kprobes: Support delayed unoptimizing
authorMasami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Fri, 3 Dec 2010 09:54:09 +0000 (18:54 +0900)
committerIngo Molnar <mingo@elte.hu>
Mon, 6 Dec 2010 16:59:30 +0000 (17:59 +0100)
Unoptimization occurs when a probe is unregistered or disabled,
and is heavy because it recovers instructions by using
stop_machine(). This patch delays unoptimization operations and
unoptimize several probes at once by using
text_poke_smp_batch(). This can avoid unexpected system slowdown
coming from stop_machine().

Changes in v5:
- Split this patch into several cleanup patches and this patch.
- Fix some text_mutex lock miss.
- Use bool instead of int for behavior flags.
- Add additional comment for (un)optimizing path.

Changes in v2:
- Use dynamic allocated buffers and params.

Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com>
Cc: Jason Baron <jbaron@redhat.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: 2nddept-manager@sdl.hitachi.co.jp
LKML-Reference: <20101203095409.2961.82733.stgit@ltc236.sdl.hitachi.co.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86/kernel/kprobes.c
kernel/kprobes.c

index 1cbd54c0df99189548a3a03f40fbb75a1703475a..da51dc8e77cb911180b88b8270a84c0a9286526d 100644 (file)
@@ -1184,6 +1184,10 @@ static void __kprobes optimized_callback(struct optimized_kprobe *op,
 {
        struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 
+       /* This is possible if op is under delayed unoptimizing */
+       if (kprobe_disabled(&op->kp))
+               return;
+
        preempt_disable();
        if (kprobe_running()) {
                kprobes_inc_nmissed_count(&op->kp);
index 1f4f9b9d5c899969d393c54bacc7b4c99e1584ea..ba4d4c0740cff601aa445e15e638674cc495f0a0 100644 (file)
@@ -354,6 +354,13 @@ 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
  */
@@ -384,6 +391,17 @@ void __kprobes opt_pre_handler(struct kprobe *p, struct pt_regs *regs)
        }
 }
 
+/* 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)
 {
@@ -397,6 +415,33 @@ static inline int kprobe_optready(struct kprobe *p)
        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).
@@ -422,9 +467,11 @@ static struct kprobe *__kprobes get_optimized_kprobe(unsigned long addr)
 
 /* 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
 
 /*
@@ -435,6 +482,11 @@ static __kprobes void do_optimize_kprobes(void)
 {
        struct optimized_kprobe *op, *tmp;
 
+       /* 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.
@@ -457,17 +509,79 @@ static __kprobes void do_optimize_kprobes(void)
        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);
+       list_for_each_entry_safe(op, tmp, &unoptimizing_list, list) {
+               /* Unoptimize kprobes */
+               arch_unoptimize_kprobe(op);
+               /* 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);
+                       /* Move only unused probes on free_list */
+                       list_move(&op->list, free_list);
+               } 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);
-       if (kprobes_all_disarmed || !kprobes_allow_optimization)
-               goto end;
 
        /*
-        * Wait for quiesence period to ensure all running interrupts
+        * 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
@@ -475,10 +589,24 @@ static __kprobes void kprobe_optimizer(struct work_struct *work)
         */
        synchronize_sched();
 
+       /* Step 3: Optimize kprobes after quiesence period */
        do_optimize_kprobes();
-end:
+
+       /* Step 4: Free cleaned kprobes after quiesence period */
+       do_free_cleaned_kprobes(&free_list);
+
        mutex_unlock(&kprobe_mutex);
        mutex_unlock(&module_mutex);
+
+       /* 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 */
@@ -504,27 +632,63 @@ static __kprobes void optimize_kprobe(struct kprobe *p)
        /* Check if it is already optimized. */
        if (op->kp.flags & KPROBE_FLAG_OPTIMIZED)
                return;
-
        op->kp.flags |= KPROBE_FLAG_OPTIMIZED;
-       list_add(&op->list, &optimizing_list);
-       if (!delayed_work_pending(&optimizing_work))
-               schedule_delayed_work(&optimizing_work, OPTIMIZE_DELAY);
+
+       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)
+static __kprobes void unoptimize_kprobe(struct kprobe *p, bool force)
 {
        struct optimized_kprobe *op;
 
-       if ((p->flags & KPROBE_FLAG_OPTIMIZED) && kprobe_aggrprobe(p)) {
-               op = container_of(p, struct optimized_kprobe, kp);
-               if (!list_empty(&op->list))
-                       /* Dequeue from the optimization queue */
+       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);
-               else
-                       /* Replace jump with break */
-                       arch_unoptimize_kprobe(op);
-               op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
+                       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();
        }
 }
 
@@ -534,12 +698,12 @@ 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 optimization queue */
+       if (!list_empty(&op->list))
+               /* Dequeue from the (un)optimization queue */
                list_del_init(&op->list);
-               op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
-       }
-       /* Don't unoptimize, because the target code will be freed. */
+
+       op->kp.flags &= ~KPROBE_FLAG_OPTIMIZED;
+       /* Don't touch the code, because it is already freed. */
        arch_remove_optimized_kprobe(op);
 }
 
@@ -552,16 +716,6 @@ static __kprobes void prepare_optimized_kprobe(struct kprobe *p)
        arch_prepare_optimized_kprobe(op);
 }
 
-/* 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);
-       kfree(op);
-}
-
 /* Allocate new optimized_kprobe and try to prepare optimized instructions */
 static __kprobes struct kprobe *alloc_aggr_kprobe(struct kprobe *p)
 {
@@ -596,7 +750,8 @@ static __kprobes void try_to_optimize_kprobe(struct kprobe *p)
        op = container_of(ap, struct optimized_kprobe, kp);
        if (!arch_prepared_optinsn(&op->optinsn)) {
                /* If failed to setup optimizing, fallback to kprobe */
-               free_aggr_kprobe(ap);
+               arch_remove_optimized_kprobe(op);
+               kfree(op);
                return;
        }
 
@@ -640,21 +795,16 @@ static void __kprobes unoptimize_all_kprobes(void)
                return;
 
        kprobes_allow_optimization = false;
-       printk(KERN_INFO "Kprobes globally unoptimized\n");
-       get_online_cpus();      /* For avoiding text_mutex deadlock */
-       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))
-                               unoptimize_kprobe(p);
+                               unoptimize_kprobe(p, false);
                }
        }
-
-       mutex_unlock(&text_mutex);
-       put_online_cpus();
-       /* Allow all currently running kprobes to complete */
-       synchronize_sched();
+       /* Wait for unoptimizing completion */
+       wait_for_kprobe_optimizer();
+       printk(KERN_INFO "Kprobes globally unoptimized\n");
 }
 
 int sysctl_kprobes_optimization;
@@ -678,6 +828,7 @@ int proc_kprobes_optimization_handler(struct ctl_table *table, int write,
 }
 #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;
@@ -685,37 +836,45 @@ static void __kprobes __arm_kprobe(struct kprobe *p)
        /* Check collision with other optimized kprobes */
        _p = get_optimized_kprobe((unsigned long)p->addr);
        if (unlikely(_p))
-               unoptimize_kprobe(_p); /* Fallback to unoptimized kprobe */
+               /* Fallback to unoptimized kprobe */
+               unoptimize_kprobe(_p, true);
 
        arch_arm_kprobe(p);
        optimize_kprobe(p);     /* Try to optimize (add kprobe to a list) */
 }
 
-static void __kprobes __disarm_kprobe(struct kprobe *p)
+/* 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);   /* Try to unoptimize */
-       arch_disarm_kprobe(p);
+       unoptimize_kprobe(p, false);    /* Try to unoptimize */
 
-       /* If another kprobe was blocked, optimize it. */
-       _p = get_optimized_kprobe((unsigned long)p->addr);
-       if (unlikely(_p))
-               optimize_kprobe(_p);
+       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)                   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)                     arch_disarm_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)
 
 static __kprobes void free_aggr_kprobe(struct kprobe *p)
 {
+       arch_remove_kprobe(p);
        kfree(p);
 }
 
@@ -741,11 +900,10 @@ static void __kprobes arm_kprobe(struct kprobe *kp)
 /* Disarm a kprobe with text_mutex */
 static void __kprobes disarm_kprobe(struct kprobe *kp)
 {
-       get_online_cpus();      /* For avoiding text_mutex deadlock */
+       /* Ditto */
        mutex_lock(&text_mutex);
-       __disarm_kprobe(kp);
+       __disarm_kprobe(kp, true);
        mutex_unlock(&text_mutex);
-       put_online_cpus();
 }
 
 /*
@@ -951,7 +1109,7 @@ 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);  /* Fall back to normal kprobe */
+               unoptimize_kprobe(ap, true);    /* Fall back to normal kprobe */
 
        if (p->break_handler) {
                if (ap->break_handler)
@@ -1014,7 +1172,9 @@ static int __kprobes register_aggr_kprobe(struct kprobe *orig_p,
                if (!ap)
                        return -ENOMEM;
                init_aggr_kprobe(ap, orig_p);
-       }
+       } else if (kprobe_unused(ap))
+               /* Busy to die */
+               return -EBUSY;
 
        if (kprobe_gone(ap)) {
                /*
@@ -1283,8 +1443,11 @@ static int __kprobes __unregister_kprobe_top(struct kprobe *p)
        /* Following process expects this probe is an aggrprobe */
        WARN_ON(!kprobe_aggrprobe(ap));
 
-       if (list_is_singular(&ap->list))
-               /* This probe is the last child of aggrprobe */
+       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 */
@@ -1313,6 +1476,7 @@ noclean:
        return 0;
 
 disarmed:
+       BUG_ON(!kprobe_disarmed(ap));
        hlist_del_rcu(&ap->hlist);
        return 0;
 }
@@ -1322,14 +1486,15 @@ static void __kprobes __unregister_kprobe_bottom(struct kprobe *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 */
+               /* This is the last child of an aggrprobe */
                ap = list_entry(p->list.next, struct kprobe, list);
                list_del(&p->list);
-               arch_remove_kprobe(ap);
                free_aggr_kprobe(ap);
        }
+       /* Otherwise, do nothing. */
 }
 
 int __kprobes register_kprobes(struct kprobe **kps, int num)
@@ -1951,36 +2116,27 @@ 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");
 
-       /*
-        * Here we call get_online_cpus() for avoiding text_mutex deadlock,
-        * because disarming may also unoptimize kprobes.
-        */
-       get_online_cpus();
        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))
-                               __disarm_kprobe(p);
+                               __disarm_kprobe(p, false);
                }
        }
-
        mutex_unlock(&text_mutex);
-       put_online_cpus();
        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();
 }
 
 /*