15aae8a53SK.Prasad /* 25aae8a53SK.Prasad * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility, 35aae8a53SK.Prasad * using the CPU's debug registers. Derived from 45aae8a53SK.Prasad * "arch/x86/kernel/hw_breakpoint.c" 55aae8a53SK.Prasad * 65aae8a53SK.Prasad * This program is free software; you can redistribute it and/or modify 75aae8a53SK.Prasad * it under the terms of the GNU General Public License as published by 85aae8a53SK.Prasad * the Free Software Foundation; either version 2 of the License, or 95aae8a53SK.Prasad * (at your option) any later version. 105aae8a53SK.Prasad * 115aae8a53SK.Prasad * This program is distributed in the hope that it will be useful, 125aae8a53SK.Prasad * but WITHOUT ANY WARRANTY; without even the implied warranty of 135aae8a53SK.Prasad * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 145aae8a53SK.Prasad * GNU General Public License for more details. 155aae8a53SK.Prasad * 165aae8a53SK.Prasad * You should have received a copy of the GNU General Public License 175aae8a53SK.Prasad * along with this program; if not, write to the Free Software 185aae8a53SK.Prasad * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. 195aae8a53SK.Prasad * 205aae8a53SK.Prasad * Copyright 2010 IBM Corporation 215aae8a53SK.Prasad * Author: K.Prasad <prasad@linux.vnet.ibm.com> 225aae8a53SK.Prasad * 235aae8a53SK.Prasad */ 245aae8a53SK.Prasad 255aae8a53SK.Prasad #include <linux/hw_breakpoint.h> 265aae8a53SK.Prasad #include <linux/notifier.h> 275aae8a53SK.Prasad #include <linux/kprobes.h> 285aae8a53SK.Prasad #include <linux/percpu.h> 295aae8a53SK.Prasad #include <linux/kernel.h> 305aae8a53SK.Prasad #include <linux/module.h> 315aae8a53SK.Prasad #include <linux/sched.h> 325aae8a53SK.Prasad #include <linux/init.h> 335aae8a53SK.Prasad #include <linux/smp.h> 345aae8a53SK.Prasad 355aae8a53SK.Prasad #include <asm/hw_breakpoint.h> 365aae8a53SK.Prasad #include <asm/processor.h> 375aae8a53SK.Prasad #include <asm/sstep.h> 385aae8a53SK.Prasad #include <asm/uaccess.h> 395aae8a53SK.Prasad 405aae8a53SK.Prasad /* 415aae8a53SK.Prasad * Stores the breakpoints currently in use on each breakpoint address 425aae8a53SK.Prasad * register for every cpu 435aae8a53SK.Prasad */ 445aae8a53SK.Prasad static DEFINE_PER_CPU(struct perf_event *, bp_per_reg); 455aae8a53SK.Prasad 465aae8a53SK.Prasad /* 475aae8a53SK.Prasad * Install a perf counter breakpoint. 485aae8a53SK.Prasad * 495aae8a53SK.Prasad * We seek a free debug address register and use it for this 505aae8a53SK.Prasad * breakpoint. 515aae8a53SK.Prasad * 525aae8a53SK.Prasad * Atomic: we hold the counter->ctx->lock and we only handle variables 535aae8a53SK.Prasad * and registers local to this cpu. 545aae8a53SK.Prasad */ 555aae8a53SK.Prasad int arch_install_hw_breakpoint(struct perf_event *bp) 565aae8a53SK.Prasad { 575aae8a53SK.Prasad struct arch_hw_breakpoint *info = counter_arch_bp(bp); 585aae8a53SK.Prasad struct perf_event **slot = &__get_cpu_var(bp_per_reg); 595aae8a53SK.Prasad 605aae8a53SK.Prasad *slot = bp; 615aae8a53SK.Prasad 625aae8a53SK.Prasad /* 635aae8a53SK.Prasad * Do not install DABR values if the instruction must be single-stepped. 645aae8a53SK.Prasad * If so, DABR will be populated in single_step_dabr_instruction(). 655aae8a53SK.Prasad */ 665aae8a53SK.Prasad if (current->thread.last_hit_ubp != bp) 675aae8a53SK.Prasad set_dabr(info->address | info->type | DABR_TRANSLATION); 685aae8a53SK.Prasad 695aae8a53SK.Prasad return 0; 705aae8a53SK.Prasad } 715aae8a53SK.Prasad 725aae8a53SK.Prasad /* 735aae8a53SK.Prasad * Uninstall the breakpoint contained in the given counter. 745aae8a53SK.Prasad * 755aae8a53SK.Prasad * First we search the debug address register it uses and then we disable 765aae8a53SK.Prasad * it. 775aae8a53SK.Prasad * 785aae8a53SK.Prasad * Atomic: we hold the counter->ctx->lock and we only handle variables 795aae8a53SK.Prasad * and registers local to this cpu. 805aae8a53SK.Prasad */ 815aae8a53SK.Prasad void arch_uninstall_hw_breakpoint(struct perf_event *bp) 825aae8a53SK.Prasad { 835aae8a53SK.Prasad struct perf_event **slot = &__get_cpu_var(bp_per_reg); 845aae8a53SK.Prasad 855aae8a53SK.Prasad if (*slot != bp) { 865aae8a53SK.Prasad WARN_ONCE(1, "Can't find the breakpoint"); 875aae8a53SK.Prasad return; 885aae8a53SK.Prasad } 895aae8a53SK.Prasad 905aae8a53SK.Prasad *slot = NULL; 915aae8a53SK.Prasad set_dabr(0); 925aae8a53SK.Prasad } 935aae8a53SK.Prasad 945aae8a53SK.Prasad /* 955aae8a53SK.Prasad * Perform cleanup of arch-specific counters during unregistration 965aae8a53SK.Prasad * of the perf-event 975aae8a53SK.Prasad */ 985aae8a53SK.Prasad void arch_unregister_hw_breakpoint(struct perf_event *bp) 995aae8a53SK.Prasad { 1005aae8a53SK.Prasad /* 1015aae8a53SK.Prasad * If the breakpoint is unregistered between a hw_breakpoint_handler() 1025aae8a53SK.Prasad * and the single_step_dabr_instruction(), then cleanup the breakpoint 1035aae8a53SK.Prasad * restoration variables to prevent dangling pointers. 1045aae8a53SK.Prasad */ 1055aae8a53SK.Prasad if (bp->ctx->task) 1065aae8a53SK.Prasad bp->ctx->task->thread.last_hit_ubp = NULL; 1075aae8a53SK.Prasad } 1085aae8a53SK.Prasad 1095aae8a53SK.Prasad /* 1105aae8a53SK.Prasad * Check for virtual address in kernel space. 1115aae8a53SK.Prasad */ 1125aae8a53SK.Prasad int arch_check_bp_in_kernelspace(struct perf_event *bp) 1135aae8a53SK.Prasad { 1145aae8a53SK.Prasad struct arch_hw_breakpoint *info = counter_arch_bp(bp); 1155aae8a53SK.Prasad 1165aae8a53SK.Prasad return is_kernel_addr(info->address); 1175aae8a53SK.Prasad } 1185aae8a53SK.Prasad 1195aae8a53SK.Prasad int arch_bp_generic_fields(int type, int *gen_bp_type) 1205aae8a53SK.Prasad { 1215aae8a53SK.Prasad switch (type) { 1225aae8a53SK.Prasad case DABR_DATA_READ: 1235aae8a53SK.Prasad *gen_bp_type = HW_BREAKPOINT_R; 1245aae8a53SK.Prasad break; 1255aae8a53SK.Prasad case DABR_DATA_WRITE: 1265aae8a53SK.Prasad *gen_bp_type = HW_BREAKPOINT_W; 1275aae8a53SK.Prasad break; 1285aae8a53SK.Prasad case (DABR_DATA_WRITE | DABR_DATA_READ): 1295aae8a53SK.Prasad *gen_bp_type = (HW_BREAKPOINT_W | HW_BREAKPOINT_R); 1305aae8a53SK.Prasad break; 1315aae8a53SK.Prasad default: 1325aae8a53SK.Prasad return -EINVAL; 1335aae8a53SK.Prasad } 1345aae8a53SK.Prasad return 0; 1355aae8a53SK.Prasad } 1365aae8a53SK.Prasad 1375aae8a53SK.Prasad /* 1385aae8a53SK.Prasad * Validate the arch-specific HW Breakpoint register settings 1395aae8a53SK.Prasad */ 1405aae8a53SK.Prasad int arch_validate_hwbkpt_settings(struct perf_event *bp) 1415aae8a53SK.Prasad { 1425aae8a53SK.Prasad int ret = -EINVAL; 1435aae8a53SK.Prasad struct arch_hw_breakpoint *info = counter_arch_bp(bp); 1445aae8a53SK.Prasad 1455aae8a53SK.Prasad if (!bp) 1465aae8a53SK.Prasad return ret; 1475aae8a53SK.Prasad 1485aae8a53SK.Prasad switch (bp->attr.bp_type) { 1495aae8a53SK.Prasad case HW_BREAKPOINT_R: 1505aae8a53SK.Prasad info->type = DABR_DATA_READ; 1515aae8a53SK.Prasad break; 1525aae8a53SK.Prasad case HW_BREAKPOINT_W: 1535aae8a53SK.Prasad info->type = DABR_DATA_WRITE; 1545aae8a53SK.Prasad break; 1555aae8a53SK.Prasad case HW_BREAKPOINT_R | HW_BREAKPOINT_W: 1565aae8a53SK.Prasad info->type = (DABR_DATA_READ | DABR_DATA_WRITE); 1575aae8a53SK.Prasad break; 1585aae8a53SK.Prasad default: 1595aae8a53SK.Prasad return ret; 1605aae8a53SK.Prasad } 1615aae8a53SK.Prasad 1625aae8a53SK.Prasad info->address = bp->attr.bp_addr; 1635aae8a53SK.Prasad info->len = bp->attr.bp_len; 1645aae8a53SK.Prasad 1655aae8a53SK.Prasad /* 1665aae8a53SK.Prasad * Since breakpoint length can be a maximum of HW_BREAKPOINT_LEN(8) 1675aae8a53SK.Prasad * and breakpoint addresses are aligned to nearest double-word 1685aae8a53SK.Prasad * HW_BREAKPOINT_ALIGN by rounding off to the lower address, the 1695aae8a53SK.Prasad * 'symbolsize' should satisfy the check below. 1705aae8a53SK.Prasad */ 1715aae8a53SK.Prasad if (info->len > 1725aae8a53SK.Prasad (HW_BREAKPOINT_LEN - (info->address & HW_BREAKPOINT_ALIGN))) 1735aae8a53SK.Prasad return -EINVAL; 1745aae8a53SK.Prasad return 0; 1755aae8a53SK.Prasad } 1765aae8a53SK.Prasad 1775aae8a53SK.Prasad /* 178*06532a67SK.Prasad * Restores the breakpoint on the debug registers. 179*06532a67SK.Prasad * Invoke this function if it is known that the execution context is 180*06532a67SK.Prasad * about to change to cause loss of MSR_SE settings. 181*06532a67SK.Prasad */ 182*06532a67SK.Prasad void thread_change_pc(struct task_struct *tsk, struct pt_regs *regs) 183*06532a67SK.Prasad { 184*06532a67SK.Prasad struct arch_hw_breakpoint *info; 185*06532a67SK.Prasad 186*06532a67SK.Prasad if (likely(!tsk->thread.last_hit_ubp)) 187*06532a67SK.Prasad return; 188*06532a67SK.Prasad 189*06532a67SK.Prasad info = counter_arch_bp(tsk->thread.last_hit_ubp); 190*06532a67SK.Prasad regs->msr &= ~MSR_SE; 191*06532a67SK.Prasad set_dabr(info->address | info->type | DABR_TRANSLATION); 192*06532a67SK.Prasad tsk->thread.last_hit_ubp = NULL; 193*06532a67SK.Prasad } 194*06532a67SK.Prasad 195*06532a67SK.Prasad /* 1965aae8a53SK.Prasad * Handle debug exception notifications. 1975aae8a53SK.Prasad */ 1985aae8a53SK.Prasad int __kprobes hw_breakpoint_handler(struct die_args *args) 1995aae8a53SK.Prasad { 2005aae8a53SK.Prasad bool is_ptrace_bp = false; 2015aae8a53SK.Prasad int rc = NOTIFY_STOP; 2025aae8a53SK.Prasad struct perf_event *bp; 2035aae8a53SK.Prasad struct pt_regs *regs = args->regs; 2045aae8a53SK.Prasad int stepped = 1; 2055aae8a53SK.Prasad struct arch_hw_breakpoint *info; 2065aae8a53SK.Prasad unsigned int instr; 2075aae8a53SK.Prasad 2085aae8a53SK.Prasad /* Disable breakpoints during exception handling */ 2095aae8a53SK.Prasad set_dabr(0); 2105aae8a53SK.Prasad /* 2115aae8a53SK.Prasad * The counter may be concurrently released but that can only 2125aae8a53SK.Prasad * occur from a call_rcu() path. We can then safely fetch 2135aae8a53SK.Prasad * the breakpoint, use its callback, touch its counter 2145aae8a53SK.Prasad * while we are in an rcu_read_lock() path. 2155aae8a53SK.Prasad */ 2165aae8a53SK.Prasad rcu_read_lock(); 2175aae8a53SK.Prasad 2185aae8a53SK.Prasad bp = __get_cpu_var(bp_per_reg); 2195aae8a53SK.Prasad if (!bp) 2205aae8a53SK.Prasad goto out; 2215aae8a53SK.Prasad info = counter_arch_bp(bp); 2225aae8a53SK.Prasad is_ptrace_bp = (bp->overflow_handler == ptrace_triggered) ? 2235aae8a53SK.Prasad true : false; 2245aae8a53SK.Prasad 2255aae8a53SK.Prasad /* 2265aae8a53SK.Prasad * Return early after invoking user-callback function without restoring 2275aae8a53SK.Prasad * DABR if the breakpoint is from ptrace which always operates in 2285aae8a53SK.Prasad * one-shot mode. The ptrace-ed process will receive the SIGTRAP signal 2295aae8a53SK.Prasad * generated in do_dabr(). 2305aae8a53SK.Prasad */ 2315aae8a53SK.Prasad if (is_ptrace_bp) { 2325aae8a53SK.Prasad perf_bp_event(bp, regs); 2335aae8a53SK.Prasad rc = NOTIFY_DONE; 2345aae8a53SK.Prasad goto out; 2355aae8a53SK.Prasad } 2365aae8a53SK.Prasad 2375aae8a53SK.Prasad /* Do not emulate user-space instructions, instead single-step them */ 2385aae8a53SK.Prasad if (user_mode(regs)) { 2395aae8a53SK.Prasad bp->ctx->task->thread.last_hit_ubp = bp; 2405aae8a53SK.Prasad regs->msr |= MSR_SE; 2415aae8a53SK.Prasad goto out; 2425aae8a53SK.Prasad } 2435aae8a53SK.Prasad 2445aae8a53SK.Prasad stepped = 0; 2455aae8a53SK.Prasad instr = 0; 2465aae8a53SK.Prasad if (!__get_user_inatomic(instr, (unsigned int *) regs->nip)) 2475aae8a53SK.Prasad stepped = emulate_step(regs, instr); 2485aae8a53SK.Prasad 2495aae8a53SK.Prasad /* 2505aae8a53SK.Prasad * emulate_step() could not execute it. We've failed in reliably 2515aae8a53SK.Prasad * handling the hw-breakpoint. Unregister it and throw a warning 2525aae8a53SK.Prasad * message to let the user know about it. 2535aae8a53SK.Prasad */ 2545aae8a53SK.Prasad if (!stepped) { 2555aae8a53SK.Prasad WARN(1, "Unable to handle hardware breakpoint. Breakpoint at " 2565aae8a53SK.Prasad "0x%lx will be disabled.", info->address); 2575aae8a53SK.Prasad perf_event_disable(bp); 2585aae8a53SK.Prasad goto out; 2595aae8a53SK.Prasad } 2605aae8a53SK.Prasad /* 2615aae8a53SK.Prasad * As a policy, the callback is invoked in a 'trigger-after-execute' 2625aae8a53SK.Prasad * fashion 2635aae8a53SK.Prasad */ 2645aae8a53SK.Prasad perf_bp_event(bp, regs); 2655aae8a53SK.Prasad 2665aae8a53SK.Prasad set_dabr(info->address | info->type | DABR_TRANSLATION); 2675aae8a53SK.Prasad out: 2685aae8a53SK.Prasad rcu_read_unlock(); 2695aae8a53SK.Prasad return rc; 2705aae8a53SK.Prasad } 2715aae8a53SK.Prasad 2725aae8a53SK.Prasad /* 2735aae8a53SK.Prasad * Handle single-step exceptions following a DABR hit. 2745aae8a53SK.Prasad */ 2755aae8a53SK.Prasad int __kprobes single_step_dabr_instruction(struct die_args *args) 2765aae8a53SK.Prasad { 2775aae8a53SK.Prasad struct pt_regs *regs = args->regs; 2785aae8a53SK.Prasad struct perf_event *bp = NULL; 2795aae8a53SK.Prasad struct arch_hw_breakpoint *bp_info; 2805aae8a53SK.Prasad 2815aae8a53SK.Prasad bp = current->thread.last_hit_ubp; 2825aae8a53SK.Prasad /* 2835aae8a53SK.Prasad * Check if we are single-stepping as a result of a 2845aae8a53SK.Prasad * previous HW Breakpoint exception 2855aae8a53SK.Prasad */ 2865aae8a53SK.Prasad if (!bp) 2875aae8a53SK.Prasad return NOTIFY_DONE; 2885aae8a53SK.Prasad 2895aae8a53SK.Prasad bp_info = counter_arch_bp(bp); 2905aae8a53SK.Prasad 2915aae8a53SK.Prasad /* 2925aae8a53SK.Prasad * We shall invoke the user-defined callback function in the single 2935aae8a53SK.Prasad * stepping handler to confirm to 'trigger-after-execute' semantics 2945aae8a53SK.Prasad */ 2955aae8a53SK.Prasad perf_bp_event(bp, regs); 2965aae8a53SK.Prasad 2975aae8a53SK.Prasad /* 2985aae8a53SK.Prasad * Do not disable MSR_SE if the process was already in 2995aae8a53SK.Prasad * single-stepping mode. 3005aae8a53SK.Prasad */ 3015aae8a53SK.Prasad if (!test_thread_flag(TIF_SINGLESTEP)) 3025aae8a53SK.Prasad regs->msr &= ~MSR_SE; 3035aae8a53SK.Prasad 3045aae8a53SK.Prasad set_dabr(bp_info->address | bp_info->type | DABR_TRANSLATION); 3055aae8a53SK.Prasad current->thread.last_hit_ubp = NULL; 3065aae8a53SK.Prasad return NOTIFY_STOP; 3075aae8a53SK.Prasad } 3085aae8a53SK.Prasad 3095aae8a53SK.Prasad /* 3105aae8a53SK.Prasad * Handle debug exception notifications. 3115aae8a53SK.Prasad */ 3125aae8a53SK.Prasad int __kprobes hw_breakpoint_exceptions_notify( 3135aae8a53SK.Prasad struct notifier_block *unused, unsigned long val, void *data) 3145aae8a53SK.Prasad { 3155aae8a53SK.Prasad int ret = NOTIFY_DONE; 3165aae8a53SK.Prasad 3175aae8a53SK.Prasad switch (val) { 3185aae8a53SK.Prasad case DIE_DABR_MATCH: 3195aae8a53SK.Prasad ret = hw_breakpoint_handler(data); 3205aae8a53SK.Prasad break; 3215aae8a53SK.Prasad case DIE_SSTEP: 3225aae8a53SK.Prasad ret = single_step_dabr_instruction(data); 3235aae8a53SK.Prasad break; 3245aae8a53SK.Prasad } 3255aae8a53SK.Prasad 3265aae8a53SK.Prasad return ret; 3275aae8a53SK.Prasad } 3285aae8a53SK.Prasad 3295aae8a53SK.Prasad /* 3305aae8a53SK.Prasad * Release the user breakpoints used by ptrace 3315aae8a53SK.Prasad */ 3325aae8a53SK.Prasad void flush_ptrace_hw_breakpoint(struct task_struct *tsk) 3335aae8a53SK.Prasad { 3345aae8a53SK.Prasad struct thread_struct *t = &tsk->thread; 3355aae8a53SK.Prasad 3365aae8a53SK.Prasad unregister_hw_breakpoint(t->ptrace_bps[0]); 3375aae8a53SK.Prasad t->ptrace_bps[0] = NULL; 3385aae8a53SK.Prasad } 3395aae8a53SK.Prasad 3405aae8a53SK.Prasad void hw_breakpoint_pmu_read(struct perf_event *bp) 3415aae8a53SK.Prasad { 3425aae8a53SK.Prasad /* TODO */ 3435aae8a53SK.Prasad } 344