fedora core 6 1.2949 + vserver 2.2.0
[linux-2.6.git] / arch / sparc64 / kernel / kprobes.c
index 7066d7b..ae221f0 100644 (file)
@@ -3,12 +3,13 @@
  * Copyright (C) 2004 David S. Miller <davem@davemloft.net>
  */
 
-#include <linux/config.h>
 #include <linux/kernel.h>
 #include <linux/kprobes.h>
-
+#include <linux/module.h>
 #include <asm/kdebug.h>
 #include <asm/signal.h>
+#include <asm/cacheflush.h>
+#include <asm/uaccess.h>
 
 /* We do not have hardware single-stepping on sparc64.
  * So we implement software single-stepping with breakpoint
  * - Mark that we are no longer actively in a kprobe.
  */
 
-int arch_prepare_kprobe(struct kprobe *p)
+DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
+DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
+
+int __kprobes arch_prepare_kprobe(struct kprobe *p)
 {
+       p->ainsn.insn[0] = *p->addr;
+       flushi(&p->ainsn.insn[0]);
+
+       p->ainsn.insn[1] = BREAKPOINT_INSTRUCTION_2;
+       flushi(&p->ainsn.insn[1]);
+
+       p->opcode = *p->addr;
        return 0;
 }
 
-void arch_copy_kprobe(struct kprobe *p)
+void __kprobes arch_arm_kprobe(struct kprobe *p)
 {
-       p->ainsn.insn[0] = *p->addr;
-       p->ainsn.insn[1] = BREAKPOINT_INSTRUCTION_2;
+       *p->addr = BREAKPOINT_INSTRUCTION;
+       flushi(p->addr);
 }
 
-void arch_remove_kprobe(struct kprobe *p)
+void __kprobes arch_disarm_kprobe(struct kprobe *p)
 {
+       *p->addr = p->opcode;
+       flushi(p->addr);
 }
 
-/* kprobe_status settings */
-#define KPROBE_HIT_ACTIVE      0x00000001
-#define KPROBE_HIT_SS          0x00000002
+static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
+{
+       kcb->prev_kprobe.kp = kprobe_running();
+       kcb->prev_kprobe.status = kcb->kprobe_status;
+       kcb->prev_kprobe.orig_tnpc = kcb->kprobe_orig_tnpc;
+       kcb->prev_kprobe.orig_tstate_pil = kcb->kprobe_orig_tstate_pil;
+}
 
-static struct kprobe *current_kprobe;
-static unsigned long current_kprobe_orig_tnpc;
-static unsigned long current_kprobe_orig_tstate_pil;
-static unsigned int 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;
+       kcb->kprobe_orig_tnpc = kcb->prev_kprobe.orig_tnpc;
+       kcb->kprobe_orig_tstate_pil = kcb->prev_kprobe.orig_tstate_pil;
+}
 
-static inline void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
+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_orig_tnpc = regs->tnpc;
+       kcb->kprobe_orig_tstate_pil = (regs->tstate & TSTATE_PIL);
+}
+
+static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs,
+                       struct kprobe_ctlblk *kcb)
 {
-       current_kprobe_orig_tnpc = regs->tnpc;
-       current_kprobe_orig_tstate_pil = (regs->tstate & TSTATE_PIL);
        regs->tstate |= TSTATE_PIL;
 
        /*single step inline, if it a breakpoint instruction*/
        if (p->opcode == BREAKPOINT_INSTRUCTION) {
                regs->tpc = (unsigned long) p->addr;
-               regs->tnpc = current_kprobe_orig_tnpc;
+               regs->tnpc = kcb->kprobe_orig_tnpc;
        } else {
                regs->tpc = (unsigned long) &p->ainsn.insn[0];
                regs->tnpc = (unsigned long) &p->ainsn.insn[1];
        }
 }
 
-static inline void disarm_kprobe(struct kprobe *p, struct pt_regs *regs)
-{
-       *p->addr = p->opcode;
-       flushi(p->addr);
-
-       regs->tpc = (unsigned long) p->addr;
-       regs->tnpc = current_kprobe_orig_tnpc;
-       regs->tstate = ((regs->tstate & ~TSTATE_PIL) |
-                       current_kprobe_orig_tstate_pil);
-}
-
-static int kprobe_handler(struct pt_regs *regs)
+static int __kprobes kprobe_handler(struct pt_regs *regs)
 {
        struct kprobe *p;
        void *addr = (void *) regs->tpc;
        int ret = 0;
+       struct kprobe_ctlblk *kcb;
 
+       /*
+        * We don't want to be preempted for the entire
+        * duration of kprobe processing
+        */
        preempt_disable();
+       kcb = get_kprobe_ctlblk();
 
        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) {
                                regs->tstate = ((regs->tstate & ~TSTATE_PIL) |
-                                       current_kprobe_orig_tstate_pil);
-                               unlock_kprobes();
+                                       kcb->kprobe_orig_tstate_pil);
                                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);
+                       kcb->kprobe_status = KPROBE_REENTER;
+                       prepare_singlestep(p, regs, kcb);
+                       return 1;
                } else {
-                       p = current_kprobe;
+                       if (*(u32 *)addr != BREAKPOINT_INSTRUCTION) {
+                       /* The breakpoint instruction was removed by
+                        * another cpu right after we hit, no further
+                        * handling of this interrupt is appropriate
+                        */
+                               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 (*(u32 *)addr != BREAKPOINT_INSTRUCTION) {
                        /*
                         * The breakpoint instruction was removed right
@@ -138,14 +171,14 @@ static int kprobe_handler(struct pt_regs *regs)
                goto no_kprobe;
        }
 
-       kprobe_status = KPROBE_HIT_ACTIVE;
-       current_kprobe = p;
+       set_current_kprobe(p, regs, kcb);
+       kcb->kprobe_status = KPROBE_HIT_ACTIVE;
        if (p->pre_handler && p->pre_handler(p, regs))
                return 1;
 
 ss_probe:
-       prepare_singlestep(p, regs);
-       kprobe_status = KPROBE_HIT_SS;
+       prepare_singlestep(p, regs, kcb);
+       kcb->kprobe_status = KPROBE_HIT_SS;
        return 1;
 
 no_kprobe:
@@ -156,15 +189,19 @@ no_kprobe:
 /* If INSN is a relative control transfer instruction,
  * return the corrected branch destination value.
  *
- * The original INSN location was REAL_PC, it actually
- * executed at PC and produced destination address NPC.
+ * regs->tpc and regs->tnpc still hold the values of the
+ * program counters at the time of trap due to the execution
+ * of the BREAKPOINT_INSTRUCTION_2 at p->ainsn.insn[1]
+ * 
  */
-static unsigned long relbranch_fixup(u32 insn, unsigned long real_pc,
-                                    unsigned long pc, unsigned long npc)
+static unsigned long __kprobes relbranch_fixup(u32 insn, struct kprobe *p,
+                                              struct pt_regs *regs)
 {
+       unsigned long real_pc = (unsigned long) p->addr;
+
        /* Branch not taken, no mods necessary.  */
-       if (npc == pc + 0x4UL)
-               return real_pc + 0x4UL;
+       if (regs->tnpc == regs->tpc + 0x4UL)
+               return real_pc + 0x8UL;
 
        /* The three cases are call, branch w/prediction,
         * and traditional branch.
@@ -172,29 +209,37 @@ static unsigned long relbranch_fixup(u32 insn, unsigned long real_pc,
        if ((insn & 0xc0000000) == 0x40000000 ||
            (insn & 0xc1c00000) == 0x00400000 ||
            (insn & 0xc1c00000) == 0x00800000) {
+               unsigned long ainsn_addr;
+
+               ainsn_addr = (unsigned long) &p->ainsn.insn[0];
+
                /* The instruction did all the work for us
                 * already, just apply the offset to the correct
                 * instruction location.
                 */
-               return (real_pc + (npc - pc));
+               return (real_pc + (regs->tnpc - ainsn_addr));
        }
 
-       return real_pc + 0x4UL;
+       /* It is jmpl or some other absolute PC modification instruction,
+        * leave NPC as-is.
+        */
+       return regs->tnpc;
 }
 
 /* If INSN is an instruction which writes it's PC location
  * into a destination register, fix that up.
  */
-static void retpc_fixup(struct pt_regs *regs, u32 insn, unsigned long real_pc)
+static void __kprobes retpc_fixup(struct pt_regs *regs, u32 insn,
+                                 unsigned long real_pc)
 {
        unsigned long *slot = NULL;
 
-       /* Simplest cast is call, which always uses %o7 */
+       /* Simplest case is 'call', which always uses %o7 */
        if ((insn & 0xc0000000) == 0x40000000) {
                slot = &regs->u_regs[UREG_I7];
        }
 
-       /* Jmpl encodes the register inside of the opcode */
+       /* 'jmpl' encodes the register inside of the opcode */
        if ((insn & 0xc1f80000) == 0x81c00000) {
                unsigned long rd = ((insn >> 25) & 0x1f);
 
@@ -216,95 +261,165 @@ static void retpc_fixup(struct pt_regs *regs, u32 insn, unsigned long real_pc)
 
 /*
  * Called after single-stepping.  p->addr is the address of the
- * instruction whose first byte has been replaced by the breakpoint
+ * instruction which has been replaced by the breakpoint
  * instruction.  To avoid the SMP problems that can occur when we
  * temporarily put back the original opcode to single-step, we
  * single-stepped a copy of the instruction.  The address of this
- * copy is p->ainsn.insn.
+ * copy is &p->ainsn.insn[0].
  *
  * This function prepares to return from the post-single-step
  * breakpoint trap.
  */
-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)
 {
        u32 insn = p->ainsn.insn[0];
 
-       regs->tpc = current_kprobe_orig_tnpc;
-       regs->tnpc = relbranch_fixup(insn,
-                                    (unsigned long) p->addr,
-                                    (unsigned long) &p->ainsn.insn[0],
-                                    regs->tnpc);
+       regs->tnpc = relbranch_fixup(insn, p, regs);
+
+       /* This assignment must occur after relbranch_fixup() */
+       regs->tpc = kcb->kprobe_orig_tnpc;
+
        retpc_fixup(regs, insn, (unsigned long) p->addr);
 
        regs->tstate = ((regs->tstate & ~TSTATE_PIL) |
-                       current_kprobe_orig_tstate_pil);
+                       kcb->kprobe_orig_tstate_pil);
 }
 
-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);
+       resume_execution(cur, regs, kcb);
 
-       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();
 
        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;
+       struct kprobe *cur = kprobe_running();
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+       const struct exception_table_entry *entry;
+
+       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 tpc points back to the probe address
+                * and allow the page fault handler to continue as a
+                * normal page fault.
+                */
+               regs->tpc = (unsigned long)cur->addr;
+               regs->tnpc = kcb->kprobe_orig_tnpc;
+               regs->tstate = ((regs->tstate & ~TSTATE_PIL) |
+                               kcb->kprobe_orig_tstate_pil);
+               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 (kprobe_status & KPROBE_HIT_SS) {
-               resume_execution(current_kprobe, regs);
+               /*
+                * In case the user-specified fault handler returned
+                * zero, try to fix up.
+                */
 
-               unlock_kprobes();
-               preempt_enable_no_resched();
+               entry = search_exception_tables(regs->tpc);
+               if (entry) {
+                       regs->tpc = entry->fixup;
+                       regs->tnpc = regs->tpc + 4;
+                       return 1;
+               }
+
+               /*
+                * fixup_exception() could not handle it,
+                * Let do_page_fault() fix it.
+                */
+               break;
+       default:
+               break;
        }
+
        return 0;
 }
 
 /*
  * 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_DEBUG:
                if (kprobe_handler(args->regs))
-                       return NOTIFY_STOP;
+                       ret = NOTIFY_STOP;
                break;
        case DIE_DEBUG_2:
                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;
 }
 
-asmlinkage void kprobe_trap(unsigned long trap_level, struct pt_regs *regs)
+asmlinkage void __kprobes kprobe_trap(unsigned long trap_level,
+                                     struct pt_regs *regs)
 {
        BUG_ON(trap_level != 0x170 && trap_level != 0x171);
 
@@ -324,24 +439,12 @@ asmlinkage void kprobe_trap(unsigned long trap_level, struct pt_regs *regs)
 }
 
 /* Jprobes support.  */
-static struct pt_regs jprobe_saved_regs;
-static struct pt_regs *jprobe_saved_regs_location;
-static struct sparc_stackf jprobe_saved_stack;
-
-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);
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 
-       jprobe_saved_regs_location = regs;
-       memcpy(&jprobe_saved_regs, regs, sizeof(*regs));
-
-       /* Save a whole stack frame, this gets arguments
-        * pushed onto the stack after using up all the
-        * arg registers.
-        */
-       memcpy(&jprobe_saved_stack,
-              (char *) (regs->u_regs[UREG_FP] + STACK_BIAS),
-              sizeof(jprobe_saved_stack));
+       memcpy(&(kcb->jprobe_saved_regs), regs, sizeof(*regs));
 
        regs->tpc  = (unsigned long) jp->entry;
        regs->tnpc = ((unsigned long) jp->entry) + 0x4UL;
@@ -350,45 +453,42 @@ int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
        return 1;
 }
 
-void jprobe_return(void)
+void __kprobes jprobe_return(void)
 {
-       preempt_enable_no_resched();
-       __asm__ __volatile__(
-               ".globl jprobe_return_trap_instruction\n"
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
+       register unsigned long orig_fp asm("g1");
+
+       orig_fp = kcb->jprobe_saved_regs.u_regs[UREG_FP];
+       __asm__ __volatile__("\n"
+"1:    cmp             %%sp, %0\n\t"
+       "blu,a,pt       %%xcc, 1b\n\t"
+       " restore\n\t"
+       ".globl         jprobe_return_trap_instruction\n"
 "jprobe_return_trap_instruction:\n\t"
-               "ta 0x70");
+       "ta             0x70"
+       : /* no outputs */
+       : "r" (orig_fp));
 }
 
 extern void jprobe_return_trap_instruction(void);
 
 extern void __show_regs(struct pt_regs * regs);
 
-int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
+int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 {
        u32 *addr = (u32 *) regs->tpc;
+       struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 
        if (addr == (u32 *) jprobe_return_trap_instruction) {
-               if (jprobe_saved_regs_location != regs) {
-                       printk("JPROBE: Current regs (%p) does not match "
-                              "saved regs (%p).\n",
-                              regs, jprobe_saved_regs_location);
-                       printk("JPROBE: Saved registers\n");
-                       __show_regs(jprobe_saved_regs_location);
-                       printk("JPROBE: Current registers\n");
-                       __show_regs(regs);
-                       BUG();
-               }
-               /* Restore old register state.  Do pt_regs
-                * first so that UREG_FP is the original one for
-                * the stack frame restore.
-                */
-               memcpy(regs, &jprobe_saved_regs, sizeof(*regs));
-
-               memcpy((char *) (regs->u_regs[UREG_FP] + STACK_BIAS),
-                      &jprobe_saved_stack,
-                      sizeof(jprobe_saved_stack));
-
+               memcpy(regs, &(kcb->jprobe_saved_regs), sizeof(*regs));
+               preempt_enable_no_resched();
                return 1;
        }
        return 0;
 }
+
+/* architecture specific initialization */
+int arch_init_kprobes(void)
+{
+       return 0;
+}