usermodehelper: check subprocess_info->path != NULL
[linux-2.6.git] / kernel / kprobes.c
index b466afa..c62b854 100644 (file)
 #include <linux/init.h>
 #include <linux/slab.h>
 #include <linux/stddef.h>
-#include <linux/module.h>
+#include <linux/export.h>
 #include <linux/moduleloader.h>
 #include <linux/kallsyms.h>
 #include <linux/freezer.h>
 #include <linux/seq_file.h>
 #include <linux/debugfs.h>
+#include <linux/sysctl.h>
 #include <linux/kdebug.h>
 #include <linux/memory.h>
+#include <linux/ftrace.h>
+#include <linux/cpu.h>
+#include <linux/jump_label.h>
 
 #include <asm-generic/sections.h>
 #include <asm/cacheflush.h>
@@ -70,13 +74,14 @@ static struct hlist_head kretprobe_inst_table[KPROBE_TABLE_SIZE];
 /* NOTE: change this value only with kprobe_mutex held */
 static bool kprobes_all_disarmed;
 
-static DEFINE_MUTEX(kprobe_mutex);     /* Protects kprobe_table */
+/* This protects kprobe_table and optimizing_list */
+static DEFINE_MUTEX(kprobe_mutex);
 static DEFINE_PER_CPU(struct kprobe *, kprobe_instance) = NULL;
 static struct {
-       spinlock_t lock ____cacheline_aligned_in_smp;
+       raw_spinlock_t lock ____cacheline_aligned_in_smp;
 } kretprobe_table_locks[KPROBE_TABLE_SIZE];
 
-static spinlock_t *kretprobe_table_lock_ptr(unsigned long hash)
+static raw_spinlock_t *kretprobe_table_lock_ptr(unsigned long hash)
 {
        return &(kretprobe_table_locks[hash].lock);
 }
@@ -93,6 +98,7 @@ static struct kprobe_blackpoint kprobe_blacklist[] = {
        {"native_get_debugreg",},
        {"irq_entries_start",},
        {"common_interrupt",},
+       {"mcount",},    /* mcount can be called from everywhere */
        {NULL}    /* Terminator */
 };
 
@@ -103,81 +109,74 @@ static struct kprobe_blackpoint kprobe_blacklist[] = {
  * stepping on the instruction on a vmalloced/kmalloced/data page
  * is a recipe for disaster
  */
-#define INSNS_PER_PAGE (PAGE_SIZE/(MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
-
 struct kprobe_insn_page {
        struct list_head list;
        kprobe_opcode_t *insns;         /* Page of instruction slots */
-       char slot_used[INSNS_PER_PAGE];
        int nused;
        int ngarbage;
+       char slot_used[];
+};
+
+#define KPROBE_INSN_PAGE_SIZE(slots)                   \
+       (offsetof(struct kprobe_insn_page, slot_used) + \
+        (sizeof(char) * (slots)))
+
+struct kprobe_insn_cache {
+       struct list_head pages; /* list of kprobe_insn_page */
+       size_t insn_size;       /* size of instruction slot */
+       int nr_garbage;
 };
 
+static int slots_per_page(struct kprobe_insn_cache *c)
+{
+       return PAGE_SIZE/(c->insn_size * sizeof(kprobe_opcode_t));
+}
+
 enum kprobe_slot_state {
        SLOT_CLEAN = 0,
        SLOT_DIRTY = 1,
        SLOT_USED = 2,
 };
 
-static DEFINE_MUTEX(kprobe_insn_mutex);        /* Protects kprobe_insn_pages */
-static LIST_HEAD(kprobe_insn_pages);
-static int kprobe_garbage_slots;
-static int collect_garbage_slots(void);
-
-static int __kprobes check_safety(void)
-{
-       int ret = 0;
-#if defined(CONFIG_PREEMPT) && defined(CONFIG_FREEZER)
-       ret = freeze_processes();
-       if (ret == 0) {
-               struct task_struct *p, *q;
-               do_each_thread(p, q) {
-                       if (p != current && p->state == TASK_RUNNING &&
-                           p->pid != 0) {
-                               printk("Check failed: %s is running\n",p->comm);
-                               ret = -1;
-                               goto loop_end;
-                       }
-               } while_each_thread(p, q);
-       }
-loop_end:
-       thaw_processes();
-#else
-       synchronize_sched();
-#endif
-       return ret;
-}
+static DEFINE_MUTEX(kprobe_insn_mutex);        /* Protects kprobe_insn_slots */
+static struct kprobe_insn_cache kprobe_insn_slots = {
+       .pages = LIST_HEAD_INIT(kprobe_insn_slots.pages),
+       .insn_size = MAX_INSN_SIZE,
+       .nr_garbage = 0,
+};
+static int __kprobes collect_garbage_slots(struct kprobe_insn_cache *c);
 
 /**
  * __get_insn_slot() - Find a slot on an executable page for an instruction.
  * We allocate an executable page if there's no room on existing ones.
  */
-static kprobe_opcode_t __kprobes *__get_insn_slot(void)
+static kprobe_opcode_t __kprobes *__get_insn_slot(struct kprobe_insn_cache *c)
 {
        struct kprobe_insn_page *kip;
 
  retry:
-       list_for_each_entry(kip, &kprobe_insn_pages, list) {
-               if (kip->nused < INSNS_PER_PAGE) {
+       list_for_each_entry(kip, &c->pages, list) {
+               if (kip->nused < slots_per_page(c)) {
                        int i;
-                       for (i = 0; i < INSNS_PER_PAGE; i++) {
+                       for (i = 0; i < slots_per_page(c); i++) {
                                if (kip->slot_used[i] == SLOT_CLEAN) {
                                        kip->slot_used[i] = SLOT_USED;
                                        kip->nused++;
-                                       return kip->insns + (i * MAX_INSN_SIZE);
+                                       return kip->insns + (i * c->insn_size);
                                }
                        }
-                       /* Surprise!  No unused slots.  Fix kip->nused. */
-                       kip->nused = INSNS_PER_PAGE;
+                       /* kip->nused is broken. Fix it. */
+                       kip->nused = slots_per_page(c);
+                       WARN_ON(1);
                }
        }
 
        /* If there are any garbage slots, collect it and try again. */
-       if (kprobe_garbage_slots && collect_garbage_slots() == 0) {
+       if (c->nr_garbage && collect_garbage_slots(c) == 0)
                goto retry;
-       }
-       /* All out of space.  Need to allocate a new page. Use slot 0. */
-       kip = kmalloc(sizeof(struct kprobe_insn_page), GFP_KERNEL);
+
+       /* All out of space.  Need to allocate a new page. */
+       kip = kmalloc(KPROBE_INSN_PAGE_SIZE(slots_per_page(c)), GFP_KERNEL);
        if (!kip)
                return NULL;
 
@@ -192,20 +191,23 @@ static kprobe_opcode_t __kprobes *__get_insn_slot(void)
                return NULL;
        }
        INIT_LIST_HEAD(&kip->list);
-       list_add(&kip->list, &kprobe_insn_pages);
-       memset(kip->slot_used, SLOT_CLEAN, INSNS_PER_PAGE);
+       memset(kip->slot_used, SLOT_CLEAN, slots_per_page(c));
        kip->slot_used[0] = SLOT_USED;
        kip->nused = 1;
        kip->ngarbage = 0;
+       list_add(&kip->list, &c->pages);
        return kip->insns;
 }
 
+
 kprobe_opcode_t __kprobes *get_insn_slot(void)
 {
-       kprobe_opcode_t *ret;
+       kprobe_opcode_t *ret = NULL;
+
        mutex_lock(&kprobe_insn_mutex);
-       ret = __get_insn_slot();
+       ret = __get_insn_slot(&kprobe_insn_slots);
        mutex_unlock(&kprobe_insn_mutex);
+
        return ret;
 }
 
@@ -221,7 +223,7 @@ static int __kprobes collect_one_slot(struct kprobe_insn_page *kip, int idx)
                 * so as not to have to set it up again the
                 * next time somebody inserts a probe.
                 */
-               if (!list_is_singular(&kprobe_insn_pages)) {
+               if (!list_is_singular(&kip->list)) {
                        list_del(&kip->list);
                        module_free(NULL, kip->insns);
                        kfree(kip);
@@ -231,63 +233,96 @@ static int __kprobes collect_one_slot(struct kprobe_insn_page *kip, int idx)
        return 0;
 }
 
-static int __kprobes collect_garbage_slots(void)
+static int __kprobes collect_garbage_slots(struct kprobe_insn_cache *c)
 {
        struct kprobe_insn_page *kip, *next;
 
-       /* Ensure no-one is preepmted on the garbages */
-       if (check_safety())
-               return -EAGAIN;
+       /* Ensure no-one is interrupted on the garbages */
+       synchronize_sched();
 
-       list_for_each_entry_safe(kip, next, &kprobe_insn_pages, list) {
+       list_for_each_entry_safe(kip, next, &c->pages, list) {
                int i;
                if (kip->ngarbage == 0)
                        continue;
                kip->ngarbage = 0;      /* we will collect all garbages */
-               for (i = 0; i < INSNS_PER_PAGE; i++) {
+               for (i = 0; i < slots_per_page(c); i++) {
                        if (kip->slot_used[i] == SLOT_DIRTY &&
                            collect_one_slot(kip, i))
                                break;
                }
        }
-       kprobe_garbage_slots = 0;
+       c->nr_garbage = 0;
        return 0;
 }
 
-void __kprobes free_insn_slot(kprobe_opcode_t * slot, int dirty)
+static void __kprobes __free_insn_slot(struct kprobe_insn_cache *c,
+                                      kprobe_opcode_t *slot, int dirty)
 {
        struct kprobe_insn_page *kip;
 
-       mutex_lock(&kprobe_insn_mutex);
-       list_for_each_entry(kip, &kprobe_insn_pages, list) {
-               if (kip->insns <= slot &&
-                   slot < kip->insns + (INSNS_PER_PAGE * MAX_INSN_SIZE)) {
-                       int i = (slot - kip->insns) / MAX_INSN_SIZE;
+       list_for_each_entry(kip, &c->pages, list) {
+               long idx = ((long)slot - (long)kip->insns) /
+                               (c->insn_size * sizeof(kprobe_opcode_t));
+               if (idx >= 0 && idx < slots_per_page(c)) {
+                       WARN_ON(kip->slot_used[idx] != SLOT_USED);
                        if (dirty) {
-                               kip->slot_used[i] = SLOT_DIRTY;
+                               kip->slot_used[idx] = SLOT_DIRTY;
                                kip->ngarbage++;
+                               if (++c->nr_garbage > slots_per_page(c))
+                                       collect_garbage_slots(c);
                        } else
-                               collect_one_slot(kip, i);
-                       break;
+                               collect_one_slot(kip, idx);
+                       return;
                }
        }
+       /* Could not free this slot. */
+       WARN_ON(1);
+}
 
-       if (dirty && ++kprobe_garbage_slots > INSNS_PER_PAGE)
-               collect_garbage_slots();
-
+void __kprobes free_insn_slot(kprobe_opcode_t * slot, int dirty)
+{
+       mutex_lock(&kprobe_insn_mutex);
+       __free_insn_slot(&kprobe_insn_slots, slot, dirty);
        mutex_unlock(&kprobe_insn_mutex);
 }
+#ifdef CONFIG_OPTPROBES
+/* For optimized_kprobe buffer */
+static DEFINE_MUTEX(kprobe_optinsn_mutex); /* Protects kprobe_optinsn_slots */
+static struct kprobe_insn_cache kprobe_optinsn_slots = {
+       .pages = LIST_HEAD_INIT(kprobe_optinsn_slots.pages),
+       /* .insn_size is initialized later */
+       .nr_garbage = 0,
+};
+/* Get a slot for optimized_kprobe buffer */
+kprobe_opcode_t __kprobes *get_optinsn_slot(void)
+{
+       kprobe_opcode_t *ret = NULL;
+
+       mutex_lock(&kprobe_optinsn_mutex);
+       ret = __get_insn_slot(&kprobe_optinsn_slots);
+       mutex_unlock(&kprobe_optinsn_mutex);
+
+       return ret;
+}
+
+void __kprobes free_optinsn_slot(kprobe_opcode_t * slot, int dirty)
+{
+       mutex_lock(&kprobe_optinsn_mutex);
+       __free_insn_slot(&kprobe_optinsn_slots, slot, dirty);
+       mutex_unlock(&kprobe_optinsn_mutex);
+}
+#endif
 #endif
 
 /* We have preemption disabled.. so it is safe to use __ versions */
 static inline void set_kprobe_instance(struct kprobe *kp)
 {
-       __get_cpu_var(kprobe_instance) = kp;
+       __this_cpu_write(kprobe_instance, kp);
 }
 
 static inline void reset_kprobe_instance(void)
 {
-       __get_cpu_var(kprobe_instance) = NULL;
+       __this_cpu_write(kprobe_instance, NULL);
 }
 
 /*
@@ -307,22 +342,590 @@ struct kprobe __kprobes *get_kprobe(void *addr)
                if (p->addr == addr)
                        return p;
        }
+
+       return NULL;
+}
+
+static int __kprobes aggr_pre_handler(struct kprobe *p, struct pt_regs *regs);
+
+/* Return true if the kprobe is an aggregator */
+static inline int kprobe_aggrprobe(struct kprobe *p)
+{
+       return p->pre_handler == aggr_pre_handler;
+}
+
+/* Return true(!0) if the kprobe is unused */
+static inline int kprobe_unused(struct kprobe *p)
+{
+       return kprobe_aggrprobe(p) && kprobe_disabled(p) &&
+              list_empty(&p->list);
+}
+
+/*
+ * Keep all fields in the kprobe consistent
+ */
+static inline void copy_kprobe(struct kprobe *ap, struct kprobe *p)
+{
+       memcpy(&p->opcode, &ap->opcode, sizeof(kprobe_opcode_t));
+       memcpy(&p->ainsn, &ap->ainsn, sizeof(struct arch_specific_insn));
+}
+
+#ifdef CONFIG_OPTPROBES
+/* NOTE: change this value only with kprobe_mutex held */
+static bool kprobes_allow_optimization;
+
+/*
+ * Call all pre_handler on the list, but ignores its return value.
+ * This must be called from arch-dep optimized caller.
+ */
+void __kprobes opt_pre_handler(struct kprobe *p, struct pt_regs *regs)
+{
+       struct kprobe *kp;
+
+       list_for_each_entry_rcu(kp, &p->list, list) {
+               if (kp->pre_handler && likely(!kprobe_disabled(kp))) {
+                       set_kprobe_instance(kp);
+                       kp->pre_handler(kp, regs);
+               }
+               reset_kprobe_instance();
+       }
+}
+
+/* Free optimized instructions and optimized_kprobe */
+static __kprobes void free_aggr_kprobe(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       op = container_of(p, struct optimized_kprobe, kp);
+       arch_remove_optimized_kprobe(op);
+       arch_remove_kprobe(p);
+       kfree(op);
+}
+
+/* Return true(!0) if the kprobe is ready for optimization. */
+static inline int kprobe_optready(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       if (kprobe_aggrprobe(p)) {
+               op = container_of(p, struct optimized_kprobe, kp);
+               return arch_prepared_optinsn(&op->optinsn);
+       }
+
+       return 0;
+}
+
+/* Return true(!0) if the kprobe is disarmed. Note: p must be on hash list */
+static inline int kprobe_disarmed(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       /* If kprobe is not aggr/opt probe, just return kprobe is disabled */
+       if (!kprobe_aggrprobe(p))
+               return kprobe_disabled(p);
+
+       op = container_of(p, struct optimized_kprobe, kp);
+
+       return kprobe_disabled(p) && list_empty(&op->list);
+}
+
+/* Return true(!0) if the probe is queued on (un)optimizing lists */
+static int __kprobes kprobe_queued(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       if (kprobe_aggrprobe(p)) {
+               op = container_of(p, struct optimized_kprobe, kp);
+               if (!list_empty(&op->list))
+                       return 1;
+       }
+       return 0;
+}
+
+/*
+ * Return an optimized kprobe whose optimizing code replaces
+ * instructions including addr (exclude breakpoint).
+ */
+static struct kprobe *__kprobes get_optimized_kprobe(unsigned long addr)
+{
+       int i;
+       struct kprobe *p = NULL;
+       struct optimized_kprobe *op;
+
+       /* Don't check i == 0, since that is a breakpoint case. */
+       for (i = 1; !p && i < MAX_OPTIMIZED_LENGTH; i++)
+               p = get_kprobe((void *)(addr - i));
+
+       if (p && kprobe_optready(p)) {
+               op = container_of(p, struct optimized_kprobe, kp);
+               if (arch_within_optimized_kprobe(op, addr))
+                       return p;
+       }
+
        return NULL;
 }
 
+/* Optimization staging list, protected by kprobe_mutex */
+static LIST_HEAD(optimizing_list);
+static LIST_HEAD(unoptimizing_list);
+
+static void kprobe_optimizer(struct work_struct *work);
+static DECLARE_DELAYED_WORK(optimizing_work, kprobe_optimizer);
+static DECLARE_COMPLETION(optimizer_comp);
+#define OPTIMIZE_DELAY 5
+
+/*
+ * Optimize (replace a breakpoint with a jump) kprobes listed on
+ * optimizing_list.
+ */
+static __kprobes void do_optimize_kprobes(void)
+{
+       /* Optimization never be done when disarmed */
+       if (kprobes_all_disarmed || !kprobes_allow_optimization ||
+           list_empty(&optimizing_list))
+               return;
+
+       /*
+        * The optimization/unoptimization refers online_cpus via
+        * stop_machine() and cpu-hotplug modifies online_cpus.
+        * And same time, text_mutex will be held in cpu-hotplug and here.
+        * This combination can cause a deadlock (cpu-hotplug try to lock
+        * text_mutex but stop_machine can not be done because online_cpus
+        * has been changed)
+        * To avoid this deadlock, we need to call get_online_cpus()
+        * for preventing cpu-hotplug outside of text_mutex locking.
+        */
+       get_online_cpus();
+       mutex_lock(&text_mutex);
+       arch_optimize_kprobes(&optimizing_list);
+       mutex_unlock(&text_mutex);
+       put_online_cpus();
+}
+
+/*
+ * Unoptimize (replace a jump with a breakpoint and remove the breakpoint
+ * if need) kprobes listed on unoptimizing_list.
+ */
+static __kprobes void do_unoptimize_kprobes(struct list_head *free_list)
+{
+       struct optimized_kprobe *op, *tmp;
+
+       /* Unoptimization must be done anytime */
+       if (list_empty(&unoptimizing_list))
+               return;
+
+       /* Ditto to do_optimize_kprobes */
+       get_online_cpus();
+       mutex_lock(&text_mutex);
+       arch_unoptimize_kprobes(&unoptimizing_list, free_list);
+       /* Loop free_list for disarming */
+       list_for_each_entry_safe(op, tmp, free_list, list) {
+               /* Disarm probes if marked disabled */
+               if (kprobe_disabled(&op->kp))
+                       arch_disarm_kprobe(&op->kp);
+               if (kprobe_unused(&op->kp)) {
+                       /*
+                        * Remove unused probes from hash list. After waiting
+                        * for synchronization, these probes are reclaimed.
+                        * (reclaiming is done by do_free_cleaned_kprobes.)
+                        */
+                       hlist_del_rcu(&op->kp.hlist);
+               } else
+                       list_del_init(&op->list);
+       }
+       mutex_unlock(&text_mutex);
+       put_online_cpus();
+}
+
+/* Reclaim all kprobes on the free_list */
+static __kprobes void do_free_cleaned_kprobes(struct list_head *free_list)
+{
+       struct optimized_kprobe *op, *tmp;
+
+       list_for_each_entry_safe(op, tmp, free_list, list) {
+               BUG_ON(!kprobe_unused(&op->kp));
+               list_del_init(&op->list);
+               free_aggr_kprobe(&op->kp);
+       }
+}
+
+/* Start optimizer after OPTIMIZE_DELAY passed */
+static __kprobes void kick_kprobe_optimizer(void)
+{
+       if (!delayed_work_pending(&optimizing_work))
+               schedule_delayed_work(&optimizing_work, OPTIMIZE_DELAY);
+}
+
+/* Kprobe jump optimizer */
+static __kprobes void kprobe_optimizer(struct work_struct *work)
+{
+       LIST_HEAD(free_list);
+
+       /* Lock modules while optimizing kprobes */
+       mutex_lock(&module_mutex);
+       mutex_lock(&kprobe_mutex);
+
+       /*
+        * Step 1: Unoptimize kprobes and collect cleaned (unused and disarmed)
+        * kprobes before waiting for quiesence period.
+        */
+       do_unoptimize_kprobes(&free_list);
+
+       /*
+        * Step 2: Wait for quiesence period to ensure all running interrupts
+        * are done. Because optprobe may modify multiple instructions
+        * there is a chance that Nth instruction is interrupted. In that
+        * case, running interrupt can return to 2nd-Nth byte of jump
+        * instruction. This wait is for avoiding it.
+        */
+       synchronize_sched();
+
+       /* Step 3: Optimize kprobes after quiesence period */
+       do_optimize_kprobes();
+
+       /* Step 4: Free cleaned kprobes after quiesence period */
+       do_free_cleaned_kprobes(&free_list);
+
+       mutex_unlock(&kprobe_mutex);
+       mutex_unlock(&module_mutex);
+
+       /* Step 5: Kick optimizer again if needed */
+       if (!list_empty(&optimizing_list) || !list_empty(&unoptimizing_list))
+               kick_kprobe_optimizer();
+       else
+               /* Wake up all waiters */
+               complete_all(&optimizer_comp);
+}
+
+/* Wait for completing optimization and unoptimization */
+static __kprobes void wait_for_kprobe_optimizer(void)
+{
+       if (delayed_work_pending(&optimizing_work))
+               wait_for_completion(&optimizer_comp);
+}
+
+/* Optimize kprobe if p is ready to be optimized */
+static __kprobes void optimize_kprobe(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       /* Check if the kprobe is disabled or not ready for optimization. */
+       if (!kprobe_optready(p) || !kprobes_allow_optimization ||
+           (kprobe_disabled(p) || kprobes_all_disarmed))
+               return;
+
+       /* Both of break_handler and post_handler are not supported. */
+       if (p->break_handler || p->post_handler)
+               return;
+
+       op = container_of(p, struct optimized_kprobe, kp);
+
+       /* Check there is no other kprobes at the optimized instructions */
+       if (arch_check_optimized_kprobe(op) < 0)
+               return;
+
+       /* Check if it is already optimized. */
+       if (op->kp.flags & KPROBE_FLAG_OPTIMIZED)
+               return;
+       op->kp.flags |= KPROBE_FLAG_OPTIMIZED;
+
+       if (!list_empty(&op->list))
+               /* This is under unoptimizing. Just dequeue the probe */
+               list_del_init(&op->list);
+       else {
+               list_add(&op->list, &optimizing_list);
+               kick_kprobe_optimizer();
+       }
+}
+
+/* Short cut to direct unoptimizing */
+static __kprobes void force_unoptimize_kprobe(struct optimized_kprobe *op)
+{
+       get_online_cpus();
+       arch_unoptimize_kprobe(op);
+       put_online_cpus();
+       if (kprobe_disabled(&op->kp))
+               arch_disarm_kprobe(&op->kp);
+}
+
+/* Unoptimize a kprobe if p is optimized */
+static __kprobes void unoptimize_kprobe(struct kprobe *p, bool force)
+{
+       struct optimized_kprobe *op;
+
+       if (!kprobe_aggrprobe(p) || kprobe_disarmed(p))
+               return; /* This is not an optprobe nor optimized */
+
+       op = container_of(p, struct optimized_kprobe, kp);
+       if (!kprobe_optimized(p)) {
+               /* Unoptimized or unoptimizing case */
+               if (force && !list_empty(&op->list)) {
+                       /*
+                        * Only if this is unoptimizing kprobe and forced,
+                        * forcibly unoptimize it. (No need to unoptimize
+                        * unoptimized kprobe again :)
+                        */
+                       list_del_init(&op->list);
+                       force_unoptimize_kprobe(op);
+               }
+               return;
+       }
+
+       op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
+       if (!list_empty(&op->list)) {
+               /* Dequeue from the optimization queue */
+               list_del_init(&op->list);
+               return;
+       }
+       /* Optimized kprobe case */
+       if (force)
+               /* Forcibly update the code: this is a special case */
+               force_unoptimize_kprobe(op);
+       else {
+               list_add(&op->list, &unoptimizing_list);
+               kick_kprobe_optimizer();
+       }
+}
+
+/* Cancel unoptimizing for reusing */
+static void reuse_unused_kprobe(struct kprobe *ap)
+{
+       struct optimized_kprobe *op;
+
+       BUG_ON(!kprobe_unused(ap));
+       /*
+        * Unused kprobe MUST be on the way of delayed unoptimizing (means
+        * there is still a relative jump) and disabled.
+        */
+       op = container_of(ap, struct optimized_kprobe, kp);
+       if (unlikely(list_empty(&op->list)))
+               printk(KERN_WARNING "Warning: found a stray unused "
+                       "aggrprobe@%p\n", ap->addr);
+       /* Enable the probe again */
+       ap->flags &= ~KPROBE_FLAG_DISABLED;
+       /* Optimize it again (remove from op->list) */
+       BUG_ON(!kprobe_optready(ap));
+       optimize_kprobe(ap);
+}
+
+/* Remove optimized instructions */
+static void __kprobes kill_optimized_kprobe(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       op = container_of(p, struct optimized_kprobe, kp);
+       if (!list_empty(&op->list))
+               /* Dequeue from the (un)optimization queue */
+               list_del_init(&op->list);
+
+       op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
+       /* Don't touch the code, because it is already freed. */
+       arch_remove_optimized_kprobe(op);
+}
+
+/* Try to prepare optimized instructions */
+static __kprobes void prepare_optimized_kprobe(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       op = container_of(p, struct optimized_kprobe, kp);
+       arch_prepare_optimized_kprobe(op);
+}
+
+/* Allocate new optimized_kprobe and try to prepare optimized instructions */
+static __kprobes struct kprobe *alloc_aggr_kprobe(struct kprobe *p)
+{
+       struct optimized_kprobe *op;
+
+       op = kzalloc(sizeof(struct optimized_kprobe), GFP_KERNEL);
+       if (!op)
+               return NULL;
+
+       INIT_LIST_HEAD(&op->list);
+       op->kp.addr = p->addr;
+       arch_prepare_optimized_kprobe(op);
+
+       return &op->kp;
+}
+
+static void __kprobes init_aggr_kprobe(struct kprobe *ap, struct kprobe *p);
+
+/*
+ * Prepare an optimized_kprobe and optimize it
+ * NOTE: p must be a normal registered kprobe
+ */
+static __kprobes void try_to_optimize_kprobe(struct kprobe *p)
+{
+       struct kprobe *ap;
+       struct optimized_kprobe *op;
+
+       ap = alloc_aggr_kprobe(p);
+       if (!ap)
+               return;
+
+       op = container_of(ap, struct optimized_kprobe, kp);
+       if (!arch_prepared_optinsn(&op->optinsn)) {
+               /* If failed to setup optimizing, fallback to kprobe */
+               arch_remove_optimized_kprobe(op);
+               kfree(op);
+               return;
+       }
+
+       init_aggr_kprobe(ap, p);
+       optimize_kprobe(ap);
+}
+
+#ifdef CONFIG_SYSCTL
+/* This should be called with kprobe_mutex locked */
+static void __kprobes optimize_all_kprobes(void)
+{
+       struct hlist_head *head;
+       struct hlist_node *node;
+       struct kprobe *p;
+       unsigned int i;
+
+       /* If optimization is already allowed, just return */
+       if (kprobes_allow_optimization)
+               return;
+
+       kprobes_allow_optimization = true;
+       for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
+               head = &kprobe_table[i];
+               hlist_for_each_entry_rcu(p, node, head, hlist)
+                       if (!kprobe_disabled(p))
+                               optimize_kprobe(p);
+       }
+       printk(KERN_INFO "Kprobes globally optimized\n");
+}
+
+/* This should be called with kprobe_mutex locked */
+static void __kprobes unoptimize_all_kprobes(void)
+{
+       struct hlist_head *head;
+       struct hlist_node *node;
+       struct kprobe *p;
+       unsigned int i;
+
+       /* If optimization is already prohibited, just return */
+       if (!kprobes_allow_optimization)
+               return;
+
+       kprobes_allow_optimization = false;
+       for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
+               head = &kprobe_table[i];
+               hlist_for_each_entry_rcu(p, node, head, hlist) {
+                       if (!kprobe_disabled(p))
+                               unoptimize_kprobe(p, false);
+               }
+       }
+       /* Wait for unoptimizing completion */
+       wait_for_kprobe_optimizer();
+       printk(KERN_INFO "Kprobes globally unoptimized\n");
+}
+
+int sysctl_kprobes_optimization;
+int proc_kprobes_optimization_handler(struct ctl_table *table, int write,
+                                     void __user *buffer, size_t *length,
+                                     loff_t *ppos)
+{
+       int ret;
+
+       mutex_lock(&kprobe_mutex);
+       sysctl_kprobes_optimization = kprobes_allow_optimization ? 1 : 0;
+       ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
+
+       if (sysctl_kprobes_optimization)
+               optimize_all_kprobes();
+       else
+               unoptimize_all_kprobes();
+       mutex_unlock(&kprobe_mutex);
+
+       return ret;
+}
+#endif /* CONFIG_SYSCTL */
+
+/* Put a breakpoint for a probe. Must be called with text_mutex locked */
+static void __kprobes __arm_kprobe(struct kprobe *p)
+{
+       struct kprobe *_p;
+
+       /* Check collision with other optimized kprobes */
+       _p = get_optimized_kprobe((unsigned long)p->addr);
+       if (unlikely(_p))
+               /* Fallback to unoptimized kprobe */
+               unoptimize_kprobe(_p, true);
+
+       arch_arm_kprobe(p);
+       optimize_kprobe(p);     /* Try to optimize (add kprobe to a list) */
+}
+
+/* Remove the breakpoint of a probe. Must be called with text_mutex locked */
+static void __kprobes __disarm_kprobe(struct kprobe *p, bool reopt)
+{
+       struct kprobe *_p;
+
+       unoptimize_kprobe(p, false);    /* Try to unoptimize */
+
+       if (!kprobe_queued(p)) {
+               arch_disarm_kprobe(p);
+               /* If another kprobe was blocked, optimize it. */
+               _p = get_optimized_kprobe((unsigned long)p->addr);
+               if (unlikely(_p) && reopt)
+                       optimize_kprobe(_p);
+       }
+       /* TODO: reoptimize others after unoptimized this probe */
+}
+
+#else /* !CONFIG_OPTPROBES */
+
+#define optimize_kprobe(p)                     do {} while (0)
+#define unoptimize_kprobe(p, f)                        do {} while (0)
+#define kill_optimized_kprobe(p)               do {} while (0)
+#define prepare_optimized_kprobe(p)            do {} while (0)
+#define try_to_optimize_kprobe(p)              do {} while (0)
+#define __arm_kprobe(p)                                arch_arm_kprobe(p)
+#define __disarm_kprobe(p, o)                  arch_disarm_kprobe(p)
+#define kprobe_disarmed(p)                     kprobe_disabled(p)
+#define wait_for_kprobe_optimizer()            do {} while (0)
+
+/* There should be no unused kprobes can be reused without optimization */
+static void reuse_unused_kprobe(struct kprobe *ap)
+{
+       printk(KERN_ERR "Error: There should be no unused kprobe here.\n");
+       BUG_ON(kprobe_unused(ap));
+}
+
+static __kprobes void free_aggr_kprobe(struct kprobe *p)
+{
+       arch_remove_kprobe(p);
+       kfree(p);
+}
+
+static __kprobes struct kprobe *alloc_aggr_kprobe(struct kprobe *p)
+{
+       return kzalloc(sizeof(struct kprobe), GFP_KERNEL);
+}
+#endif /* CONFIG_OPTPROBES */
+
 /* Arm a kprobe with text_mutex */
 static void __kprobes arm_kprobe(struct kprobe *kp)
 {
+       /*
+        * Here, since __arm_kprobe() doesn't use stop_machine(),
+        * this doesn't cause deadlock on text_mutex. So, we don't
+        * need get_online_cpus().
+        */
        mutex_lock(&text_mutex);
-       arch_arm_kprobe(kp);
+       __arm_kprobe(kp);
        mutex_unlock(&text_mutex);
 }
 
 /* Disarm a kprobe with text_mutex */
 static void __kprobes disarm_kprobe(struct kprobe *kp)
 {
+       /* Ditto */
        mutex_lock(&text_mutex);
-       arch_disarm_kprobe(kp);
+       __disarm_kprobe(kp, true);
        mutex_unlock(&text_mutex);
 }
 
@@ -362,7 +965,7 @@ static void __kprobes aggr_post_handler(struct kprobe *p, struct pt_regs *regs,
 static int __kprobes aggr_fault_handler(struct kprobe *p, struct pt_regs *regs,
                                        int trapnr)
 {
-       struct kprobe *cur = __get_cpu_var(kprobe_instance);
+       struct kprobe *cur = __this_cpu_read(kprobe_instance);
 
        /*
         * if we faulted "during" the execution of a user specified
@@ -377,7 +980,7 @@ static int __kprobes aggr_fault_handler(struct kprobe *p, struct pt_regs *regs,
 
 static int __kprobes aggr_break_handler(struct kprobe *p, struct pt_regs *regs)
 {
-       struct kprobe *cur = __get_cpu_var(kprobe_instance);
+       struct kprobe *cur = __this_cpu_read(kprobe_instance);
        int ret = 0;
 
        if (cur && cur->break_handler) {
@@ -392,7 +995,7 @@ static int __kprobes aggr_break_handler(struct kprobe *p, struct pt_regs *regs)
 void __kprobes kprobes_inc_nmissed_count(struct kprobe *p)
 {
        struct kprobe *kp;
-       if (p->pre_handler != aggr_pre_handler) {
+       if (!kprobe_aggrprobe(p)) {
                p->nmissed++;
        } else {
                list_for_each_entry_rcu(kp, &p->list, list)
@@ -410,9 +1013,9 @@ void __kprobes recycle_rp_inst(struct kretprobe_instance *ri,
        hlist_del(&ri->hlist);
        INIT_HLIST_NODE(&ri->hlist);
        if (likely(rp)) {
-               spin_lock(&rp->lock);
+               raw_spin_lock(&rp->lock);
                hlist_add_head(&ri->hlist, &rp->free_instances);
-               spin_unlock(&rp->lock);
+               raw_spin_unlock(&rp->lock);
        } else
                /* Unregistering */
                hlist_add_head(&ri->hlist, head);
@@ -420,36 +1023,41 @@ void __kprobes recycle_rp_inst(struct kretprobe_instance *ri,
 
 void __kprobes kretprobe_hash_lock(struct task_struct *tsk,
                         struct hlist_head **head, unsigned long *flags)
+__acquires(hlist_lock)
 {
        unsigned long hash = hash_ptr(tsk, KPROBE_HASH_BITS);
-       spinlock_t *hlist_lock;
+       raw_spinlock_t *hlist_lock;
 
        *head = &kretprobe_inst_table[hash];
        hlist_lock = kretprobe_table_lock_ptr(hash);
-       spin_lock_irqsave(hlist_lock, *flags);
+       raw_spin_lock_irqsave(hlist_lock, *flags);
 }
 
 static void __kprobes kretprobe_table_lock(unsigned long hash,
        unsigned long *flags)
+__acquires(hlist_lock)
 {
-       spinlock_t *hlist_lock = kretprobe_table_lock_ptr(hash);
-       spin_lock_irqsave(hlist_lock, *flags);
+       raw_spinlock_t *hlist_lock = kretprobe_table_lock_ptr(hash);
+       raw_spin_lock_irqsave(hlist_lock, *flags);
 }
 
 void __kprobes kretprobe_hash_unlock(struct task_struct *tsk,
        unsigned long *flags)
+__releases(hlist_lock)
 {
        unsigned long hash = hash_ptr(tsk, KPROBE_HASH_BITS);
-       spinlock_t *hlist_lock;
+       raw_spinlock_t *hlist_lock;
 
        hlist_lock = kretprobe_table_lock_ptr(hash);
-       spin_unlock_irqrestore(hlist_lock, *flags);
+       raw_spin_unlock_irqrestore(hlist_lock, *flags);
 }
 
-void __kprobes kretprobe_table_unlock(unsigned long hash, unsigned long *flags)
+static void __kprobes kretprobe_table_unlock(unsigned long hash,
+       unsigned long *flags)
+__releases(hlist_lock)
 {
-       spinlock_t *hlist_lock = kretprobe_table_lock_ptr(hash);
-       spin_unlock_irqrestore(hlist_lock, *flags);
+       raw_spinlock_t *hlist_lock = kretprobe_table_lock_ptr(hash);
+       raw_spin_unlock_irqrestore(hlist_lock, *flags);
 }
 
 /*
@@ -469,6 +1077,7 @@ void __kprobes kprobe_flush_task(struct task_struct *tk)
                /* Early boot.  kretprobe_table_locks not yet initialized. */
                return;
 
+       INIT_HLIST_HEAD(&empty_rp);
        hash = hash_ptr(tk, KPROBE_HASH_BITS);
        head = &kretprobe_inst_table[hash];
        kretprobe_table_lock(hash, &flags);
@@ -477,7 +1086,6 @@ void __kprobes kprobe_flush_task(struct task_struct *tk)
                        recycle_rp_inst(ri, &empty_rp);
        }
        kretprobe_table_unlock(hash, &flags);
-       INIT_HLIST_HEAD(&empty_rp);
        hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
                hlist_del(&ri->hlist);
                kfree(ri);
@@ -516,21 +1124,16 @@ static void __kprobes cleanup_rp_inst(struct kretprobe *rp)
 }
 
 /*
- * Keep all fields in the kprobe consistent
- */
-static inline void copy_kprobe(struct kprobe *old_p, struct kprobe *p)
-{
-       memcpy(&p->opcode, &old_p->opcode, sizeof(kprobe_opcode_t));
-       memcpy(&p->ainsn, &old_p->ainsn, sizeof(struct arch_specific_insn));
-}
-
-/*
 * Add the new probe to ap->list. Fail if this is the
 * second jprobe at the address - two jprobes can't coexist
 */
 static int __kprobes add_new_kprobe(struct kprobe *ap, struct kprobe *p)
 {
        BUG_ON(kprobe_gone(ap) || kprobe_gone(p));
+
+       if (p->break_handler || p->post_handler)
+               unoptimize_kprobe(ap, true);    /* Fall back to normal kprobe */
+
        if (p->break_handler) {
                if (ap->break_handler)
                        return -EEXIST;
@@ -545,7 +1148,7 @@ static int __kprobes add_new_kprobe(struct kprobe *ap, struct kprobe *p)
                ap->flags &= ~KPROBE_FLAG_DISABLED;
                if (!kprobes_all_disarmed)
                        /* Arm the breakpoint again. */
-                       arm_kprobe(ap);
+                       __arm_kprobe(ap);
        }
        return 0;
 }
@@ -554,12 +1157,13 @@ static int __kprobes add_new_kprobe(struct kprobe *ap, struct kprobe *p)
  * Fill in the required fields of the "manager kprobe". Replace the
  * earlier kprobe in the hlist with the manager kprobe
  */
-static inline void add_aggr_kprobe(struct kprobe *ap, struct kprobe *p)
+static void __kprobes init_aggr_kprobe(struct kprobe *ap, struct kprobe *p)
 {
+       /* Copy p's insn slot to ap */
        copy_kprobe(p, ap);
        flush_insn_slot(ap);
        ap->addr = p->addr;
-       ap->flags = p->flags;
+       ap->flags = p->flags & ~KPROBE_FLAG_OPTIMIZED;
        ap->pre_handler = aggr_pre_handler;
        ap->fault_handler = aggr_fault_handler;
        /* We don't care the kprobe which has gone. */
@@ -569,8 +1173,9 @@ static inline void add_aggr_kprobe(struct kprobe *ap, struct kprobe *p)
                ap->break_handler = aggr_break_handler;
 
        INIT_LIST_HEAD(&ap->list);
-       list_add_rcu(&p->list, &ap->list);
+       INIT_HLIST_NODE(&ap->hlist);
 
+       list_add_rcu(&p->list, &ap->list);
        hlist_replace_rcu(&p->hlist, &ap->hlist);
 }
 
@@ -578,19 +1183,21 @@ static inline void add_aggr_kprobe(struct kprobe *ap, struct kprobe *p)
  * This is the second or subsequent kprobe at the address - handle
  * the intricacies
  */
-static int __kprobes register_aggr_kprobe(struct kprobe *old_p,
+static int __kprobes register_aggr_kprobe(struct kprobe *orig_p,
                                          struct kprobe *p)
 {
        int ret = 0;
-       struct kprobe *ap = old_p;
+       struct kprobe *ap = orig_p;
 
-       if (old_p->pre_handler != aggr_pre_handler) {
-               /* If old_p is not an aggr_probe, create new aggr_kprobe. */
-               ap = kzalloc(sizeof(struct kprobe), GFP_KERNEL);
+       if (!kprobe_aggrprobe(orig_p)) {
+               /* If orig_p is not an aggr_kprobe, create new aggr_kprobe. */
+               ap = alloc_aggr_kprobe(orig_p);
                if (!ap)
                        return -ENOMEM;
-               add_aggr_kprobe(ap, old_p);
-       }
+               init_aggr_kprobe(ap, orig_p);
+       } else if (kprobe_unused(ap))
+               /* This probe is going to die. Rescue it */
+               reuse_unused_kprobe(ap);
 
        if (kprobe_gone(ap)) {
                /*
@@ -608,6 +1215,9 @@ static int __kprobes register_aggr_kprobe(struct kprobe *old_p,
                         */
                        return ret;
 
+               /* Prepare optimized instructions if possible. */
+               prepare_optimized_kprobe(ap);
+
                /*
                 * Clear gone flag to prevent allocating new slot again, and
                 * set disabled flag because it is not armed yet.
@@ -616,27 +1226,11 @@ static int __kprobes register_aggr_kprobe(struct kprobe *old_p,
                            | KPROBE_FLAG_DISABLED;
        }
 
+       /* Copy ap's insn slot to p */
        copy_kprobe(ap, p);
        return add_new_kprobe(ap, p);
 }
 
-/* Try to disable aggr_kprobe, and return 1 if succeeded.*/
-static int __kprobes try_to_disable_aggr_kprobe(struct kprobe *p)
-{
-       struct kprobe *kp;
-
-       list_for_each_entry_rcu(kp, &p->list, list) {
-               if (!kprobe_disabled(kp))
-                       /*
-                        * There is an active probe on the list.
-                        * We can't disable aggr_kprobe.
-                        */
-                       return 0;
-       }
-       p->flags |= KPROBE_FLAG_DISABLED;
-       return 1;
-}
-
 static int __kprobes in_kprobes_functions(unsigned long addr)
 {
        struct kprobe_blackpoint *kb;
@@ -661,52 +1255,61 @@ static int __kprobes in_kprobes_functions(unsigned long addr)
 /*
  * If we have a symbol_name argument, look it up and add the offset field
  * to it. This way, we can specify a relative address to a symbol.
+ * This returns encoded errors if it fails to look up symbol or invalid
+ * combination of parameters.
  */
 static kprobe_opcode_t __kprobes *kprobe_addr(struct kprobe *p)
 {
        kprobe_opcode_t *addr = p->addr;
+
+       if ((p->symbol_name && p->addr) ||
+           (!p->symbol_name && !p->addr))
+               goto invalid;
+
        if (p->symbol_name) {
-               if (addr)
-                       return NULL;
                kprobe_lookup_name(p->symbol_name, addr);
+               if (!addr)
+                       return ERR_PTR(-ENOENT);
        }
 
-       if (!addr)
-               return NULL;
-       return (kprobe_opcode_t *)(((char *)addr) + p->offset);
+       addr = (kprobe_opcode_t *)(((char *)addr) + p->offset);
+       if (addr)
+               return addr;
+
+invalid:
+       return ERR_PTR(-EINVAL);
 }
 
 /* Check passed kprobe is valid and return kprobe in kprobe_table. */
 static struct kprobe * __kprobes __get_valid_kprobe(struct kprobe *p)
 {
-       struct kprobe *old_p, *list_p;
+       struct kprobe *ap, *list_p;
 
-       old_p = get_kprobe(p->addr);
-       if (unlikely(!old_p))
+       ap = get_kprobe(p->addr);
+       if (unlikely(!ap))
                return NULL;
 
-       if (p != old_p) {
-               list_for_each_entry_rcu(list_p, &old_p->list, list)
+       if (p != ap) {
+               list_for_each_entry_rcu(list_p, &ap->list, list)
                        if (list_p == p)
                        /* kprobe p is a valid probe */
                                goto valid;
                return NULL;
        }
 valid:
-       return old_p;
+       return ap;
 }
 
 /* Return error if the kprobe is being re-registered */
 static inline int check_kprobe_rereg(struct kprobe *p)
 {
        int ret = 0;
-       struct kprobe *old_p;
 
        mutex_lock(&kprobe_mutex);
-       old_p = __get_valid_kprobe(p);
-       if (old_p)
+       if (__get_valid_kprobe(p))
                ret = -EINVAL;
        mutex_unlock(&kprobe_mutex);
+
        return ret;
 }
 
@@ -718,19 +1321,22 @@ int __kprobes register_kprobe(struct kprobe *p)
        kprobe_opcode_t *addr;
 
        addr = kprobe_addr(p);
-       if (!addr)
-               return -EINVAL;
+       if (IS_ERR(addr))
+               return PTR_ERR(addr);
        p->addr = addr;
 
        ret = check_kprobe_rereg(p);
        if (ret)
                return ret;
 
+       jump_label_lock();
        preempt_disable();
        if (!kernel_text_address((unsigned long) p->addr) ||
-           in_kprobes_functions((unsigned long) p->addr)) {
-               preempt_enable();
-               return -EINVAL;
+           in_kprobes_functions((unsigned long) p->addr) ||
+           ftrace_text_reserved(p->addr, p->addr) ||
+           jump_label_text_reserved(p->addr, p->addr)) {
+               ret = -EINVAL;
+               goto cannot_probe;
        }
 
        /* User can pass only KPROBE_FLAG_DISABLED to register_kprobe */
@@ -741,14 +1347,15 @@ int __kprobes register_kprobe(struct kprobe *p)
         */
        probed_mod = __module_text_address((unsigned long) p->addr);
        if (probed_mod) {
+               /* Return -ENOENT if fail. */
+               ret = -ENOENT;
                /*
                 * We must hold a refcount of the probed module while updating
                 * its code to prohibit unexpected unloading.
                 */
-               if (unlikely(!try_module_get(probed_mod))) {
-                       preempt_enable();
-                       return -EINVAL;
-               }
+               if (unlikely(!try_module_get(probed_mod)))
+                       goto cannot_probe;
+
                /*
                 * If the module freed .init.text, we couldn't insert
                 * kprobes in there.
@@ -756,101 +1363,176 @@ int __kprobes register_kprobe(struct kprobe *p)
                if (within_module_init((unsigned long)p->addr, probed_mod) &&
                    probed_mod->state != MODULE_STATE_COMING) {
                        module_put(probed_mod);
-                       preempt_enable();
-                       return -EINVAL;
+                       goto cannot_probe;
                }
+               /* ret will be updated by following code */
        }
        preempt_enable();
+       jump_label_unlock();
 
        p->nmissed = 0;
        INIT_LIST_HEAD(&p->list);
        mutex_lock(&kprobe_mutex);
+
+       jump_label_lock(); /* needed to call jump_label_text_reserved() */
+
+       get_online_cpus();      /* For avoiding text_mutex deadlock. */
+       mutex_lock(&text_mutex);
+
        old_p = get_kprobe(p->addr);
        if (old_p) {
+               /* Since this may unoptimize old_p, locking text_mutex. */
                ret = register_aggr_kprobe(old_p, p);
                goto out;
        }
 
-       mutex_lock(&text_mutex);
        ret = arch_prepare_kprobe(p);
        if (ret)
-               goto out_unlock_text;
+               goto out;
 
        INIT_HLIST_NODE(&p->hlist);
        hlist_add_head_rcu(&p->hlist,
                       &kprobe_table[hash_ptr(p->addr, KPROBE_HASH_BITS)]);
 
        if (!kprobes_all_disarmed && !kprobe_disabled(p))
-               arch_arm_kprobe(p);
+               __arm_kprobe(p);
+
+       /* Try to optimize kprobe */
+       try_to_optimize_kprobe(p);
 
-out_unlock_text:
-       mutex_unlock(&text_mutex);
 out:
+       mutex_unlock(&text_mutex);
+       put_online_cpus();
+       jump_label_unlock();
        mutex_unlock(&kprobe_mutex);
 
        if (probed_mod)
                module_put(probed_mod);
 
        return ret;
+
+cannot_probe:
+       preempt_enable();
+       jump_label_unlock();
+       return ret;
 }
 EXPORT_SYMBOL_GPL(register_kprobe);
 
+/* Check if all probes on the aggrprobe are disabled */
+static int __kprobes aggr_kprobe_disabled(struct kprobe *ap)
+{
+       struct kprobe *kp;
+
+       list_for_each_entry_rcu(kp, &ap->list, list)
+               if (!kprobe_disabled(kp))
+                       /*
+                        * There is an active probe on the list.
+                        * We can't disable this ap.
+                        */
+                       return 0;
+
+       return 1;
+}
+
+/* Disable one kprobe: Make sure called under kprobe_mutex is locked */
+static struct kprobe *__kprobes __disable_kprobe(struct kprobe *p)
+{
+       struct kprobe *orig_p;
+
+       /* Get an original kprobe for return */
+       orig_p = __get_valid_kprobe(p);
+       if (unlikely(orig_p == NULL))
+               return NULL;
+
+       if (!kprobe_disabled(p)) {
+               /* Disable probe if it is a child probe */
+               if (p != orig_p)
+                       p->flags |= KPROBE_FLAG_DISABLED;
+
+               /* Try to disarm and disable this/parent probe */
+               if (p == orig_p || aggr_kprobe_disabled(orig_p)) {
+                       disarm_kprobe(orig_p);
+                       orig_p->flags |= KPROBE_FLAG_DISABLED;
+               }
+       }
+
+       return orig_p;
+}
+
 /*
  * Unregister a kprobe without a scheduler synchronization.
  */
 static int __kprobes __unregister_kprobe_top(struct kprobe *p)
 {
-       struct kprobe *old_p, *list_p;
+       struct kprobe *ap, *list_p;
 
-       old_p = __get_valid_kprobe(p);
-       if (old_p == NULL)
+       /* Disable kprobe. This will disarm it if needed. */
+       ap = __disable_kprobe(p);
+       if (ap == NULL)
                return -EINVAL;
 
-       if (old_p == p ||
-           (old_p->pre_handler == aggr_pre_handler &&
-            list_is_singular(&old_p->list))) {
+       if (ap == p)
                /*
-                * Only probe on the hash list. Disarm only if kprobes are
-                * enabled and not gone - otherwise, the breakpoint would
-                * already have been removed. We save on flushing icache.
+                * This probe is an independent(and non-optimized) kprobe
+                * (not an aggrprobe). Remove from the hash list.
                 */
-               if (!kprobes_all_disarmed && !kprobe_disabled(old_p))
-                       disarm_kprobe(p);
-               hlist_del_rcu(&old_p->hlist);
-       } else {
+               goto disarmed;
+
+       /* Following process expects this probe is an aggrprobe */
+       WARN_ON(!kprobe_aggrprobe(ap));
+
+       if (list_is_singular(&ap->list) && kprobe_disarmed(ap))
+               /*
+                * !disarmed could be happen if the probe is under delayed
+                * unoptimizing.
+                */
+               goto disarmed;
+       else {
+               /* If disabling probe has special handlers, update aggrprobe */
                if (p->break_handler && !kprobe_gone(p))
-                       old_p->break_handler = NULL;
+                       ap->break_handler = NULL;
                if (p->post_handler && !kprobe_gone(p)) {
-                       list_for_each_entry_rcu(list_p, &old_p->list, list) {
+                       list_for_each_entry_rcu(list_p, &ap->list, list) {
                                if ((list_p != p) && (list_p->post_handler))
                                        goto noclean;
                        }
-                       old_p->post_handler = NULL;
+                       ap->post_handler = NULL;
                }
 noclean:
+               /*
+                * Remove from the aggrprobe: this path will do nothing in
+                * __unregister_kprobe_bottom().
+                */
                list_del_rcu(&p->list);
-               if (!kprobe_disabled(old_p)) {
-                       try_to_disable_aggr_kprobe(old_p);
-                       if (!kprobes_all_disarmed && kprobe_disabled(old_p))
-                               disarm_kprobe(old_p);
-               }
+               if (!kprobe_disabled(ap) && !kprobes_all_disarmed)
+                       /*
+                        * Try to optimize this probe again, because post
+                        * handler may have been changed.
+                        */
+                       optimize_kprobe(ap);
        }
        return 0;
+
+disarmed:
+       BUG_ON(!kprobe_disarmed(ap));
+       hlist_del_rcu(&ap->hlist);
+       return 0;
 }
 
 static void __kprobes __unregister_kprobe_bottom(struct kprobe *p)
 {
-       struct kprobe *old_p;
+       struct kprobe *ap;
 
        if (list_empty(&p->list))
+               /* This is an independent kprobe */
                arch_remove_kprobe(p);
        else if (list_is_singular(&p->list)) {
-               /* "p" is the last child of an aggr_kprobe */
-               old_p = list_entry(p->list.next, struct kprobe, list);
+               /* This is the last child of an aggrprobe */
+               ap = list_entry(p->list.next, struct kprobe, list);
                list_del(&p->list);
-               arch_remove_kprobe(old_p);
-               kfree(old_p);
+               free_aggr_kprobe(ap);
        }
+       /* Otherwise, do nothing. */
 }
 
 int __kprobes register_kprobes(struct kprobe **kps, int num)
@@ -914,18 +1596,19 @@ int __kprobes register_jprobes(struct jprobe **jps, int num)
        if (num <= 0)
                return -EINVAL;
        for (i = 0; i < num; i++) {
-               unsigned long addr;
+               unsigned long addr, offset;
                jp = jps[i];
                addr = arch_deref_entry_point(jp->entry);
 
-               if (!kernel_text_address(addr))
-                       ret = -EINVAL;
-               else {
-                       /* Todo: Verify probepoint is a function entry point */
+               /* Verify probepoint is a function entry point */
+               if (kallsyms_lookup_size_offset(addr, NULL, &offset) &&
+                   offset == 0) {
                        jp->kp.pre_handler = setjmp_pre_handler;
                        jp->kp.break_handler = longjmp_break_handler;
                        ret = register_kprobe(&jp->kp);
-               }
+               } else
+                       ret = -EINVAL;
+
                if (ret < 0) {
                        if (i > 0)
                                unregister_jprobes(jps, i);
@@ -982,18 +1665,22 @@ static int __kprobes pre_handler_kretprobe(struct kprobe *p,
 
        /*TODO: consider to only swap the RA after the last pre_handler fired */
        hash = hash_ptr(current, KPROBE_HASH_BITS);
-       spin_lock_irqsave(&rp->lock, flags);
+       raw_spin_lock_irqsave(&rp->lock, flags);
        if (!hlist_empty(&rp->free_instances)) {
                ri = hlist_entry(rp->free_instances.first,
                                struct kretprobe_instance, hlist);
                hlist_del(&ri->hlist);
-               spin_unlock_irqrestore(&rp->lock, flags);
+               raw_spin_unlock_irqrestore(&rp->lock, flags);
 
                ri->rp = rp;
                ri->task = current;
 
-               if (rp->entry_handler && rp->entry_handler(ri, regs))
+               if (rp->entry_handler && rp->entry_handler(ri, regs)) {
+                       raw_spin_lock_irqsave(&rp->lock, flags);
+                       hlist_add_head(&ri->hlist, &rp->free_instances);
+                       raw_spin_unlock_irqrestore(&rp->lock, flags);
                        return 0;
+               }
 
                arch_prepare_kretprobe(ri, regs);
 
@@ -1004,7 +1691,7 @@ static int __kprobes pre_handler_kretprobe(struct kprobe *p,
                kretprobe_table_unlock(hash, &flags);
        } else {
                rp->nmissed++;
-               spin_unlock_irqrestore(&rp->lock, flags);
+               raw_spin_unlock_irqrestore(&rp->lock, flags);
        }
        return 0;
 }
@@ -1018,8 +1705,8 @@ int __kprobes register_kretprobe(struct kretprobe *rp)
 
        if (kretprobe_blacklist_size) {
                addr = kprobe_addr(&rp->kp);
-               if (!addr)
-                       return -EINVAL;
+               if (IS_ERR(addr))
+                       return PTR_ERR(addr);
 
                for (i = 0; kretprobe_blacklist[i].name != NULL; i++) {
                        if (kretprobe_blacklist[i].addr == addr)
@@ -1035,12 +1722,12 @@ int __kprobes register_kretprobe(struct kretprobe *rp)
        /* Pre-allocate memory for max kretprobe instances */
        if (rp->maxactive <= 0) {
 #ifdef CONFIG_PREEMPT
-               rp->maxactive = max(10, 2 * NR_CPUS);
+               rp->maxactive = max_t(unsigned int, 10, 2*num_possible_cpus());
 #else
-               rp->maxactive = NR_CPUS;
+               rp->maxactive = num_possible_cpus();
 #endif
        }
-       spin_lock_init(&rp->lock);
+       raw_spin_lock_init(&rp->lock);
        INIT_HLIST_HEAD(&rp->free_instances);
        for (i = 0; i < rp->maxactive; i++) {
                inst = kmalloc(sizeof(struct kretprobe_instance) +
@@ -1145,7 +1832,7 @@ static void __kprobes kill_kprobe(struct kprobe *p)
        struct kprobe *kp;
 
        p->flags |= KPROBE_FLAG_GONE;
-       if (p->pre_handler == aggr_pre_handler) {
+       if (kprobe_aggrprobe(p)) {
                /*
                 * If this is an aggr_kprobe, we have to list all the
                 * chained probes and mark them GONE.
@@ -1154,6 +1841,7 @@ static void __kprobes kill_kprobe(struct kprobe *p)
                        kp->flags |= KPROBE_FLAG_GONE;
                p->post_handler = NULL;
                p->break_handler = NULL;
+               kill_optimized_kprobe(p);
        }
        /*
         * Here, we can remove insn_slot safely, because no thread calls
@@ -1162,6 +1850,56 @@ static void __kprobes kill_kprobe(struct kprobe *p)
        arch_remove_kprobe(p);
 }
 
+/* Disable one kprobe */
+int __kprobes disable_kprobe(struct kprobe *kp)
+{
+       int ret = 0;
+
+       mutex_lock(&kprobe_mutex);
+
+       /* Disable this kprobe */
+       if (__disable_kprobe(kp) == NULL)
+               ret = -EINVAL;
+
+       mutex_unlock(&kprobe_mutex);
+       return ret;
+}
+EXPORT_SYMBOL_GPL(disable_kprobe);
+
+/* Enable one kprobe */
+int __kprobes enable_kprobe(struct kprobe *kp)
+{
+       int ret = 0;
+       struct kprobe *p;
+
+       mutex_lock(&kprobe_mutex);
+
+       /* Check whether specified probe is valid. */
+       p = __get_valid_kprobe(kp);
+       if (unlikely(p == NULL)) {
+               ret = -EINVAL;
+               goto out;
+       }
+
+       if (kprobe_gone(kp)) {
+               /* This kprobe has gone, we couldn't enable it. */
+               ret = -EINVAL;
+               goto out;
+       }
+
+       if (p != kp)
+               kp->flags &= ~KPROBE_FLAG_DISABLED;
+
+       if (!kprobes_all_disarmed && kprobe_disabled(p)) {
+               p->flags &= ~KPROBE_FLAG_DISABLED;
+               arm_kprobe(p);
+       }
+out:
+       mutex_unlock(&kprobe_mutex);
+       return ret;
+}
+EXPORT_SYMBOL_GPL(enable_kprobe);
+
 void __kprobes dump_kprobe(struct kprobe *kp)
 {
        printk(KERN_WARNING "Dumping kprobe:\n");
@@ -1227,7 +1965,7 @@ static int __init init_kprobes(void)
        for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
                INIT_HLIST_HEAD(&kprobe_table[i]);
                INIT_HLIST_HEAD(&kretprobe_inst_table[i]);
-               spin_lock_init(&(kretprobe_table_locks[i].lock));
+               raw_spin_lock_init(&(kretprobe_table_locks[i].lock));
        }
 
        /*
@@ -1263,6 +2001,15 @@ static int __init init_kprobes(void)
                }
        }
 
+#if defined(CONFIG_OPTPROBES)
+#if defined(__ARCH_WANT_KPROBES_INSN_SLOT)
+       /* Init kprobe_optinsn_slots */
+       kprobe_optinsn_slots.insn_size = MAX_OPTINSN_SIZE;
+#endif
+       /* By default, kprobes can be optimized */
+       kprobes_allow_optimization = true;
+#endif
+
        /* By default, kprobes are armed */
        kprobes_all_disarmed = false;
 
@@ -1281,7 +2028,7 @@ static int __init init_kprobes(void)
 
 #ifdef CONFIG_DEBUG_FS
 static void __kprobes report_probe(struct seq_file *pi, struct kprobe *p,
-               const char *sym, int offset,char *modname)
+               const char *sym, int offset, char *modname, struct kprobe *pp)
 {
        char *kprobe_type;
 
@@ -1291,19 +2038,21 @@ static void __kprobes report_probe(struct seq_file *pi, struct kprobe *p,
                kprobe_type = "j";
        else
                kprobe_type = "k";
+
        if (sym)
-               seq_printf(pi, "%p  %s  %s+0x%x  %s %s%s\n",
+               seq_printf(pi, "%p  %s  %s+0x%x  %s ",
                        p->addr, kprobe_type, sym, offset,
-                       (modname ? modname : " "),
-                       (kprobe_gone(p) ? "[GONE]" : ""),
-                       ((kprobe_disabled(p) && !kprobe_gone(p)) ?
-                        "[DISABLED]" : ""));
+                       (modname ? modname : " "));
        else
-               seq_printf(pi, "%p  %s  %p %s%s\n",
-                       p->addr, kprobe_type, p->addr,
-                       (kprobe_gone(p) ? "[GONE]" : ""),
-                       ((kprobe_disabled(p) && !kprobe_gone(p)) ?
-                        "[DISABLED]" : ""));
+               seq_printf(pi, "%p  %s  %p ",
+                       p->addr, kprobe_type, p->addr);
+
+       if (!pp)
+               pp = p;
+       seq_printf(pi, "%s%s%s\n",
+               (kprobe_gone(p) ? "[GONE]" : ""),
+               ((kprobe_disabled(p) && !kprobe_gone(p)) ?  "[DISABLED]" : ""),
+               (kprobe_optimized(pp) ? "[OPTIMIZED]" : ""));
 }
 
 static void __kprobes *kprobe_seq_start(struct seq_file *f, loff_t *pos)
@@ -1339,11 +2088,11 @@ static int __kprobes show_kprobe_addr(struct seq_file *pi, void *v)
        hlist_for_each_entry_rcu(p, node, head, hlist) {
                sym = kallsyms_lookup((unsigned long)p->addr, NULL,
                                        &offset, &modname, namebuf);
-               if (p->pre_handler == aggr_pre_handler) {
+               if (kprobe_aggrprobe(p)) {
                        list_for_each_entry_rcu(kp, &p->list, list)
-                               report_probe(pi, kp, sym, offset, modname);
+                               report_probe(pi, kp, sym, offset, modname, p);
                } else
-                       report_probe(pi, p, sym, offset, modname);
+                       report_probe(pi, p, sym, offset, modname, NULL);
        }
        preempt_enable();
        return 0;
@@ -1361,78 +2110,13 @@ static int __kprobes kprobes_open(struct inode *inode, struct file *filp)
        return seq_open(filp, &kprobes_seq_ops);
 }
 
-static struct file_operations debugfs_kprobes_operations = {
+static const struct file_operations debugfs_kprobes_operations = {
        .open           = kprobes_open,
        .read           = seq_read,
        .llseek         = seq_lseek,
        .release        = seq_release,
 };
 
-/* Disable one kprobe */
-int __kprobes disable_kprobe(struct kprobe *kp)
-{
-       int ret = 0;
-       struct kprobe *p;
-
-       mutex_lock(&kprobe_mutex);
-
-       /* Check whether specified probe is valid. */
-       p = __get_valid_kprobe(kp);
-       if (unlikely(p == NULL)) {
-               ret = -EINVAL;
-               goto out;
-       }
-
-       /* If the probe is already disabled (or gone), just return */
-       if (kprobe_disabled(kp))
-               goto out;
-
-       kp->flags |= KPROBE_FLAG_DISABLED;
-       if (p != kp)
-               /* When kp != p, p is always enabled. */
-               try_to_disable_aggr_kprobe(p);
-
-       if (!kprobes_all_disarmed && kprobe_disabled(p))
-               disarm_kprobe(p);
-out:
-       mutex_unlock(&kprobe_mutex);
-       return ret;
-}
-EXPORT_SYMBOL_GPL(disable_kprobe);
-
-/* Enable one kprobe */
-int __kprobes enable_kprobe(struct kprobe *kp)
-{
-       int ret = 0;
-       struct kprobe *p;
-
-       mutex_lock(&kprobe_mutex);
-
-       /* Check whether specified probe is valid. */
-       p = __get_valid_kprobe(kp);
-       if (unlikely(p == NULL)) {
-               ret = -EINVAL;
-               goto out;
-       }
-
-       if (kprobe_gone(kp)) {
-               /* This kprobe has gone, we couldn't enable it. */
-               ret = -EINVAL;
-               goto out;
-       }
-
-       if (!kprobes_all_disarmed && kprobe_disabled(p))
-               arm_kprobe(p);
-
-       p->flags &= ~KPROBE_FLAG_DISABLED;
-       if (p != kp)
-               kp->flags &= ~KPROBE_FLAG_DISABLED;
-out:
-       mutex_unlock(&kprobe_mutex);
-       return ret;
-}
-EXPORT_SYMBOL_GPL(enable_kprobe);
-
 static void __kprobes arm_all_kprobes(void)
 {
        struct hlist_head *head;
@@ -1446,12 +2130,13 @@ static void __kprobes arm_all_kprobes(void)
        if (!kprobes_all_disarmed)
                goto already_enabled;
 
+       /* Arming kprobes doesn't optimize kprobe itself */
        mutex_lock(&text_mutex);
        for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
                head = &kprobe_table[i];
                hlist_for_each_entry_rcu(p, node, head, hlist)
                        if (!kprobe_disabled(p))
-                               arch_arm_kprobe(p);
+                               __arm_kprobe(p);
        }
        mutex_unlock(&text_mutex);
 
@@ -1473,29 +2158,27 @@ static void __kprobes disarm_all_kprobes(void)
        mutex_lock(&kprobe_mutex);
 
        /* If kprobes are already disarmed, just return */
-       if (kprobes_all_disarmed)
-               goto already_disabled;
+       if (kprobes_all_disarmed) {
+               mutex_unlock(&kprobe_mutex);
+               return;
+       }
 
        kprobes_all_disarmed = true;
        printk(KERN_INFO "Kprobes globally disabled\n");
+
        mutex_lock(&text_mutex);
        for (i = 0; i < KPROBE_TABLE_SIZE; i++) {
                head = &kprobe_table[i];
                hlist_for_each_entry_rcu(p, node, head, hlist) {
                        if (!arch_trampoline_kprobe(p) && !kprobe_disabled(p))
-                               arch_disarm_kprobe(p);
+                               __disarm_kprobe(p, false);
                }
        }
-
        mutex_unlock(&text_mutex);
        mutex_unlock(&kprobe_mutex);
-       /* Allow all currently running kprobes to complete */
-       synchronize_sched();
-       return;
 
-already_disabled:
-       mutex_unlock(&kprobe_mutex);
-       return;
+       /* Wait for disarming all kprobes by optimizer */
+       wait_for_kprobe_optimizer();
 }
 
 /*
@@ -1521,7 +2204,7 @@ static ssize_t write_enabled_file_bool(struct file *file,
               const char __user *user_buf, size_t count, loff_t *ppos)
 {
        char buf[32];
-       int buf_size;
+       size_t buf_size;
 
        buf_size = min(count, (sizeof(buf)-1));
        if (copy_from_user(buf, user_buf, buf_size))
@@ -1543,9 +2226,10 @@ static ssize_t write_enabled_file_bool(struct file *file,
        return count;
 }
 
-static struct file_operations fops_kp = {
+static const struct file_operations fops_kp = {
        .read =         read_enabled_file_bool,
        .write =        write_enabled_file_bool,
+       .llseek =       default_llseek,
 };
 
 static int __kprobes debugfs_kprobe_init(void)