linux 2.6.16.38 w/ vs2.0.3-rc1
[linux-2.6.git] / arch / i386 / kernel / kprobes.c
index 38806f4..694a139 100644 (file)
 #include <asm/cacheflush.h>
 #include <asm/kdebug.h>
 #include <asm/desc.h>
-#include <asm/uaccess.h>
 
 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 int __kprobes is_IF_modifier(kprobe_opcode_t opcode)
+static inline int is_IF_modifier(kprobe_opcode_t opcode)
 {
        switch (opcode) {
        case 0xfa:              /* cli */
@@ -109,11 +65,6 @@ int __kprobes arch_prepare_kprobe(struct kprobe *p)
 
        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;
 }
 
@@ -133,12 +84,12 @@ void __kprobes arch_disarm_kprobe(struct kprobe *p)
 
 void __kprobes arch_remove_kprobe(struct kprobe *p)
 {
-       mutex_lock(&kprobe_mutex);
+       down(&kprobe_mutex);
        free_insn_slot(p->ainsn.insn);
-       mutex_unlock(&kprobe_mutex);
+       up(&kprobe_mutex);
 }
 
-static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
+static inline void save_previous_kprobe(struct kprobe_ctlblk *kcb)
 {
        kcb->prev_kprobe.kp = kprobe_running();
        kcb->prev_kprobe.status = kcb->kprobe_status;
@@ -146,7 +97,7 @@ static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
        kcb->prev_kprobe.saved_eflags = kcb->kprobe_saved_eflags;
 }
 
-static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
+static inline void restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 {
        __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
        kcb->kprobe_status = kcb->prev_kprobe.status;
@@ -154,7 +105,7 @@ static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
        kcb->kprobe_saved_eflags = kcb->prev_kprobe.saved_eflags;
 }
 
-static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
+static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
                                struct kprobe_ctlblk *kcb)
 {
        __get_cpu_var(current_kprobe) = p;
@@ -164,7 +115,7 @@ static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
                kcb->kprobe_saved_eflags &= ~IF_MASK;
 }
 
-static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
+static inline void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 {
        regs->eflags |= TF_MASK;
        regs->eflags &= ~IF_MASK;
@@ -204,13 +155,9 @@ static int __kprobes kprobe_handler(struct pt_regs *regs)
 {
        struct kprobe *p;
        int ret = 0;
-       kprobe_opcode_t *addr;
+       kprobe_opcode_t *addr = NULL;
+       unsigned long *lp;
        struct kprobe_ctlblk *kcb;
-#ifdef CONFIG_PREEMPT
-       unsigned pre_preempt_count = preempt_count();
-#endif /* CONFIG_PREEMPT */
-
-       addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
 
        /*
         * We don't want to be preempted for the entire
@@ -219,6 +166,17 @@ static int __kprobes kprobe_handler(struct pt_regs *regs)
        preempt_disable();
        kcb = get_kprobe_ctlblk();
 
+       /* 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.
+        */
+       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));
+       }
        /* Check we're not actually recursing */
        if (kprobe_running()) {
                p = get_kprobe(addr);
@@ -242,6 +200,10 @@ static int __kprobes kprobe_handler(struct pt_regs *regs)
                        kcb->kprobe_status = KPROBE_REENTER;
                        return 1;
                } else {
+                       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 by
                         * another cpu right after we hit, no further
@@ -261,6 +223,11 @@ static int __kprobes kprobe_handler(struct pt_regs *regs)
 
        p = get_kprobe(addr);
        if (!p) {
+               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
@@ -285,21 +252,6 @@ static int __kprobes kprobe_handler(struct pt_regs *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);
        kcb->kprobe_status = KPROBE_HIT_SS;
@@ -315,44 +267,17 @@ no_kprobe:
  * 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)
+ void kretprobe_trampoline_holder(void)
  {
-       asm volatile ( ".global kretprobe_trampoline\n"
+       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");
-}
+                       "nop\n");
+ }
 
 /*
- * Called from kretprobe_trampoline
+ * Called when we hit the probe point at kretprobe_trampoline
  */
-fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
+int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
 {
         struct kretprobe_instance *ri = NULL;
         struct hlist_head *head;
@@ -381,11 +306,8 @@ fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
                        /* another task is sharing our hash bucket */
                         continue;
 
-               if (ri->rp && ri->rp->handler){
-                       __get_cpu_var(current_kprobe) = &ri->rp->kp;
+               if (ri->rp && ri->rp->handler)
                        ri->rp->handler(ri, regs);
-                       __get_cpu_var(current_kprobe) = NULL;
-               }
 
                orig_ret_address = (unsigned long)ri->ret_addr;
                recycle_rp_inst(ri);
@@ -400,10 +322,18 @@ fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
        }
 
        BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
+       regs->eip = orig_ret_address;
 
+       reset_current_kprobe();
        spin_unlock_irqrestore(&kretprobe_lock, flags);
+       preempt_enable_no_resched();
 
-       return (void*)orig_ret_address;
+       /*
+        * By returning a non-zero value, we are telling
+        * kprobe_handler() that we don't want the post_handler
+        * to run (and have re-enabled preemption)
+        */
+        return 1;
 }
 
 /*
@@ -427,82 +357,62 @@ fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
  * 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 __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 orig_eip = (unsigned long)p->addr;
 
-       regs->eflags &= ~TF_MASK;
        switch (p->ainsn.insn[0]) {
        case 0x9c:              /* pushfl */
                *tos &= ~(TF_MASK | IF_MASK);
                *tos |= kcb->kprobe_old_eflags;
                break;
-       case 0xc2:              /* iret/ret/lret */
-       case 0xc3:
-       case 0xca:
+       case 0xc3:              /* ret/lret */
        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 0xc2:
+       case 0xca:
+               regs->eflags &= ~TF_MASK;
+               /* eip is already adjusted, no more changes required*/
+               return;
        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.
-                        * But this is not boostable
-                        */
+                       /* call absolute, indirect */
+                       /* Fix return addr; eip is correct. */
+                       next_eip = regs->eip;
                        *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. And this is boostable */
-                       p->ainsn.boostable = 1;
-                       goto no_change;
+                       /* eip is correct. */
+                       next_eip = regs->eip;
                }
+               break;
+       case 0xea:              /* jmp absolute -- eip is correct */
+               next_eip = regs->eip;
+               break;
        default:
                break;
        }
 
-       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->eflags &= ~TF_MASK;
+       if (next_eip) {
+               regs->eip = next_eip;
+       } else {
+               regs->eip = orig_eip + (regs->eip - copy_eip);
        }
-
-       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.
  */
-static int __kprobes post_kprobe_handler(struct pt_regs *regs)
+static inline int post_kprobe_handler(struct pt_regs *regs)
 {
        struct kprobe *cur = kprobe_running();
        struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
@@ -538,62 +448,20 @@ out:
        return 1;
 }
 
-static int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
+static inline int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 {
        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;
+       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;
+       if (kcb->kprobe_status & KPROBE_HIT_SS) {
+               resume_execution(cur, regs, kcb);
+               regs->eflags |= kcb->kprobe_old_eflags;
 
-               /*
-                * fixup_exception() could not handle it,
-                * Let do_page_fault() fix it.
-                */
-               break;
-       default:
-               break;
+               reset_current_kprobe();
+               preempt_enable_no_resched();
        }
        return 0;
 }
@@ -607,9 +475,6 @@ int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
        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))
@@ -699,7 +564,12 @@ int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
        return 0;
 }
 
+static struct kprobe trampoline_p = {
+       .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
+       .pre_handler = trampoline_probe_handler
+};
+
 int __init arch_init_kprobes(void)
 {
-       return 0;
+       return register_kprobe(&trampoline_p);
 }