the interrupts are still * disabled. */ + unblock_gc_signals(); 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 @@ -322,7 +688,7 @@ interrupt_internal_error(int signal, siginfo_t *info, os_context_t *context, * 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 */ @@ -330,114 +696,184 @@ interrupt_internal_error(int signal, siginfo_t *info, os_context_t *context, 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) { - struct thread *thread; - struct interrupt_data *data; + /* 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 + * the GC (in alloc()) decides that a GC is required, it will set + * *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; + + if (arch_pseudo_atomic_atomic(context)) { + lose("Handling pending interrupt in pseduo atomic."); + } + + FSHOW_SIGNAL((stderr, "/entering interrupt_handle_pending\n")); check_blockables_blocked_or_lose(); - thread=arch_os_get_current_thread(); - data=thread->interrupt_data; + /* 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. */ - interrupt_maybe_gc_int(0,NULL,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(); + + /* 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); } - /* we may be here only to do the gc stuff, if interrupts are - * enabled run the pending handler */ - if (!((SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) || - ( -#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) - (!foreign_function_call_active) && -#endif - arch_pseudo_atomic_atomic(context)))) { - - /* 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); - } +#ifndef LISP_FEATURE_WIN32 + /* 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); } + /* 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")); +#endif } -/* - * 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, void *void_context) +interrupt_handle_now(int signal, siginfo_t *info, os_context_t *context) { - os_context_t *context = (os_context_t*)void_context; -#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) +#ifdef FOREIGN_FUNCTION_CALL_FLAG boolean were_in_lisp; #endif union interrupt_handler handler; + check_blockables_blocked_or_lose(); + +#ifndef LISP_FEATURE_WIN32 if (sigismember(&deferrable_sigset,signal)) check_interrupts_enabled_or_lose(context); - -#ifdef LISP_FEATURE_LINUX - /* 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)) { return; } -#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) +#ifdef FOREIGN_FUNCTION_CALL_FLAG were_in_lisp = !foreign_function_call_active; if (were_in_lisp) #endif @@ -454,7 +890,7 @@ interrupt_handle_now(int signal, siginfo_t *info, void *void_context) /* This can happen if someone tries to ignore or default one * of the signals we need for runtime support, and the runtime * support decides to pass on it. */ - lose("no handler for signal %d in interrupt_handle_now(..)", signal); + lose("no handler for signal %d in interrupt_handle_now(..)\n", signal); } else if (lowtag_of(handler.lisp) == FUN_POINTER_LOWTAG) { /* Once we've decided what to do about contexts in a @@ -469,13 +905,15 @@ interrupt_handle_now(int signal, siginfo_t *info, void *void_context) * all our allocation from C now goes through a PA wrapper, * but still, doesn't hurt. * - * Yeah, but non-gencgc platforms that don't really wrap - * allocation in PA. MG - 2005-08-29 */ - - lispobj info_sap,context_sap = alloc_sap(context); + * Yeah, but non-gencgc platforms don't really wrap allocation + * in PA. MG - 2005-08-29 */ + + lispobj info_sap, context_sap; + /* Leave deferrable signals blocked, the handler itself will + * allow signals again when it sees fit. */ + unblock_gc_signals(); + context_sap = alloc_sap(context); info_sap = alloc_sap(info); - /* Allow signals again. */ - thread_sigmask(SIG_SETMASK, os_context_sigmask_addr(context), 0); FSHOW_SIGNAL((stderr,"/calling Lisp-level handler\n")); @@ -484,16 +922,18 @@ interrupt_handle_now(int signal, siginfo_t *info, void *void_context) info_sap, context_sap); } else { + /* This cannot happen in sane circumstances. */ FSHOW_SIGNAL((stderr,"/calling C-level handler\n")); +#ifndef LISP_FEATURE_WIN32 /* Allow signals again. */ thread_sigmask(SIG_SETMASK, os_context_sigmask_addr(context), 0); - - (*handler.c)(signal, info, void_context); +#endif + (*handler.c)(signal, info, context); } -#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) +#ifdef FOREIGN_FUNCTION_CALL_FLAG if (were_in_lisp) #endif { @@ -510,18 +950,22 @@ interrupt_handle_now(int signal, siginfo_t *info, void *void_context) * far as C or the kernel is concerned we dealt with the signal * already; we're just doing the Lisp-level processing now that we * put off then */ - -void -run_deferred_handler(struct interrupt_data *data, void *v_context) { +static void +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 boolean maybe_defer_handler(void *handler, struct interrupt_data *data, int signal, siginfo_t *info, os_context_t *context) @@ -531,46 +975,46 @@ maybe_defer_handler(void *handler, struct interrupt_data *data, check_blockables_blocked_or_lose(); if (SymbolValue(INTERRUPT_PENDING,thread) != NIL) - lose("interrupt already pending"); + 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) { + if ((SymbolValue(INTERRUPTS_ENABLED,thread) == NIL) || + in_leaving_without_gcing_race_p(thread)) { store_signal_data_for_later(data,handler,signal,info,context); SetSymbolValue(INTERRUPT_PENDING, T,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))); + 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 ( -#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64) - /* 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)) { 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)); + "/maybe_defer_handler(%x,%d): deferred(PA)\n", + (unsigned int)handler,signal)); + 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; } @@ -588,146 +1032,134 @@ store_signal_data_for_later (struct interrupt_data *data, void *handler, data->pending_signal = signal; if(info) memcpy(&(data->pending_info), info, sizeof(siginfo_t)); - 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)); - } + + FSHOW_SIGNAL((stderr, "/store_signal_data_for_later: signal: %d\n", + signal)); + + if(!context) + lose("Null 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); - struct thread *thread=arch_os_get_current_thread(); - struct interrupt_data *data=thread->interrupt_data; -#ifdef LISP_FEATURE_LINUX - os_restore_fp_control(context); -#endif - if(maybe_defer_handler(interrupt_handle_now,data,signal,info,context)) - return; - interrupt_handle_now(signal, info, context); -#ifdef LISP_FEATURE_DARWIN - /* Work around G5 bug */ - DARWIN_FIX_CONTEXT(context); -#endif + SAVE_ERRNO(context,void_context); + struct thread *thread = arch_os_get_current_thread(); + struct interrupt_data *data = thread->interrupt_data; + + if(!maybe_defer_handler(interrupt_handle_now,data,signal,info,context)) + interrupt_handle_now(signal, info, context); + RESTORE_ERRNO; } static void -low_level_interrupt_handle_now(int signal, siginfo_t *info, void *void_context) +low_level_interrupt_handle_now(int signal, siginfo_t *info, + os_context_t *context) { - os_context_t *context = (os_context_t*)void_context; - -#ifdef LISP_FEATURE_LINUX - os_restore_fp_control(context); -#endif + /* No FP control fixage needed, caller has done that. */ check_blockables_blocked_or_lose(); check_interrupts_enabled_or_lose(context); - interrupt_low_level_handlers[signal](signal, info, void_context); -#ifdef LISP_FEATURE_DARWIN - /* Work around G5 bug */ - DARWIN_FIX_CONTEXT(context); -#endif + (*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); - struct thread *thread=arch_os_get_current_thread(); - struct interrupt_data *data=thread->interrupt_data; -#ifdef LISP_FEATURE_LINUX - os_restore_fp_control(context); -#endif - if(maybe_defer_handler(low_level_interrupt_handle_now,data, - signal,info,context)) - return; - low_level_interrupt_handle_now(signal, info, context); -#ifdef LISP_FEATURE_DARWIN - /* Work around G5 bug */ - DARWIN_FIX_CONTEXT(context); -#endif + SAVE_ERRNO(context,void_context); + struct thread *thread = arch_os_get_current_thread(); + struct interrupt_data *data = thread->interrupt_data; + + if(!maybe_defer_handler(low_level_interrupt_handle_now,data, + signal,info,context)) + low_level_interrupt_handle_now(signal, info, context); + 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) { 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); + FSHOW_SIGNAL((stderr, "sig_stop_for_gc deferred (*GC-INHIBIT*)\n")); + return; + } else if (arch_pseudo_atomic_atomic(context)) { + 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)); + FSHOW_SIGNAL((stderr,"sig_stop_for_gc deferred (PA)\n")); + return; + } - sigfillset(&ss); /* Block everything. */ - thread_sigmask(SIG_BLOCK,&ss,0); + FSHOW_SIGNAL((stderr, "/sig_stop_for_gc_handler\n")); - /* The GC can't tell if a thread is a zombie, so this would be a - * good time to let the kernel reap any of our children in that - * awful state, to stop them from being waited for indefinitely. - * Userland reaping is done later when GC is finished */ - 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)); - - sigemptyset(&ss); sigaddset(&ss,SIG_STOP_FOR_GC); - sigwaitinfo(&ss,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)); - } + /* Not PA and GC not inhibited -- we can stop now. */ - undo_fake_foreign_function_call(context); + /* need the context stored so it can have registers scavenged */ + fake_foreign_function_call(context); + + /* Block everything. */ + sigfillset(&ss); + thread_sigmask(SIG_BLOCK,&ss,0); + + /* Not pending anymore. */ + SetSymbolValue(GC_PENDING,NIL,thread); + SetSymbolValue(STOP_FOR_GC_PENDING,NIL,thread); + + 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")); + + 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); - interrupt_handle_now(signal, info, context); -#ifdef LISP_FEATURE_DARWIN - DARWIN_FIX_CONTEXT(context); + SAVE_ERRNO(context,void_context); +#ifndef LISP_FEATURE_WIN32 + if ((signal == SIGILL) || (signal == SIGBUS) +#ifndef LISP_FEATURE_LINUX + || (signal == SIGEMT) #endif -} - -/* - * stuff to detect and handle hitting the GC trigger - */ - -#ifndef LISP_FEATURE_GENCGC -/* since GENCGC has its own way to record trigger */ -static boolean -gc_trigger_hit(int signal, siginfo_t *info, os_context_t *context) -{ - if (current_auto_gc_trigger == NULL) - return 0; - else{ - void *badaddr=arch_get_bad_addr(signal,info,context); - return (badaddr >= (void *)current_auto_gc_trigger && - badaddr <((void *)current_dynamic_space + DYNAMIC_SPACE_SIZE)); - } -} + ) + 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 * returns, it will call function instead of whatever it was doing @@ -735,13 +1167,19 @@ gc_trigger_hit(int signal, siginfo_t *info, os_context_t *context) */ #if (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)) -int *context_eflags_addr(os_context_t *context); +extern int *context_eflags_addr(os_context_t *context); #endif extern lispobj call_into_lisp(lispobj fun, lispobj *args, int nargs); extern void post_signal_tramp(void); -void arrange_return_to_lisp_function(os_context_t *context, lispobj function) +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_in_sigset_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); @@ -750,8 +1188,14 @@ void 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 @@ -784,6 +1228,37 @@ void arrange_return_to_lisp_function(os_context_t *context, lispobj function) u32 *sp=(u32 *)*os_context_register_addr(context,reg_ESP); +#if defined(LISP_FEATURE_DARWIN) + u32 *register_save_area = (u32 *)os_validate(0, 0x40); + + FSHOW_SIGNAL((stderr, "/arrange_return_to_lisp_function: preparing to go to function %x, sp: %x\n", function, sp)); + FSHOW_SIGNAL((stderr, "/arrange_return_to_lisp_function: context: %x, &context %x\n", context, &context)); + + /* 1. os_validate (malloc/mmap) register_save_block + * 2. copy register state into register_save_block + * 3. put a pointer to register_save_block in a register in the context + * 4. set the context's EIP to point to a trampoline which: + * a. builds the fake stack frame from the block + * b. frees the block + * c. calls the function + */ + + *register_save_area = *os_context_pc_addr(context); + *(register_save_area + 1) = function; + *(register_save_area + 2) = *os_context_register_addr(context,reg_EDI); + *(register_save_area + 3) = *os_context_register_addr(context,reg_ESI); + *(register_save_area + 4) = *os_context_register_addr(context,reg_EDX); + *(register_save_area + 5) = *os_context_register_addr(context,reg_ECX); + *(register_save_area + 6) = *os_context_register_addr(context,reg_EBX); + *(register_save_area + 7) = *os_context_register_addr(context,reg_EAX); + *(register_save_area + 8) = *context_eflags_addr(context); + + *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: */ *(sp-15) = (u32)post_signal_tramp; *(sp-14) = function; /* args for call_into_lisp : function*/ @@ -805,8 +1280,11 @@ void arrange_return_to_lisp_function(os_context_t *context, lispobj function) *(sp-2)=*os_context_register_addr(context,reg_EBP); *(sp-1)=*os_context_pc_addr(context); +#endif + #elif defined(LISP_FEATURE_X86_64) u64 *sp=(u64 *)*os_context_register_addr(context,reg_RSP); + /* return address for call_into_lisp: */ *(sp-18) = (u64)post_signal_tramp; @@ -839,6 +1317,8 @@ void arrange_return_to_lisp_function(os_context_t *context, lispobj function) #endif #ifdef LISP_FEATURE_X86 + +#if !defined(LISP_FEATURE_DARWIN) *os_context_pc_addr(context) = (os_context_register_t)call_into_lisp; *os_context_register_addr(context,reg_ECX) = 0; *os_context_register_addr(context,reg_EBP) = (os_context_register_t)(sp-2); @@ -847,7 +1327,9 @@ void arrange_return_to_lisp_function(os_context_t *context, lispobj function) (os_context_register_t)(sp-15); #else *os_context_register_addr(context,reg_ESP) = (os_context_register_t)(sp-15); -#endif +#endif /* __NETBSD__ */ +#endif /* LISP_FEATURE_DARWIN */ + #elif defined(LISP_FEATURE_X86_64) *os_context_pc_addr(context) = (os_context_register_t)call_into_lisp; *os_context_register_addr(context,reg_RCX) = 0; @@ -871,31 +1353,24 @@ void 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); - 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, * that has the added benefit of automatically discriminating between * functions and variables. */ -void undefined_alien_function() { - funcall0(SymbolFunction(UNDEFINED_ALIEN_FUNCTION_ERROR)); +void +undefined_alien_function(void) +{ + funcall0(StaticSymbolFunction(UNDEFINED_ALIEN_FUNCTION_ERROR)); } -boolean handle_guard_page_triggered(os_context_t *context,os_vm_address_t addr) +boolean +handle_guard_page_triggered(os_context_t *context,os_vm_address_t addr) { struct thread *th=arch_os_get_current_thread(); @@ -907,11 +1382,17 @@ boolean handle_guard_page_triggered(os_context_t *context,os_vm_address_t addr) * 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); + corruption_warning_and_maybe_lose("Control stack exhausted"); + protect_control_stack_guard_page(0, NULL); + protect_control_stack_return_guard_page(1, NULL); +#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) && @@ -920,134 +1401,189 @@ boolean 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); + fprintf(stderr, "INFO: Control stack guard page reprotected\n"); + protect_control_stack_guard_page(1, NULL); + protect_control_stack_return_guard_page(0, NULL); + return 1; + } + else if(addr >= BINDING_STACK_GUARD_PAGE(th) && + addr < BINDING_STACK_GUARD_PAGE(th) + os_vm_page_size) { + corruption_warning_and_maybe_lose("Binding stack exhausted"); + protect_binding_stack_guard_page(0, NULL); + protect_binding_stack_return_guard_page(1, NULL); + + /* 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) { + fprintf(stderr, "INFO: Binding stack guard page reprotected\n"); + protect_binding_stack_guard_page(1, NULL); + protect_binding_stack_return_guard_page(0, NULL); + return 1; + } + else if(addr >= ALIEN_STACK_GUARD_PAGE(th) && + addr < ALIEN_STACK_GUARD_PAGE(th) + os_vm_page_size) { + corruption_warning_and_maybe_lose("Alien stack exhausted"); + protect_alien_stack_guard_page(0, NULL); + protect_alien_stack_return_guard_page(1, NULL); + + /* 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) { + fprintf(stderr, "INFO: Alien stack guard page reprotected\n"); + protect_alien_stack_guard_page(1, NULL); + protect_alien_stack_return_guard_page(0, NULL); 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; } + +/* + * noise to install handlers + */ + +#ifndef LISP_FEATURE_WIN32 +/* In Linux 2.4 synchronous signals (sigtrap & co) can be delivered if + * they are blocked, in Linux 2.6 the default handler is invoked + * instead that usually coredumps. One might hastily think that adding + * SA_NODEFER helps, but until ~2.6.13 if SA_NODEFER is specified then + * the whole sa_mask is ignored and instead of not adding the signal + * in question to the mask. That means if it's not blockable the + * signal must be unblocked at the beginning of signal handlers. + * + * It turns out that NetBSD's SA_NODEFER doesn't DTRT in a different + * way: if SA_NODEFER is set and the signal is in sa_mask, the signal + * will be unblocked in the sigmask during the signal handler. -- RMK + * X-mas day, 2005 + */ +static volatile int sigaction_nodefer_works = -1; -#ifndef LISP_FEATURE_GENCGC -/* This function gets called from the SIGSEGV (for e.g. Linux, NetBSD, & - * OpenBSD) or SIGBUS (for e.g. FreeBSD) handler. Here we check - * whether the signal was due to treading on the mprotect()ed zone - - * and if so, arrange for a GC to happen. */ -extern unsigned long bytes_consed_between_gcs; /* gc-common.c */ +#define SA_NODEFER_TEST_BLOCK_SIGNAL SIGABRT +#define SA_NODEFER_TEST_KILL_SIGNAL SIGUSR1 -boolean -interrupt_maybe_gc(int signal, siginfo_t *info, void *void_context) +static void +sigaction_nodefer_test_handler(int signal, siginfo_t *info, void *void_context) { - os_context_t *context=(os_context_t *) void_context; - - if(!foreign_function_call_active && gc_trigger_hit(signal, info, context)){ - struct thread *thread=arch_os_get_current_thread(); - clear_auto_gc_trigger(); - /* Don't flood the system with interrupts if the need to gc is - * already noted. This can happen for example when SUB-GC - * allocates or after a gc triggered in a WITHOUT-GCING. */ - if (SymbolValue(GC_PENDING,thread) == NIL) { - if (SymbolValue(GC_INHIBIT,thread) == NIL) { - if (arch_pseudo_atomic_atomic(context)) { - /* set things up so that GC happens when we finish - * the PA section */ - SetSymbolValue(GC_PENDING,T,thread); - arch_set_pseudo_atomic_interrupted(context); - } else { - interrupt_maybe_gc_int(signal,info,void_context); - } - } else { - SetSymbolValue(GC_PENDING,T,thread); - } + sigset_t empty, current; + int i; + sigemptyset(&empty); + thread_sigmask(SIG_BLOCK, &empty, ¤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 + * 2.6.13 or so also doesn't block the other signal when + * SA_NODEFER is set. */ + for(i = 1; i < NSIG; i++) + if (sigismember(¤t, i) != + (((i == SA_NODEFER_TEST_BLOCK_SIGNAL) || (i == signal)) ? 1 : 0)) { + FSHOW_SIGNAL((stderr, "SA_NODEFER doesn't work, signal %d\n", i)); + sigaction_nodefer_works = 0; } - return 1; - } - return 0; + if (sigaction_nodefer_works == -1) + sigaction_nodefer_works = 1; } -#endif - -/* this is also used by gencgc, in alloc() */ -boolean -interrupt_maybe_gc_int(int signal, siginfo_t *info, void *void_context) +static void +see_if_sigaction_nodefer_works(void) { - os_context_t *context=(os_context_t *) void_context; - struct thread *thread=arch_os_get_current_thread(); + struct sigaction sa, old_sa; + + sa.sa_flags = SA_SIGINFO | SA_NODEFER; + sa.sa_sigaction = sigaction_nodefer_test_handler; + sigemptyset(&sa.sa_mask); + sigaddset(&sa.sa_mask, SA_NODEFER_TEST_BLOCK_SIGNAL); + sigaddset(&sa.sa_mask, SA_NODEFER_TEST_KILL_SIGNAL); + sigaction(SA_NODEFER_TEST_KILL_SIGNAL, &sa, &old_sa); + /* Make sure no signals are blocked. */ + { + sigset_t empty; + sigemptyset(&empty); + thread_sigmask(SIG_SETMASK, &empty, 0); + } + kill(getpid(), SA_NODEFER_TEST_KILL_SIGNAL); + while (sigaction_nodefer_works == -1); + sigaction(SA_NODEFER_TEST_KILL_SIGNAL, &old_sa, NULL); +} - fake_foreign_function_call(context); +#undef SA_NODEFER_TEST_BLOCK_SIGNAL +#undef SA_NODEFER_TEST_KILL_SIGNAL - /* SUB-GC may return without GCing if *GC-INHIBIT* is set, in - * which case we will be running with no gc trigger barrier - * thing for a while. But it shouldn't be long until the end - * of WITHOUT-GCING. - * - * FIXME: It would be good to protect the end of dynamic space - * and signal a storage condition from there. - */ +static void +unblock_me_trampoline(int signal, siginfo_t *info, void *void_context) +{ + SAVE_ERRNO(context,void_context); + sigset_t unblock; - /* Restore the signal mask from the interrupted context before - * calling into Lisp if interrupts are enabled. Why not always? - * - * Suppose there is a WITHOUT-INTERRUPTS block far, far out. If an - * interrupt hits while in SUB-GC, it is deferred and the - * os_context_sigmask of that interrupt is set to block further - * deferrable interrupts (until the first one is - * handled). Unfortunately, that context refers to this place and - * when we return from here the signals will not be blocked. - * - * A kludgy alternative is to propagate the sigmask change to the - * outer context. - */ - if(SymbolValue(INTERRUPTS_ENABLED,thread)!=NIL) - thread_sigmask(SIG_SETMASK, os_context_sigmask_addr(context), 0); -#ifdef LISP_FEATURE_SB_THREAD - else { - sigset_t new; - sigaddset(&new,SIG_STOP_FOR_GC); - thread_sigmask(SIG_UNBLOCK,&new,0); - } -#endif - funcall0(SymbolFunction(SUB_GC)); + sigemptyset(&unblock); + sigaddset(&unblock, signal); + thread_sigmask(SIG_UNBLOCK, &unblock, 0); + interrupt_handle_now(signal, info, context); + RESTORE_ERRNO; +} - undo_fake_foreign_function_call(context); - return 1; +static void +low_level_unblock_me_trampoline(int signal, siginfo_t *info, void *void_context) +{ + SAVE_ERRNO(context,void_context); + sigset_t unblock; + + sigemptyset(&unblock); + sigaddset(&unblock, signal); + thread_sigmask(SIG_UNBLOCK, &unblock, 0); + (*interrupt_low_level_handlers[signal])(signal, info, context); + RESTORE_ERRNO; } - -/* - * noise to install handlers - */ +static void +low_level_handle_now_handler(int signal, siginfo_t *info, void *void_context) +{ + SAVE_ERRNO(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; if (0 > signal || signal >= NSIG) { - lose("bad signal number %d", signal); + lose("bad signal number %d\n", signal); } - if (sigismember(&deferrable_sigset,signal)) + if (ARE_SAME_HANDLER(handler, SIG_DFL)) + sa.sa_sigaction = (void (*)(int, siginfo_t*, void*))handler; + else if (sigismember(&deferrable_sigset,signal)) sa.sa_sigaction = low_level_maybe_now_maybe_later; + else if (!sigaction_nodefer_works && + !sigismember(&blockable_sigset, signal)) + sa.sa_sigaction = low_level_unblock_me_trampoline; 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; + 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_INTERRUPT_THREAD - || (signal==SIG_INTERRUPT_THREAD) -#endif - ) + if((signal==SIG_MEMORY_FAULT)) sa.sa_flags |= SA_ONSTACK; #endif @@ -1055,11 +1591,13 @@ undoably_install_low_level_interrupt_handler (int signal, interrupt_low_level_handlers[signal] = (ARE_SAME_HANDLER(handler, SIG_DFL) ? 0 : handler); } +#endif /* 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; union interrupt_handler oldhandler; @@ -1074,16 +1612,19 @@ install_handler(int signal, void handler(int, siginfo_t*, void*)) (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; - } else if (sigismember(&deferrable_sigset, signal)) { + ARE_SAME_HANDLER(handler, SIG_IGN)) + sa.sa_sigaction = (void (*)(int, siginfo_t*, void*))handler; + else if (sigismember(&deferrable_sigset, signal)) sa.sa_sigaction = maybe_now_maybe_later; - } else { + else if (!sigaction_nodefer_works && + !sigismember(&blockable_sigset, signal)) + sa.sa_sigaction = unblock_me_trampoline; + else sa.sa_sigaction = interrupt_handle_now_handler; - } sigcopyset(&sa.sa_mask, &blockable_sigset); - sa.sa_flags = SA_SIGINFO | SA_RESTART; + sa.sa_flags = SA_SIGINFO | SA_RESTART | + (sigaction_nodefer_works ? SA_NODEFER : 0); sigaction(signal, &sa, NULL); } @@ -1095,17 +1636,33 @@ install_handler(int signal, void handler(int, siginfo_t*, void*)) FSHOW((stderr, "/leaving POSIX install_handler(%d, ..)\n", signal)); return (unsigned long)oldhandler.lisp; +#else + /* Probably-wrong Win32 hack */ + return 0; +#endif } +/* This must not go through lisp as it's allowed anytime, even when on + * the altstack. */ void -interrupt_init() +sigabrt_handler(int signal, siginfo_t *info, os_context_t *context) { + lose("SIGABRT received.\n"); +} + +void +interrupt_init(void) +{ +#ifndef LISP_FEATURE_WIN32 int i; SHOW("entering 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++) { @@ -1115,8 +1672,94 @@ 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 +} + +#ifndef LISP_FEATURE_WIN32 +int +siginfo_code(siginfo_t *info) +{ + return info->si_code; +} +os_vm_address_t current_memory_fault_address; + +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 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; + /* To allow debugging memory faults in signal handlers and such. */ + corruption_warning_and_maybe_lose("Memory fault at %x", addr); + 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 + +static void +unhandled_trap_error(os_context_t *context) +{ + lispobj context_sap; + fake_foreign_function_call(context); + unblock_gc_signals(); + context_sap = alloc_sap(context); +#ifndef LISP_FEATURE_WIN32 + thread_sigmask(SIG_SETMASK, os_context_sigmask_addr(context), 0); +#endif + funcall1(StaticSymbolFunction(UNHANDLED_TRAP_ERROR), context_sap); + lose("UNHANDLED-TRAP-ERROR fell through"); +} + +/* 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) +{ + switch(trap) { + case trap_PendingInterrupt: + FSHOW((stderr, "/\n")); + arch_skip_instruction(context); + interrupt_handle_pending(context); + break; + case trap_Error: + case trap_Cerror: + FSHOW((stderr, "/\n", trap)); + interrupt_internal_error(context, trap==trap_Cerror); + break; + case trap_Breakpoint: + arch_handle_breakpoint(context); + break; + case trap_FunEndBreakpoint: + arch_handle_fun_end_breakpoint(context); + break; +#ifdef trap_AfterBreakpoint + case trap_AfterBreakpoint: + arch_handle_after_breakpoint(context); + break; +#endif +#ifdef trap_SingleStepAround + case trap_SingleStepAround: + case trap_SingleStepBefore: + arch_handle_single_step_trap(context, trap); + break; +#endif + case trap_Halt: + fake_foreign_function_call(context); + lose("%%PRIMITIVE HALT called; the party is over.\n"); + default: + unhandled_trap_error(context); + } }