Fedora kernel-2.6.17-1.2142_FC4 patched with stable patch-2.6.17.4-vs2.0.2-rc26.diff
[linux-2.6.git] / arch / i386 / kernel / kprobes.c
index 59ff9b4..38806f4 100644 (file)
  *             Rusty Russell).
  * 2004-July   Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
  *             interface to access function arguments.
+ * 2005-May    Hien Nguyen <hien@us.ibm.com>, Jim Keniston
+ *             <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
+ *             <prasanna@in.ibm.com> added function-return probes.
  */
 
 #include <linux/config.h>
 #include <linux/kprobes.h>
 #include <linux/ptrace.h>
-#include <linux/spinlock.h>
 #include <linux/preempt.h>
+#include <asm/cacheflush.h>
 #include <asm/kdebug.h>
 #include <asm/desc.h>
+#include <asm/uaccess.h>
 
-/* kprobe_status settings */
-#define KPROBE_HIT_ACTIVE      0x00000001
-#define KPROBE_HIT_SS          0x00000002
-
-static struct kprobe *current_kprobe;
-static unsigned long kprobe_status, kprobe_old_eflags, kprobe_saved_eflags;
-static struct pt_regs jprobe_saved_regs;
-static long *jprobe_saved_esp;
-/* copy of the kernel stack at the probe fire time */
-static kprobe_opcode_t jprobes_stack[MAX_STACK_SIZE];
 void jprobe_return_end(void);
 
+DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
+DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
+
+/* insert a jmp code */
+static __always_inline void set_jmp_op(void *from, void *to)
+{
+       struct __arch_jmp_op {
+               char op;
+               long raddr;
+       } __attribute__((packed)) *jop;
+       jop = (struct __arch_jmp_op *)from;
+       jop->raddr = (long)(to) - ((long)(from) + 5);
+       jop->op = RELATIVEJUMP_INSTRUCTION;
+}
+
+/*
+ * returns non-zero if opcodes can be boosted.
+ */
+static __always_inline int can_boost(kprobe_opcode_t opcode)
+{
+       switch (opcode & 0xf0 ) {
+       case 0x70:
+               return 0; /* can't boost conditional jump */
+       case 0x90:
+               /* can't boost call and pushf */
+               return opcode != 0x9a && opcode != 0x9c;
+       case 0xc0:
+               /* can't boost undefined opcodes and soft-interruptions */
+               return (0xc1 < opcode && opcode < 0xc6) ||
+                       (0xc7 < opcode && opcode < 0xcc) || opcode == 0xcf;
+       case 0xd0:
+               /* can boost AA* and XLAT */
+               return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
+       case 0xe0:
+               /* can boost in/out and (may be) jmps */
+               return (0xe3 < opcode && opcode != 0xe8);
+       case 0xf0:
+               /* clear and set flags can be boost */
+               return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
+       default:
+               /* currently, can't boost 2 bytes opcodes */
+               return opcode != 0x0f;
+       }
+}
+
+
 /*
  * returns non-zero if opcode modifies the interrupt flag.
  */
-static inline int is_IF_modifier(kprobe_opcode_t opcode)
+static int __kprobes is_IF_modifier(kprobe_opcode_t opcode)
 {
        switch (opcode) {
        case 0xfa:              /* cli */
@@ -60,27 +100,71 @@ static inline int is_IF_modifier(kprobe_opcode_t opcode)
        return 0;
 }
 
-int arch_prepare_kprobe(struct kprobe *p)
+int __kprobes arch_prepare_kprobe(struct kprobe *p)
 {
+       /* insn: must be on special executable page on i386. */
+       p->ainsn.insn = get_insn_slot();
+       if (!p->ainsn.insn)
+               return -ENOMEM;
+
+       memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
+       p->opcode = *p->addr;
+       if (can_boost(p->opcode)) {
+               p->ainsn.boostable = 0;
+       } else {
+               p->ainsn.boostable = -1;
+       }
        return 0;
 }
 
-void arch_copy_kprobe(struct kprobe *p)
+void __kprobes arch_arm_kprobe(struct kprobe *p)
 {
-       memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
+       *p->addr = BREAKPOINT_INSTRUCTION;
+       flush_icache_range((unsigned long) p->addr,
+                          (unsigned long) p->addr + sizeof(kprobe_opcode_t));
+}
+
+void __kprobes arch_disarm_kprobe(struct kprobe *p)
+{
+       *p->addr = p->opcode;
+       flush_icache_range((unsigned long) p->addr,
+                          (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 }
 
-void arch_remove_kprobe(struct kprobe *p)
+void __kprobes arch_remove_kprobe(struct kprobe *p)
 {
+       mutex_lock(&kprobe_mutex);
+       free_insn_slot(p->ainsn.insn);
+       mutex_unlock(&kprobe_mutex);
 }
 
-static inline void disarm_kprobe(struct kprobe *p, struct pt_regs *regs)
+static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 {
-       *p->addr = p->opcode;
-       regs->eip = (unsigned long)p->addr;
+       kcb->prev_kprobe.kp = kprobe_running();
+       kcb->prev_kprobe.status = kcb->kprobe_status;
+       kcb->prev_kprobe.old_eflags = kcb->kprobe_old_eflags;
+       kcb->prev_kprobe.saved_eflags = kcb->kprobe_saved_eflags;
+}
+
+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;
+       kcb->kprobe_old_eflags = kcb->prev_kprobe.old_eflags;
+       kcb->kprobe_saved_eflags = kcb->prev_kprobe.saved_eflags;
+}
+
+static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
+                               struct kprobe_ctlblk *kcb)
+{
+       __get_cpu_var(current_kprobe) = p;
+       kcb->kprobe_saved_eflags = kcb->kprobe_old_eflags
+               = (regs->eflags & (TF_MASK | IF_MASK));
+       if (is_IF_modifier(p->opcode))
+               kcb->kprobe_saved_eflags &= ~IF_MASK;
 }
 
-static inline void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
+static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 {
        regs->eflags |= TF_MASK;
        regs->eflags &= ~IF_MASK;
@@ -88,66 +172,95 @@ static inline void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
        if (p->opcode == BREAKPOINT_INSTRUCTION)
                regs->eip = (unsigned long)p->addr;
        else
-               regs->eip = (unsigned long)&p->ainsn.insn;
+               regs->eip = (unsigned long)p->ainsn.insn;
+}
+
+/* Called with kretprobe_lock held */
+void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
+                                     struct pt_regs *regs)
+{
+       unsigned long *sara = (unsigned long *)&regs->esp;
+        struct kretprobe_instance *ri;
+
+        if ((ri = get_free_rp_inst(rp)) != NULL) {
+                ri->rp = rp;
+                ri->task = current;
+               ri->ret_addr = (kprobe_opcode_t *) *sara;
+
+               /* Replace the return addr with trampoline addr */
+               *sara = (unsigned long) &kretprobe_trampoline;
+
+                add_rp_inst(ri);
+        } else {
+                rp->nmissed++;
+        }
 }
 
 /*
  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
  * remain disabled thorough out this function.
  */
-static int kprobe_handler(struct pt_regs *regs)
+static int __kprobes kprobe_handler(struct pt_regs *regs)
 {
        struct kprobe *p;
        int ret = 0;
-       kprobe_opcode_t *addr = NULL;
-       unsigned long *lp;
+       kprobe_opcode_t *addr;
+       struct kprobe_ctlblk *kcb;
+#ifdef CONFIG_PREEMPT
+       unsigned pre_preempt_count = preempt_count();
+#endif /* CONFIG_PREEMPT */
 
-       /* We're in an interrupt, but this is clear and BUG()-safe. */
-       preempt_disable();
-       /* Check if the application is using LDT entry for its code segment and
-        * calculate the address by reading the base address from the LDT entry.
+       addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
+
+       /*
+        * We don't want to be preempted for the entire
+        * duration of kprobe processing
         */
-       if ((regs->xcs & 4) && (current->mm)) {
-               lp = (unsigned long *) ((unsigned long)((regs->xcs >> 3) * 8)
-                                       + (char *) current->mm->context.ldt);
-               addr = (kprobe_opcode_t *) (get_desc_base(lp) + regs->eip -
-                                               sizeof(kprobe_opcode_t));
-       } else {
-               addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
-       }
+       preempt_disable();
+       kcb = get_kprobe_ctlblk();
+
        /* Check we're not actually recursing */
        if (kprobe_running()) {
-               /* We *are* holding lock here, so this is safe.
-                  Disarm the probe we just hit, and ignore it. */
                p = get_kprobe(addr);
                if (p) {
-                       if (kprobe_status == KPROBE_HIT_SS) {
+                       if (kcb->kprobe_status == KPROBE_HIT_SS &&
+                               *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
                                regs->eflags &= ~TF_MASK;
-                               regs->eflags |= kprobe_saved_eflags;
-                               unlock_kprobes();
+                               regs->eflags |= kcb->kprobe_saved_eflags;
                                goto no_kprobe;
                        }
-                       disarm_kprobe(p, regs);
-                       ret = 1;
+                       /* We have reentered the kprobe_handler(), since
+                        * another probe was hit while within the handler.
+                        * We here save the original kprobes variables and
+                        * just single step on the instruction of the new probe
+                        * without calling any user handlers.
+                        */
+                       save_previous_kprobe(kcb);
+                       set_current_kprobe(p, regs, kcb);
+                       kprobes_inc_nmissed_count(p);
+                       prepare_singlestep(p, regs);
+                       kcb->kprobe_status = KPROBE_REENTER;
+                       return 1;
                } else {
-                       p = current_kprobe;
+                       if (*addr != BREAKPOINT_INSTRUCTION) {
+                       /* The breakpoint instruction was removed by
+                        * another cpu right after we hit, no further
+                        * handling of this interrupt is appropriate
+                        */
+                               regs->eip -= sizeof(kprobe_opcode_t);
+                               ret = 1;
+                               goto no_kprobe;
+                       }
+                       p = __get_cpu_var(current_kprobe);
                        if (p->break_handler && p->break_handler(p, regs)) {
                                goto ss_probe;
                        }
                }
-               /* If it's not ours, can't be delete race, (we hold lock). */
                goto no_kprobe;
        }
 
-       lock_kprobes();
        p = get_kprobe(addr);
        if (!p) {
-               unlock_kprobes();
-               if (regs->eflags & VM_MASK) {
-                       /* We are in virtual-8086 mode. Return 0 */
-                       goto no_kprobe;
-               }
-
                if (*addr != BREAKPOINT_INSTRUCTION) {
                        /*
                         * The breakpoint instruction was removed right
@@ -155,27 +268,41 @@ static int kprobe_handler(struct pt_regs *regs)
                         * either a probepoint or a debugger breakpoint
                         * at this address.  In either case, no further
                         * handling of this interrupt is appropriate.
+                        * Back up over the (now missing) int3 and run
+                        * the original instruction.
                         */
+                       regs->eip -= sizeof(kprobe_opcode_t);
                        ret = 1;
                }
                /* Not one of ours: let kernel handle it */
                goto no_kprobe;
        }
 
-       kprobe_status = KPROBE_HIT_ACTIVE;
-       current_kprobe = p;
-       kprobe_saved_eflags = kprobe_old_eflags
-           = (regs->eflags & (TF_MASK | IF_MASK));
-       if (is_IF_modifier(p->opcode))
-               kprobe_saved_eflags &= ~IF_MASK;
+       set_current_kprobe(p, regs, kcb);
+       kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 
        if (p->pre_handler && p->pre_handler(p, regs))
                /* handler has already set things up, so skip ss setup */
                return 1;
 
+       if (p->ainsn.boostable == 1 &&
+#ifdef CONFIG_PREEMPT
+           !(pre_preempt_count) && /*
+                                      * This enables booster when the direct
+                                      * execution path aren't preempted.
+                                      */
+#endif /* CONFIG_PREEMPT */
+           !p->post_handler && !p->break_handler ) {
+               /* Boost up -- we can execute copied instructions directly */
+               reset_current_kprobe();
+               regs->eip = (unsigned long)p->ainsn.insn;
+               preempt_enable_no_resched();
+               return 1;
+       }
+
 ss_probe:
        prepare_singlestep(p, regs);
-       kprobe_status = KPROBE_HIT_SS;
+       kcb->kprobe_status = KPROBE_HIT_SS;
        return 1;
 
 no_kprobe:
@@ -183,6 +310,102 @@ no_kprobe:
        return ret;
 }
 
+/*
+ * For function-return probes, init_kprobes() establishes a probepoint
+ * here. When a retprobed function returns, this probe is hit and
+ * trampoline_probe_handler() runs, calling the kretprobe's handler.
+ */
+ void __kprobes kretprobe_trampoline_holder(void)
+ {
+       asm volatile ( ".global kretprobe_trampoline\n"
+                       "kretprobe_trampoline: \n"
+                       "       pushf\n"
+                       /* skip cs, eip, orig_eax, es, ds */
+                       "       subl $20, %esp\n"
+                       "       pushl %eax\n"
+                       "       pushl %ebp\n"
+                       "       pushl %edi\n"
+                       "       pushl %esi\n"
+                       "       pushl %edx\n"
+                       "       pushl %ecx\n"
+                       "       pushl %ebx\n"
+                       "       movl %esp, %eax\n"
+                       "       call trampoline_handler\n"
+                       /* move eflags to cs */
+                       "       movl 48(%esp), %edx\n"
+                       "       movl %edx, 44(%esp)\n"
+                       /* save true return address on eflags */
+                       "       movl %eax, 48(%esp)\n"
+                       "       popl %ebx\n"
+                       "       popl %ecx\n"
+                       "       popl %edx\n"
+                       "       popl %esi\n"
+                       "       popl %edi\n"
+                       "       popl %ebp\n"
+                       "       popl %eax\n"
+                       /* skip eip, orig_eax, es, ds */
+                       "       addl $16, %esp\n"
+                       "       popf\n"
+                       "       ret\n");
+}
+
+/*
+ * Called from kretprobe_trampoline
+ */
+fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
+{
+        struct kretprobe_instance *ri = NULL;
+        struct hlist_head *head;
+        struct hlist_node *node, *tmp;
+       unsigned long flags, orig_ret_address = 0;
+       unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
+
+       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 an multiple functions in the call path
+        * have a return probe installed on them, and/or more then one return
+        * 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;
+                       ri->rp->handler(ri, regs);
+                       __get_cpu_var(current_kprobe) = NULL;
+               }
+
+               orig_ret_address = (unsigned long)ri->ret_addr;
+               recycle_rp_inst(ri);
+
+               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;
+       }
+
+       BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
+
+       spin_unlock_irqrestore(&kretprobe_lock, flags);
+
+       return (void*)orig_ret_address;
+}
+
 /*
  * Called after single-stepping.  p->addr is the address of the
  * instruction whose first byte has been replaced by the "int 3"
@@ -204,72 +427,104 @@ no_kprobe:
  * 2) If the single-stepped instruction was a call, the return address
  * that is atop the stack is the address following the copied instruction.
  * We need to make it the address following the original instruction.
+ *
+ * This function also checks instruction size for preparing direct execution.
  */
-static void resume_execution(struct kprobe *p, struct pt_regs *regs)
+static void __kprobes resume_execution(struct kprobe *p,
+               struct pt_regs *regs, struct kprobe_ctlblk *kcb)
 {
        unsigned long *tos = (unsigned long *)&regs->esp;
-       unsigned long next_eip = 0;
-       unsigned long copy_eip = (unsigned long)&p->ainsn.insn;
+       unsigned long copy_eip = (unsigned long)p->ainsn.insn;
        unsigned long orig_eip = (unsigned long)p->addr;
 
+       regs->eflags &= ~TF_MASK;
        switch (p->ainsn.insn[0]) {
        case 0x9c:              /* pushfl */
                *tos &= ~(TF_MASK | IF_MASK);
-               *tos |= kprobe_old_eflags;
+               *tos |= kcb->kprobe_old_eflags;
                break;
-       case 0xc3:              /* ret/lret */
-       case 0xcb:
-       case 0xc2:
+       case 0xc2:              /* iret/ret/lret */
+       case 0xc3:
        case 0xca:
-               regs->eflags &= ~TF_MASK;
-               /* eip is already adjusted, no more changes required*/
-               return;
+       case 0xcb:
+       case 0xcf:
+       case 0xea:              /* jmp absolute -- eip is correct */
+               /* eip is already adjusted, no more changes required */
+               p->ainsn.boostable = 1;
+               goto no_change;
        case 0xe8:              /* call relative - Fix return addr */
                *tos = orig_eip + (*tos - copy_eip);
                break;
+       case 0x9a:              /* call absolute -- same as call absolute, indirect */
+               *tos = orig_eip + (*tos - copy_eip);
+               goto no_change;
        case 0xff:
                if ((p->ainsn.insn[1] & 0x30) == 0x10) {
-                       /* call absolute, indirect */
-                       /* Fix return addr; eip is correct. */
-                       next_eip = regs->eip;
+                       /*
+                        * call absolute, indirect
+                        * Fix return addr; eip is correct.
+                        * But this is not boostable
+                        */
                        *tos = orig_eip + (*tos - copy_eip);
+                       goto no_change;
                } else if (((p->ainsn.insn[1] & 0x31) == 0x20) ||       /* jmp near, absolute indirect */
                           ((p->ainsn.insn[1] & 0x31) == 0x21)) {       /* jmp far, absolute indirect */
-                       /* eip is correct. */
-                       next_eip = regs->eip;
+                       /* eip is correct. And this is boostable */
+                       p->ainsn.boostable = 1;
+                       goto no_change;
                }
-               break;
-       case 0xea:              /* jmp absolute -- eip is correct */
-               next_eip = regs->eip;
-               break;
        default:
                break;
        }
 
-       regs->eflags &= ~TF_MASK;
-       if (next_eip) {
-               regs->eip = next_eip;
-       } else {
-               regs->eip = orig_eip + (regs->eip - copy_eip);
+       if (p->ainsn.boostable == 0) {
+               if ((regs->eip > copy_eip) &&
+                   (regs->eip - copy_eip) + 5 < MAX_INSN_SIZE) {
+                       /*
+                        * These instructions can be executed directly if it
+                        * jumps back to correct address.
+                        */
+                       set_jmp_op((void *)regs->eip,
+                                  (void *)orig_eip + (regs->eip - copy_eip));
+                       p->ainsn.boostable = 1;
+               } else {
+                       p->ainsn.boostable = -1;
+               }
        }
+
+       regs->eip = orig_eip + (regs->eip - copy_eip);
+
+no_change:
+       return;
 }
 
 /*
  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
- * remain disabled thoroughout this function.  And we hold kprobe lock.
+ * remain disabled thoroughout this function.
  */
-static inline int post_kprobe_handler(struct pt_regs *regs)
+static int __kprobes post_kprobe_handler(struct pt_regs *regs)
 {
-       if (!kprobe_running())
+       struct kprobe *cur = kprobe_running();
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
+       if (!cur)
                return 0;
 
-       if (current_kprobe->post_handler)
-               current_kprobe->post_handler(current_kprobe, regs, 0);
+       if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
+               kcb->kprobe_status = KPROBE_HIT_SSDONE;
+               cur->post_handler(cur, regs, 0);
+       }
 
-       resume_execution(current_kprobe, regs);
-       regs->eflags |= kprobe_saved_eflags;
+       resume_execution(cur, regs, kcb);
+       regs->eflags |= kcb->kprobe_saved_eflags;
 
-       unlock_kprobes();
+       /*Restore back the original saved kprobes variables and continue. */
+       if (kcb->kprobe_status == KPROBE_REENTER) {
+               restore_previous_kprobe(kcb);
+               goto out;
+       }
+       reset_current_kprobe();
+out:
        preempt_enable_no_resched();
 
        /*
@@ -283,19 +538,62 @@ static inline int post_kprobe_handler(struct pt_regs *regs)
        return 1;
 }
 
-/* Interrupts disabled, kprobe_lock held. */
-static inline int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
+static int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 {
-       if (current_kprobe->fault_handler
-           && current_kprobe->fault_handler(current_kprobe, regs, trapnr))
-               return 1;
-
-       if (kprobe_status & KPROBE_HIT_SS) {
-               resume_execution(current_kprobe, regs);
-               regs->eflags |= kprobe_old_eflags;
-
-               unlock_kprobes();
+       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 eip points back to the probe address
+                * and allow the page fault handler to continue as a
+                * normal page fault.
+                */
+               regs->eip = (unsigned long)cur->addr;
+               regs->eflags |= kcb->kprobe_old_eflags;
+               if (kcb->kprobe_status == KPROBE_REENTER)
+                       restore_previous_kprobe(kcb);
+               else
+                       reset_current_kprobe();
                preempt_enable_no_resched();
+               break;
+       case KPROBE_HIT_ACTIVE:
+       case KPROBE_HIT_SSDONE:
+               /*
+                * We increment the nmissed count for accounting,
+                * we can also use npre/npostfault count for accouting
+                * 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 first.
+                */
+               if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
+                       return 1;
+
+               /*
+                * In case the user-specified fault handler returned
+                * zero, try to fix up.
+                */
+               if (fixup_exception(regs))
+                       return 1;
+
+               /*
+                * fixup_exception() could not handle it,
+                * Let do_page_fault() fix it.
+                */
+               break;
+       default:
+               break;
        }
        return 0;
 }
@@ -303,43 +601,48 @@ static inline int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 /*
  * Wrapper routine to for handling exceptions.
  */
-int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
-                            void *data)
+int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
+                                      unsigned long val, void *data)
 {
        struct die_args *args = (struct die_args *)data;
+       int ret = NOTIFY_DONE;
+
+       if (args->regs && user_mode(args->regs))
+               return ret;
+
        switch (val) {
        case DIE_INT3:
                if (kprobe_handler(args->regs))
-                       return NOTIFY_STOP;
+                       ret = NOTIFY_STOP;
                break;
        case DIE_DEBUG:
                if (post_kprobe_handler(args->regs))
-                       return NOTIFY_STOP;
+                       ret = NOTIFY_STOP;
                break;
        case DIE_GPF:
-               if (kprobe_running() &&
-                   kprobe_fault_handler(args->regs, args->trapnr))
-                       return NOTIFY_STOP;
-               break;
        case DIE_PAGE_FAULT:
+               /* kprobe_running() needs smp_processor_id() */
+               preempt_disable();
                if (kprobe_running() &&
                    kprobe_fault_handler(args->regs, args->trapnr))
-                       return NOTIFY_STOP;
+                       ret = NOTIFY_STOP;
+               preempt_enable();
                break;
        default:
                break;
        }
-       return NOTIFY_DONE;
+       return ret;
 }
 
-int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
+int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 {
        struct jprobe *jp = container_of(p, struct jprobe, kp);
        unsigned long addr;
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 
-       jprobe_saved_regs = *regs;
-       jprobe_saved_esp = &regs->esp;
-       addr = (unsigned long)jprobe_saved_esp;
+       kcb->jprobe_saved_regs = *regs;
+       kcb->jprobe_saved_esp = &regs->esp;
+       addr = (unsigned long)(kcb->jprobe_saved_esp);
 
        /*
         * TBD: As Linus pointed out, gcc assumes that the callee
@@ -348,45 +651,55 @@ int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
         * we also save and restore enough stack bytes to cover
         * the argument area.
         */
-       memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr));
+       memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
+                       MIN_STACK_SIZE(addr));
        regs->eflags &= ~IF_MASK;
        regs->eip = (unsigned long)(jp->entry);
        return 1;
 }
 
-void jprobe_return(void)
+void __kprobes jprobe_return(void)
 {
-       preempt_enable_no_resched();
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+
        asm volatile ("       xchgl   %%ebx,%%esp     \n"
                      "       int3                      \n"
                      "       .globl jprobe_return_end  \n"
                      "       jprobe_return_end:        \n"
                      "       nop                       \n"::"b"
-                     (jprobe_saved_esp):"memory");
+                     (kcb->jprobe_saved_esp):"memory");
 }
 
-int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
+int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 {
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
        u8 *addr = (u8 *) (regs->eip - 1);
-       unsigned long stack_addr = (unsigned long)jprobe_saved_esp;
+       unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_esp);
        struct jprobe *jp = container_of(p, struct jprobe, kp);
 
        if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
-               if (&regs->esp != jprobe_saved_esp) {
+               if (&regs->esp != kcb->jprobe_saved_esp) {
                        struct pt_regs *saved_regs =
-                           container_of(jprobe_saved_esp, struct pt_regs, esp);
+                           container_of(kcb->jprobe_saved_esp,
+                                           struct pt_regs, esp);
                        printk("current esp %p does not match saved esp %p\n",
-                              &regs->esp, jprobe_saved_esp);
+                              &regs->esp, kcb->jprobe_saved_esp);
                        printk("Saved registers for jprobe %p\n", jp);
                        show_registers(saved_regs);
                        printk("Current registers\n");
                        show_registers(regs);
                        BUG();
                }
-               *regs = jprobe_saved_regs;
-               memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack,
+               *regs = kcb->jprobe_saved_regs;
+               memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
                       MIN_STACK_SIZE(stack_addr));
+               preempt_enable_no_resched();
                return 1;
        }
        return 0;
 }
+
+int __init arch_init_kprobes(void)
+{
+       return 0;
+}