X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fppc-arch.c;h=5e6b8851ecffa9527553ac38808ef6623cd2e10a;hb=02d7495253b9075e4d86275590c3e827c814b596;hp=fb5958ecf389ca632ef8770af5257e2a218f7b83;hpb=343ef95fda9da33830d5ab6aabe5368c898f0918;p=sbcl.git diff --git a/src/runtime/ppc-arch.c b/src/runtime/ppc-arch.c index fb5958e..5e6b885 100644 --- a/src/runtime/ppc-arch.c +++ b/src/runtime/ppc-arch.c @@ -1,3 +1,14 @@ +/* + * This software is part of the SBCL system. See the README file for + * more information. + * + * This software is derived from the CMU CL system, which was + * written at Carnegie Mellon University and released into the + * public domain. The software is in the public domain and is + * provided with absolutely no warranty. See the COPYING and CREDITS + * files for more information. + */ + #include #include "sbcl.h" @@ -9,6 +20,12 @@ #include "signal.h" #include "interrupt.h" #include "interr.h" +#include "breakpoint.h" +#include "alloc.h" + +#if defined(LISP_FEATURE_GENCGC) +#include "gencgc-alloc-region.h" +#endif /* The header files may not define PT_DAR/PT_DSISR. This definition is correct for all versions of ppc linux >= 2.0.30 @@ -20,43 +37,37 @@ Even with the patch, the DSISR may not have its 'write' bit set correctly (it tends not to be set if the fault was caused by something other than a protection violation.) - + Caveat callers. */ +#if defined (LISP_FEATURE_DARWIN) || defined(LISP_FEATURE_LINUX) #ifndef PT_DAR -#define PT_DAR 41 +#define PT_DAR 41 #endif #ifndef PT_DSISR -#define PT_DSISR 42 +#define PT_DSISR 42 +#endif #endif void arch_init() { } -os_vm_address_t +os_vm_address_t arch_get_bad_addr(int sig, siginfo_t *code, os_context_t *context) { - unsigned int *pc = (unsigned int *)(*os_context_pc_addr(context)); os_vm_address_t addr; - - - /* Make sure it's not the pc thats bogus, and that it was lisp code */ - /* that caused the fault. */ - if ((((unsigned long)pc) & 3) != 0 || - ((pc < READ_ONLY_SPACE_START || - pc >= READ_ONLY_SPACE_START+READ_ONLY_SPACE_SIZE) && - ((lispobj *)pc < current_dynamic_space || - (lispobj *)pc >= current_dynamic_space + DYNAMIC_SPACE_SIZE))) - return 0; - - + +#if defined(LISP_FEATURE_NETBSD) + addr = (os_vm_address_t) (code->si_addr); +#else addr = (os_vm_address_t) (*os_context_register_addr(context,PT_DAR)); +#endif return addr; } - -void + +void arch_skip_instruction(os_context_t *context) { char** pcptr; @@ -71,190 +82,393 @@ arch_internal_error_arguments(os_context_t *context) } -boolean +boolean arch_pseudo_atomic_atomic(os_context_t *context) { - return ((*os_context_register_addr(context,reg_ALLOC)) & 4); + /* 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 + * + * The foreign_function_call_active used to live at each call-site + * to arch_pseudo_atomic_atomic, but this seems clearer. + * --NS 2007-05-15 */ + return (!foreign_function_call_active) + && ((*os_context_register_addr(context,reg_ALLOC)) & 4); } -#define PSEUDO_ATOMIC_INTERRUPTED_BIAS 0x7f000000 - -void +void arch_set_pseudo_atomic_interrupted(os_context_t *context) { - *os_context_register_addr(context,reg_NL3) - += PSEUDO_ATOMIC_INTERRUPTED_BIAS; + *os_context_register_addr(context,reg_ALLOC) |= 1; } -unsigned long +void +arch_clear_pseudo_atomic_interrupted(os_context_t *context) +{ + *os_context_register_addr(context,reg_ALLOC) &= ~1; +} + +unsigned int arch_install_breakpoint(void *pc) { - unsigned long *ptr = (unsigned long *)pc; - unsigned long result = *ptr; + unsigned int *ptr = (unsigned int *)pc; + unsigned int result = *ptr; *ptr = (3<<26) | (5 << 21) | trap_Breakpoint; - os_flush_icache((os_vm_address_t) pc, sizeof(unsigned long)); + os_flush_icache((os_vm_address_t) pc, sizeof(unsigned int)); return result; } -void -arch_remove_breakpoint(void *pc, unsigned long orig_inst) +void +arch_remove_breakpoint(void *pc, unsigned int orig_inst) { - *(unsigned long *)pc = orig_inst; - os_flush_icache((os_vm_address_t) pc, sizeof(unsigned long)); + *(unsigned int *)pc = orig_inst; + os_flush_icache((os_vm_address_t) pc, sizeof(unsigned int)); } -static unsigned long *skipped_break_addr, displaced_after_inst; +/* + * Perform the instruction that we overwrote with a breakpoint. As we + * don't have a single-step facility, this means we have to: + * - put the instruction back + * - put a second breakpoint at the following instruction, + * set after_breakpoint and continue execution. + * + * When the second breakpoint is hit (very shortly thereafter, we hope) + * sigtrap_handler gets called again, but follows the AfterBreakpoint + * arm, which + * - puts a bpt back in the first breakpoint place (running across a + * breakpoint shouldn't cause it to be uninstalled) + * - replaces the second bpt with the instruction it was meant to be + * - carries on + * + * Clear? + */ +static unsigned int *skipped_break_addr, displaced_after_inst; static sigset_t orig_sigmask; -void -arch_do_displaced_inst(os_context_t *context,unsigned int orig_inst) +void +arch_do_displaced_inst(os_context_t *context, unsigned int orig_inst) { /* not sure how we ensure that we get the breakpoint reinstalled * after doing this -dan */ - unsigned long *pc = (unsigned long *)(*os_context_pc_addr(context)); - + unsigned int *pc = (unsigned int *)(*os_context_pc_addr(context)); + orig_sigmask = *os_context_sigmask_addr(context); sigaddset_blockable(os_context_sigmask_addr(context)); - + *pc = orig_inst; - os_flush_icache((os_vm_address_t) pc, sizeof(unsigned long)); + os_flush_icache((os_vm_address_t) pc, sizeof(unsigned int)); skipped_break_addr = pc; + + /* FIXME: we should apparently be installing the after-breakpoint + * here, but would need to find the next instruction address for + * it first. alpha-arch.c shows how to do it. --NS 2007-04-02 */ } -static void -sigtrap_handler(int signal, siginfo_t *siginfo, os_context_t *context) +#ifdef LISP_FEATURE_GENCGC +/* + * Return non-zero if the current instruction is an allocation trap + */ +static int +allocation_trap_p(os_context_t * context) { - u32 code; - sigset_t *mask; -#ifdef LISP_FEATURE_LINUX - os_restore_fp_control(context); + int result; + unsigned int *pc; + unsigned inst; + unsigned opcode; + unsigned src; + unsigned dst; + + result = 0; + + /* + * First, the instruction has to be a TWLGE temp, NL3, which has the + * format. + * | 6| 5| 5 | 5 | 10|1| width + * |31|5 |dst|src| 4|0| field + */ + pc = (unsigned int *) (*os_context_pc_addr(context)); + inst = *pc; + +#if 0 + fprintf(stderr, "allocation_trap_p at %p: inst = 0x%08x\n", pc, inst); #endif - mask=(os_context_sigmask_addr(context)); - sigsetmask(mask); - code=*((u32 *)(*os_context_pc_addr(context))); - if (code == ((3 << 26) | (16 << 21) | (reg_ALLOC << 16))) { - /* twlti reg_ALLOC,0 - check for deferred interrupt */ - *os_context_register_addr(context,reg_ALLOC) - -= PSEUDO_ATOMIC_INTERRUPTED_BIAS; - arch_skip_instruction(context); - /* interrupt or GC was requested in PA; now we're done with the - PA section we may as well get around to it */ - interrupt_handle_pending(context); - return; - + + opcode = inst >> 26; + src = (inst >> 11) & 0x1f; + dst = (inst >> 16) & 0x1f; + if ((opcode == 31) && (src == reg_NL3) && (5 == ((inst >> 21) & 0x1f)) + && (4 == ((inst >> 1) & 0x3ff))) { + /* + * We got the instruction. Now, look back to make sure it was + * proceeded by what we expected. 2 instructions back should be + * an ADD or ADDI instruction. + */ + unsigned int add_inst; + + add_inst = pc[-3]; +#if 0 + fprintf(stderr, " add inst at %p: inst = 0x%08x\n", + pc - 3, add_inst); +#endif + opcode = add_inst >> 26; + if ((opcode == 31) && (266 == ((add_inst >> 1) & 0x1ff))) { + return 1; + } else if ((opcode == 14)) { + return 1; + } else { + fprintf(stderr, + "Whoa! Got allocation trap but could not find ADD or ADDI instruction: 0x%08x in the proper place\n", + add_inst); + } } - if ((code >> 16) == ((3 << 10) | (6 << 5))) { - /* twllei reg_ZERO,N will always trap if reg_ZERO = 0 */ - int trap = code & 0x1f; - - switch (trap) { - case trap_Halt: - fake_foreign_function_call(context); - lose("%%primitive halt called; the party is over.\n"); - - case trap_Error: - case trap_Cerror: - interrupt_internal_error(signal, code, context, trap == trap_Cerror); - break; - - case trap_PendingInterrupt: - /* when do we run this branch instead of the twlti code above? */ - arch_skip_instruction(context); - interrupt_handle_pending(context); - break; - - case trap_Breakpoint: - handle_breakpoint(signal, code, context); - break; - - case trap_FunEndBreakpoint: - *os_context_pc_addr(context) - =(int)handle_fun_end_breakpoint(signal, code, context); - break; - - case trap_AfterBreakpoint: - *skipped_break_addr = trap_Breakpoint; - skipped_break_addr = NULL; - *(unsigned long *)*os_context_pc_addr(context) - = displaced_after_inst; - *os_context_sigmask_addr(context)= orig_sigmask; - - os_flush_icache((os_vm_address_t) *os_context_pc_addr(context), - sizeof(unsigned long)); - break; - - default: - interrupt_handle_now(signal, code, context); - break; - } -#ifdef LISP_FEATURE_DARWIN - DARWIN_FIX_CONTEXT(context); + return 0; +} + +extern struct alloc_region boxed_region; + +void +handle_allocation_trap(os_context_t * context) +{ + unsigned int *pc; + unsigned int inst; + unsigned int target, target_ptr, end_addr; + unsigned int opcode; + int size; + boolean were_in_lisp; + char *memory; + + target = 0; + size = 0; + +#if 0 + fprintf(stderr, "In handle_allocation_trap\n"); #endif - return; + + /* I don't think it's possible for us NOT to be in lisp when we get + * here. Remove this later? */ + were_in_lisp = !foreign_function_call_active; + + if (were_in_lisp) { + fake_foreign_function_call(context); + } else { + fprintf(stderr, "**** Whoa! allocation trap and we weren't in lisp!\n"); } - if (((code >> 26) == 3) && (((code >> 21) & 31) == 24)) { - interrupt_internal_error(signal, code, context, 0); -#ifdef LISP_FEATURE_DARWIN - DARWIN_FIX_CONTEXT(context); + + /* + * Look at current instruction: TWNE temp, NL3. We're here because + * temp > NL3 and temp is the end of the allocation, and NL3 is + * current-region-end-addr. + * + * We need to adjust temp and alloc-tn. + */ + + pc = (unsigned int *) (*os_context_pc_addr(context)); + inst = pc[0]; + end_addr = (inst >> 11) & 0x1f; + target = (inst >> 16) & 0x1f; + + target_ptr = *os_context_register_addr(context, target); + +#if 0 + fprintf(stderr, "handle_allocation_trap at %p:\n", pc); + fprintf(stderr, "boxed_region.free_pointer: %p\n", boxed_region.free_pointer); + fprintf(stderr, "boxed_region.end_addr: %p\n", boxed_region.end_addr); + fprintf(stderr, "target reg: %d, end_addr reg: %d\n", target, end_addr); + fprintf(stderr, "target: %x\n", *os_context_register_addr(context, target)); + fprintf(stderr, "end_addr: %x\n", *os_context_register_addr(context, end_addr)); +#endif + +#if 0 + fprintf(stderr, "handle_allocation_trap at %p:\n", pc); + fprintf(stderr, " trap inst = 0x%08x\n", inst); + fprintf(stderr, " target reg = %s\n", lisp_register_names[target]); +#endif + + /* + * Go back and look at the add/addi instruction. The second src arg + * is the size of the allocation. Get it and call alloc to allocate + * new space. + */ + inst = pc[-3]; + opcode = inst >> 26; +#if 0 + fprintf(stderr, " add inst = 0x%08x, opcode = %d\n", inst, opcode); +#endif + if (opcode == 14) { + /* + * ADDI temp-tn, alloc-tn, size + * + * Extract the size + */ + size = (inst & 0xffff); + } else if (opcode == 31) { + /* + * ADD temp-tn, alloc-tn, size-tn + * + * Extract the size + */ + int reg; + + reg = (inst >> 11) & 0x1f; +#if 0 + fprintf(stderr, " add, reg = %s\n", lisp_register_names[reg]); +#endif + size = *os_context_register_addr(context, reg); + + } + +#if 0 + fprintf(stderr, "Alloc %d to %s\n", size, lisp_register_names[target]); #endif - return; + +#if INLINE_ALLOC_DEBUG + if ((((unsigned long)boxed_region.end_addr + size) / PAGE_SIZE) == + (((unsigned long)boxed_region.end_addr) / PAGE_SIZE)) { + fprintf(stderr,"*** possibly bogus trap allocation of %d bytes at %p\n", + size, target_ptr); + fprintf(stderr, " dynamic_space_free_pointer: %p, boxed_region.end_addr %p\n", + dynamic_space_free_pointer, boxed_region.end_addr); } - - interrupt_handle_now(signal, code, context); -#ifdef LISP_FEATURE_DARWIN - /* Work around G5 bug */ - DARWIN_FIX_CONTEXT(context); #endif -} +#if 0 + fprintf(stderr, "Ready to alloc\n"); + fprintf(stderr, "free_pointer = 0x%08x\n", + dynamic_space_free_pointer); +#endif -void arch_install_interrupt_handlers() -{ - undoably_install_low_level_interrupt_handler(SIGILL,sigtrap_handler); - undoably_install_low_level_interrupt_handler(SIGTRAP,sigtrap_handler); -} + /* + * alloc-tn was incremented by size. Need to decrement it by size + * to restore its original value. This is not true on GENCGC + * anymore. d_s_f_p and reg_alloc get out of sync, but the p_a + * bits stay intact and we set it to the proper value when it + * needs to be. Keep this comment here for the moment in case + * somebody tries to figure out what happened here. + */ + /* dynamic_space_free_pointer = + (lispobj *) ((long) dynamic_space_free_pointer - size); + */ +#if 0 + fprintf(stderr, "free_pointer = 0x%08x new\n", + dynamic_space_free_pointer); +#endif + memory = (char *) alloc(size); -extern lispobj call_into_lisp(lispobj fun, lispobj *args, int nargs); +#if 0 + fprintf(stderr, "alloc returned %p\n", memory); + fprintf(stderr, "free_pointer = 0x%08x\n", + dynamic_space_free_pointer); +#endif -lispobj funcall0(lispobj function) -{ - lispobj *args = current_control_stack_pointer; + /* + * The allocation macro wants the result to point to the end of the + * object! + */ + memory += size; - return call_into_lisp(function, args, 0); -} +#if 0 + fprintf(stderr, "object end at %p\n", memory); +#endif -lispobj funcall1(lispobj function, lispobj arg0) -{ - lispobj *args = current_control_stack_pointer; + *os_context_register_addr(context, target) = (unsigned long) memory; + *os_context_register_addr(context, reg_ALLOC) = + (unsigned long) dynamic_space_free_pointer + | (*os_context_register_addr(context, reg_ALLOC) + & LOWTAG_MASK); + + if (were_in_lisp) { + undo_fake_foreign_function_call(context); + } - current_control_stack_pointer += 1; - args[0] = arg0; - return call_into_lisp(function, args, 1); } +#endif -lispobj funcall2(lispobj function, lispobj arg0, lispobj arg1) +void +arch_handle_breakpoint(os_context_t *context) { - lispobj *args = current_control_stack_pointer; + handle_breakpoint(context); +} - current_control_stack_pointer += 2; - args[0] = arg0; - args[1] = arg1; +void +arch_handle_fun_end_breakpoint(os_context_t *context) +{ + *os_context_pc_addr(context) + =(int)handle_fun_end_breakpoint(context); +} - return call_into_lisp(function, args, 2); +void +arch_handle_after_breakpoint(os_context_t *context) +{ + *skipped_break_addr = trap_Breakpoint; + skipped_break_addr = NULL; + *(unsigned int *)*os_context_pc_addr(context) + = displaced_after_inst; + *os_context_sigmask_addr(context)= orig_sigmask; + os_flush_icache((os_vm_address_t) *os_context_pc_addr(context), + sizeof(unsigned int)); } -lispobj funcall3(lispobj function, lispobj arg0, lispobj arg1, lispobj arg2) +void +arch_handle_single_step_trap(os_context_t *context, int trap) { - lispobj *args = current_control_stack_pointer; + unsigned int code = *((u32 *)(*os_context_pc_addr(context))); + int register_offset = code >> 5 & 0x1f; + handle_single_step_trap(context, trap, register_offset); + arch_skip_instruction(context); +} - current_control_stack_pointer += 3; - args[0] = arg0; - args[1] = arg1; - args[2] = arg2; +static void +sigtrap_handler(int signal, siginfo_t *siginfo, void *void_context) +{ + unsigned int code; + os_context_t *context = void_context; + +#ifdef LISP_FEATURE_LINUX + os_restore_fp_control(context); +#endif + code=*((u32 *)(*os_context_pc_addr(context))); + if (code == ((3 << 26) | (0x18 << 21) | (reg_NL3 << 16))) { + arch_clear_pseudo_atomic_interrupted(context); + arch_skip_instruction(context); + /* interrupt or GC was requested in PA; now we're done with the + PA section we may as well get around to it */ + interrupt_handle_pending(context); + return; + } - return call_into_lisp(function, args, 3); +#ifdef LISP_FEATURE_GENCGC + /* Is this an allocation trap? */ + if (allocation_trap_p(context)) { + handle_allocation_trap(context); + arch_skip_instruction(context); + return; + } +#endif + + if ((code >> 16) == ((3 << 10) | (6 << 5))) { + /* twllei reg_ZERO,N will always trap if reg_ZERO = 0 */ + int trap = code & 0x1f; + handle_trap(context,trap); + return; + } + if (((code >> 26) == 3) && (((code >> 21) & 31) == 24)) { + interrupt_internal_error(context, 0); + return; + } + + interrupt_handle_now(signal, (siginfo_t *)code, context); +} + + +void arch_install_interrupt_handlers() +{ + undoably_install_low_level_interrupt_handler(SIGILL, sigtrap_handler); + undoably_install_low_level_interrupt_handler(SIGTRAP, sigtrap_handler); } void @@ -302,7 +516,7 @@ arch_write_linkage_table_jmp(void* reloc_addr, void *target_addr) * ori 13, 13, (low part of addr) * mtctr 13 * bctr - * + * */ int* inst_ptr; unsigned long hi; /* Top 16 bits of address */ @@ -322,7 +536,7 @@ arch_write_linkage_table_jmp(void* reloc_addr, void *target_addr) /* * addis 13, 0, (hi part) */ - + inst = (15 << 26) | (LINKAGE_TEMP_REG << 21) | (0 << 16) | hi; *inst_ptr++ = inst; @@ -332,7 +546,7 @@ arch_write_linkage_table_jmp(void* reloc_addr, void *target_addr) inst = (24 << 26) | (LINKAGE_TEMP_REG << 21) | (LINKAGE_TEMP_REG << 16) | lo; *inst_ptr++ = inst; - + /* * mtctr 13 */ @@ -349,11 +563,11 @@ arch_write_linkage_table_jmp(void* reloc_addr, void *target_addr) *inst_ptr++ = inst; - + os_flush_icache((os_vm_address_t) reloc_addr, (char*) inst_ptr - (char*) reloc_addr); } -void +void arch_write_linkage_table_ref(void * reloc_addr, void *target_addr) { *(unsigned long *)reloc_addr = (unsigned long)target_addr;