]> err.no Git - linux-2.6/commitdiff
ARM kprobes: core code
authorAbhishek Sagar <sagar.abhishek@gmail.com>
Mon, 11 Jun 2007 22:20:10 +0000 (22:20 +0000)
committerRussell King <rmk+kernel@arm.linux.org.uk>
Sat, 26 Jan 2008 15:25:16 +0000 (15:25 +0000)
This is a full implementation of Kprobes including Jprobes and
Kretprobes support.

This ARM implementation does not follow the usual kprobes double-
exception model. The traditional model is where the initial kprobes
breakpoint calls kprobe_handler(), which returns from exception to
execute the instruction in its original context, then immediately
re-enters after a second breakpoint (or single-stepping exception)
into post_kprobe_handler(), each time the probe is hit..  The ARM
implementation only executes one kprobes exception per hit, so no
post_kprobe_handler() phase. All side-effects from the kprobe'd
instruction are resolved before returning from the initial exception.
As a result, all instructions are _always_ effectively boosted
regardless of the type of instruction, and even regardless of whether
or not there is a post-handler for the probe.

Signed-off-by: Abhishek Sagar <sagar.abhishek@gmail.com>
Signed-off-by: Quentin Barnes <qbarnes@gmail.com>
Signed-off-by: Nicolas Pitre <nico@marvell.com>
arch/arm/kernel/Makefile
arch/arm/kernel/kprobes.c [new file with mode: 0644]
include/asm-arm/kprobes.h

index 9e0eebacd3d961246cbea82b4c12aa9b01e25e6c..faa7619211534c541f61e370985cf1480f247f52 100644 (file)
@@ -19,7 +19,7 @@ obj-$(CONFIG_ISA_DMA)         += dma-isa.o
 obj-$(CONFIG_PCI)              += bios32.o isa.o
 obj-$(CONFIG_SMP)              += smp.o
 obj-$(CONFIG_KEXEC)            += machine_kexec.o relocate_kernel.o
-obj-$(CONFIG_KPROBES)          += kprobes-decode.o
+obj-$(CONFIG_KPROBES)          += kprobes.o kprobes-decode.o
 obj-$(CONFIG_OABI_COMPAT)      += sys_oabi-compat.o
 
 obj-$(CONFIG_CRUNCH)           += crunch.o crunch-bits.o
diff --git a/arch/arm/kernel/kprobes.c b/arch/arm/kernel/kprobes.c
new file mode 100644 (file)
index 0000000..a18a845
--- /dev/null
@@ -0,0 +1,453 @@
+/*
+ * arch/arm/kernel/kprobes.c
+ *
+ * Kprobes on ARM
+ *
+ * Abhishek Sagar <sagar.abhishek@gmail.com>
+ * Copyright (C) 2006, 2007 Motorola Inc.
+ *
+ * Nicolas Pitre <nico@marvell.com>
+ * Copyright (C) 2007 Marvell Ltd.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/kprobes.h>
+#include <linux/module.h>
+#include <linux/stringify.h>
+#include <asm/traps.h>
+#include <asm/cacheflush.h>
+
+/*
+ * This undefined instruction must be unique and
+ * reserved solely for kprobes' use.
+ */
+#define KPROBE_BREAKPOINT_INSTRUCTION  0xe7f001f8
+
+#define MIN_STACK_SIZE(addr)                           \
+       min((unsigned long)MAX_STACK_SIZE,              \
+           (unsigned long)current_thread_info() + THREAD_START_SP - (addr))
+
+#define flush_insns(addr, cnt)                                 \
+       flush_icache_range((unsigned long)(addr),       \
+                          (unsigned long)(addr) +      \
+                          sizeof(kprobe_opcode_t) * (cnt))
+
+/* Used as a marker in ARM_pc to note when we're in a jprobe. */
+#define JPROBE_MAGIC_ADDR              0xffffffff
+
+DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
+DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
+
+
+int __kprobes arch_prepare_kprobe(struct kprobe *p)
+{
+       kprobe_opcode_t insn;
+       kprobe_opcode_t tmp_insn[MAX_INSN_SIZE];
+       unsigned long addr = (unsigned long)p->addr;
+       int is;
+
+       if (addr & 0x3)
+               return -EINVAL;
+
+       insn = *p->addr;
+       p->opcode = insn;
+       p->ainsn.insn = tmp_insn;
+
+       switch (arm_kprobe_decode_insn(insn, &p->ainsn)) {
+       case INSN_REJECTED:     /* not supported */
+               return -EINVAL;
+
+       case INSN_GOOD:         /* instruction uses slot */
+               p->ainsn.insn = get_insn_slot();
+               if (!p->ainsn.insn)
+                       return -ENOMEM;
+               for (is = 0; is < MAX_INSN_SIZE; ++is)
+                       p->ainsn.insn[is] = tmp_insn[is];
+               flush_insns(&p->ainsn.insn, MAX_INSN_SIZE);
+               break;
+
+       case INSN_GOOD_NO_SLOT: /* instruction doesn't need insn slot */
+               p->ainsn.insn = NULL;
+               break;
+       }
+
+       return 0;
+}
+
+void __kprobes arch_arm_kprobe(struct kprobe *p)
+{
+       *p->addr = KPROBE_BREAKPOINT_INSTRUCTION;
+       flush_insns(p->addr, 1);
+}
+
+void __kprobes arch_disarm_kprobe(struct kprobe *p)
+{
+       *p->addr = p->opcode;
+       flush_insns(p->addr, 1);
+}
+
+void __kprobes arch_remove_kprobe(struct kprobe *p)
+{
+       if (p->ainsn.insn) {
+               mutex_lock(&kprobe_mutex);
+               free_insn_slot(p->ainsn.insn, 0);
+               mutex_unlock(&kprobe_mutex);
+               p->ainsn.insn = NULL;
+       }
+}
+
+static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
+{
+       kcb->prev_kprobe.kp = kprobe_running();
+       kcb->prev_kprobe.status = kcb->kprobe_status;
+}
+
+static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
+{
+       __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
+       kcb->kprobe_status = kcb->prev_kprobe.status;
+}
+
+static void __kprobes set_current_kprobe(struct kprobe *p)
+{
+       __get_cpu_var(current_kprobe) = p;
+}
+
+static void __kprobes singlestep(struct kprobe *p, struct pt_regs *regs,
+                                struct kprobe_ctlblk *kcb)
+{
+       regs->ARM_pc += 4;
+       p->ainsn.insn_handler(p, regs);
+}
+
+/*
+ * Called with IRQs disabled. IRQs must remain disabled from that point
+ * all the way until processing this kprobe is complete.  The current
+ * kprobes implementation cannot process more than one nested level of
+ * kprobe, and that level is reserved for user kprobe handlers, so we can't
+ * risk encountering a new kprobe in an interrupt handler.
+ */
+void __kprobes kprobe_handler(struct pt_regs *regs)
+{
+       struct kprobe *p, *cur;
+       struct kprobe_ctlblk *kcb;
+       kprobe_opcode_t *addr = (kprobe_opcode_t *)regs->ARM_pc;
+
+       kcb = get_kprobe_ctlblk();
+       cur = kprobe_running();
+       p = get_kprobe(addr);
+
+       if (p) {
+               if (cur) {
+                       /* Kprobe is pending, so we're recursing. */
+                       switch (kcb->kprobe_status) {
+                       case KPROBE_HIT_ACTIVE:
+                       case KPROBE_HIT_SSDONE:
+                               /* A pre- or post-handler probe got us here. */
+                               kprobes_inc_nmissed_count(p);
+                               save_previous_kprobe(kcb);
+                               set_current_kprobe(p);
+                               kcb->kprobe_status = KPROBE_REENTER;
+                               singlestep(p, regs, kcb);
+                               restore_previous_kprobe(kcb);
+                               break;
+                       default:
+                               /* impossible cases */
+                               BUG();
+                       }
+               } else {
+                       set_current_kprobe(p);
+                       kcb->kprobe_status = KPROBE_HIT_ACTIVE;
+
+                       /*
+                        * If we have no pre-handler or it returned 0, we
+                        * continue with normal processing.  If we have a
+                        * pre-handler and it returned non-zero, it prepped
+                        * for calling the break_handler below on re-entry,
+                        * so get out doing nothing more here.
+                        */
+                       if (!p->pre_handler || !p->pre_handler(p, regs)) {
+                               kcb->kprobe_status = KPROBE_HIT_SS;
+                               singlestep(p, regs, kcb);
+                               if (p->post_handler) {
+                                       kcb->kprobe_status = KPROBE_HIT_SSDONE;
+                                       p->post_handler(p, regs, 0);
+                               }
+                               reset_current_kprobe();
+                       }
+               }
+       } else if (cur) {
+               /* We probably hit a jprobe.  Call its break handler. */
+               if (cur->break_handler && cur->break_handler(cur, regs)) {
+                       kcb->kprobe_status = KPROBE_HIT_SS;
+                       singlestep(cur, regs, kcb);
+                       if (cur->post_handler) {
+                               kcb->kprobe_status = KPROBE_HIT_SSDONE;
+                               cur->post_handler(cur, regs, 0);
+                       }
+               }
+               reset_current_kprobe();
+       } else {
+               /*
+                * The probe was removed and a race is in progress.
+                * There is nothing we can do about it.  Let's restart
+                * the instruction.  By the time we can restart, the
+                * real instruction will be there.
+                */
+       }
+}
+
+static int kprobe_trap_handler(struct pt_regs *regs, unsigned int instr)
+{
+       kprobe_handler(regs);
+       return 0;
+}
+
+int __kprobes kprobe_fault_handler(struct pt_regs *regs, unsigned int fsr)
+{
+       struct kprobe *cur = kprobe_running();
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+       switch (kcb->kprobe_status) {
+       case KPROBE_HIT_SS:
+       case KPROBE_REENTER:
+               /*
+                * We are here because the instruction being single
+                * stepped caused a page fault. We reset the current
+                * kprobe and the PC to point back to the probe address
+                * and allow the page fault handler to continue as a
+                * normal page fault.
+                */
+               regs->ARM_pc = (long)cur->addr;
+               if (kcb->kprobe_status == KPROBE_REENTER) {
+                       restore_previous_kprobe(kcb);
+               } else {
+                       reset_current_kprobe();
+               }
+               break;
+
+       case KPROBE_HIT_ACTIVE:
+       case KPROBE_HIT_SSDONE:
+               /*
+                * We increment the nmissed count for accounting,
+                * we can also use npre/npostfault count for accounting
+                * these specific fault cases.
+                */
+               kprobes_inc_nmissed_count(cur);
+
+               /*
+                * We come here because instructions in the pre/post
+                * handler caused the page_fault, this could happen
+                * if handler tries to access user space by
+                * copy_from_user(), get_user() etc. Let the
+                * user-specified handler try to fix it.
+                */
+               if (cur->fault_handler && cur->fault_handler(cur, regs, fsr))
+                       return 1;
+               break;
+
+       default:
+               break;
+       }
+
+       return 0;
+}
+
+int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
+                                      unsigned long val, void *data)
+{
+       /*
+        * notify_die() is currently never called on ARM,
+        * so this callback is currently empty.
+        */
+       return NOTIFY_DONE;
+}
+
+/*
+ * When a retprobed function returns, trampoline_handler() is called,
+ * calling the kretprobe's handler. We construct a struct pt_regs to
+ * give a view of registers r0-r11 to the user return-handler.  This is
+ * not a complete pt_regs structure, but that should be plenty sufficient
+ * for kretprobe handlers which should normally be interested in r0 only
+ * anyway.
+ */
+static void __attribute__((naked)) __kprobes kretprobe_trampoline(void)
+{
+       __asm__ __volatile__ (
+               "stmdb  sp!, {r0 - r11}         \n\t"
+               "mov    r0, sp                  \n\t"
+               "bl     trampoline_handler      \n\t"
+               "mov    lr, r0                  \n\t"
+               "ldmia  sp!, {r0 - r11}         \n\t"
+               "mov    pc, lr                  \n\t"
+               : : : "memory");
+}
+
+/* Called from kretprobe_trampoline */
+static __used __kprobes void *trampoline_handler(struct pt_regs *regs)
+{
+       struct kretprobe_instance *ri = NULL;
+       struct hlist_head *head, empty_rp;
+       struct hlist_node *node, *tmp;
+       unsigned long flags, orig_ret_address = 0;
+       unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
+
+       INIT_HLIST_HEAD(&empty_rp);
+       spin_lock_irqsave(&kretprobe_lock, flags);
+       head = kretprobe_inst_table_head(current);
+
+       /*
+        * It is possible to have multiple instances associated with a given
+        * task either because multiple functions in the call path have
+        * a return probe installed on them, and/or more than one return
+        * probe was registered for a target function.
+        *
+        * We can handle this because:
+        *     - instances are always inserted at the head of the list
+        *     - when multiple return probes are registered for the same
+        *       function, the first instance's ret_addr will point to the
+        *       real return address, and all the rest will point to
+        *       kretprobe_trampoline
+        */
+       hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
+               if (ri->task != current)
+                       /* another task is sharing our hash bucket */
+                       continue;
+
+               if (ri->rp && ri->rp->handler) {
+                       __get_cpu_var(current_kprobe) = &ri->rp->kp;
+                       get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
+                       ri->rp->handler(ri, regs);
+                       __get_cpu_var(current_kprobe) = NULL;
+               }
+
+               orig_ret_address = (unsigned long)ri->ret_addr;
+               recycle_rp_inst(ri, &empty_rp);
+
+               if (orig_ret_address != trampoline_address)
+                       /*
+                        * This is the real return address. Any other
+                        * instances associated with this task are for
+                        * other calls deeper on the call stack
+                        */
+                       break;
+       }
+
+       kretprobe_assert(ri, orig_ret_address, trampoline_address);
+       spin_unlock_irqrestore(&kretprobe_lock, flags);
+
+       hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
+               hlist_del(&ri->hlist);
+               kfree(ri);
+       }
+
+       return (void *)orig_ret_address;
+}
+
+/* Called with kretprobe_lock held. */
+void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
+                                     struct pt_regs *regs)
+{
+       ri->ret_addr = (kprobe_opcode_t *)regs->ARM_lr;
+
+       /* Replace the return addr with trampoline addr. */
+       regs->ARM_lr = (unsigned long)&kretprobe_trampoline;
+}
+
+int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
+{
+       struct jprobe *jp = container_of(p, struct jprobe, kp);
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+       long sp_addr = regs->ARM_sp;
+
+       kcb->jprobe_saved_regs = *regs;
+       memcpy(kcb->jprobes_stack, (void *)sp_addr, MIN_STACK_SIZE(sp_addr));
+       regs->ARM_pc = (long)jp->entry;
+       regs->ARM_cpsr |= PSR_I_BIT;
+       preempt_disable();
+       return 1;
+}
+
+void __kprobes jprobe_return(void)
+{
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+       __asm__ __volatile__ (
+               /*
+                * Setup an empty pt_regs. Fill SP and PC fields as
+                * they're needed by longjmp_break_handler.
+                */
+               "sub    sp, %0, %1              \n\t"
+               "ldr    r0, ="__stringify(JPROBE_MAGIC_ADDR)"\n\t"
+               "str    %0, [sp, %2]            \n\t"
+               "str    r0, [sp, %3]            \n\t"
+               "mov    r0, sp                  \n\t"
+               "bl     kprobe_handler          \n\t"
+
+               /*
+                * Return to the context saved by setjmp_pre_handler
+                * and restored by longjmp_break_handler.
+                */
+               "ldr    r0, [sp, %4]            \n\t"
+               "msr    cpsr_cxsf, r0           \n\t"
+               "ldmia  sp, {r0 - pc}           \n\t"
+               :
+               : "r" (kcb->jprobe_saved_regs.ARM_sp),
+                 "I" (sizeof(struct pt_regs)),
+                 "J" (offsetof(struct pt_regs, ARM_sp)),
+                 "J" (offsetof(struct pt_regs, ARM_pc)),
+                 "J" (offsetof(struct pt_regs, ARM_cpsr))
+               : "memory", "cc");
+}
+
+int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
+{
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+       long stack_addr = kcb->jprobe_saved_regs.ARM_sp;
+       long orig_sp = regs->ARM_sp;
+       struct jprobe *jp = container_of(p, struct jprobe, kp);
+
+       if (regs->ARM_pc == JPROBE_MAGIC_ADDR) {
+               if (orig_sp != stack_addr) {
+                       struct pt_regs *saved_regs =
+                               (struct pt_regs *)kcb->jprobe_saved_regs.ARM_sp;
+                       printk("current sp %lx does not match saved sp %lx\n",
+                              orig_sp, stack_addr);
+                       printk("Saved registers for jprobe %p\n", jp);
+                       show_regs(saved_regs);
+                       printk("Current registers\n");
+                       show_regs(regs);
+                       BUG();
+               }
+               *regs = kcb->jprobe_saved_regs;
+               memcpy((void *)stack_addr, kcb->jprobes_stack,
+                      MIN_STACK_SIZE(stack_addr));
+               preempt_enable_no_resched();
+               return 1;
+       }
+       return 0;
+}
+
+static struct undef_hook kprobes_break_hook = {
+       .instr_mask     = 0xffffffff,
+       .instr_val      = KPROBE_BREAKPOINT_INSTRUCTION,
+       .cpsr_mask      = MODE_MASK,
+       .cpsr_val       = SVC_MODE,
+       .fn             = kprobe_trap_handler,
+};
+
+int __init arch_init_kprobes()
+{
+       arm_kprobe_decode_init();
+       register_undef_hook(&kprobes_break_hook);
+       return 0;
+}
index 951322328793e13613b771dc507bd9032a7c2b19..273f37413ee6fd9cd2921775d6803d434a3fabaa 100644 (file)
 
 #include <linux/types.h>
 #include <linux/ptrace.h>
+#include <linux/percpu.h>
+
+#define ARCH_SUPPORTS_KRETPROBES
+#define __ARCH_WANT_KPROBES_INSN_SLOT
+#define MAX_INSN_SIZE                  2
+#define MAX_STACK_SIZE                 64      /* 32 would probably be OK */
+
+#define regs_return_value(regs)                ((regs)->ARM_r0)
+#define flush_insn_slot(p)             do { } while (0)
+#define kretprobe_blacklist_size       0
 
 typedef u32 kprobe_opcode_t;
 
@@ -30,6 +40,25 @@ struct arch_specific_insn {
        kprobe_insn_handler_t   *insn_handler;
 };
 
+struct prev_kprobe {
+       struct kprobe *kp;
+       unsigned int status;
+};
+
+/* per-cpu kprobe control block */
+struct kprobe_ctlblk {
+       unsigned int kprobe_status;
+       struct prev_kprobe prev_kprobe;
+       struct pt_regs jprobe_saved_regs;
+       char jprobes_stack[MAX_STACK_SIZE];
+};
+
+void arch_remove_kprobe(struct kprobe *);
+
+int kprobe_fault_handler(struct pt_regs *regs, unsigned int fsr);
+int kprobe_exceptions_notify(struct notifier_block *self,
+                            unsigned long val, void *data);
+
 enum kprobe_insn {
        INSN_REJECTED,
        INSN_GOOD,