perf_counter: unify and fix delayed counter wakeup

While going over the wakeup code I noticed delayed wakeups only work
for hardware counters but basically all software counters rely on
them.

This patch unifies and generalizes the delayed wakeup to fix this
issue.

Since we're dealing with NMI context bits here, use a cmpxchg() based
single link list implementation to track counters that have pending
wakeups.

[ This should really be generic code for delayed wakeups, but since we
  cannot use cmpxchg()/xchg() in generic code, I've let it live in the
  perf_counter code. -- Eric Dumazet could use it to aggregate the
  network wakeups. ]

Furthermore, the x86 method of using TIF flags was flawed in that its
quite possible to end up setting the bit on the idle task, loosing the
wakeup.

The powerpc method uses per-cpu storage and does appear to be
sufficient.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Paul Mackerras <paulus@samba.org>
Orig-LKML-Reference: <20090330171023.153932974@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

10 files changed, 142 insertions, 76 deletions

diff --git a/arch/powerpc/include/asm/hw_irq.h b/arch/powerpc/include/asm/hw_irq.h
index cb32d571c9c7..20a44d0c9fdd 100644
--- a/arch/powerpc/include/asm/hw_irq.h
+++ b/arch/powerpc/include/asm/hw_irq.h
@@ -132,7 +132,7 @@ static inline int irqs_disabled_flags(unsigned long flags)
 struct irq_chip;
 
 #ifdef CONFIG_PERF_COUNTERS
-static inline unsigned long get_perf_counter_pending(void)
+static inline unsigned long test_perf_counter_pending(void)
 {
         unsigned long x;
 
@@ -160,7 +160,7 @@ extern void perf_counter_do_pending(void);
 
 #else
 
-static inline unsigned long get_perf_counter_pending(void)
+static inline unsigned long test_perf_counter_pending(void)
 {
         return 0;
 }
diff --git a/arch/powerpc/kernel/irq.c b/arch/powerpc/kernel/irq.c
index 469e9635ff04..2cd471f92fe6 100644
--- a/arch/powerpc/kernel/irq.c
+++ b/arch/powerpc/kernel/irq.c
@@ -135,7 +135,7 @@ notrace void raw_local_irq_restore(unsigned long en)
                         iseries_handle_interrupts();
         }
 
-        if (get_perf_counter_pending()) {
+        if (test_perf_counter_pending()) {
                 clear_perf_counter_pending();
                 perf_counter_do_pending();
         }
diff --git a/arch/powerpc/kernel/perf_counter.c b/arch/powerpc/kernel/perf_counter.c
index df007fe0cc0b..cde720fc495c 100644
--- a/arch/powerpc/kernel/perf_counter.c
+++ b/arch/powerpc/kernel/perf_counter.c
@@ -650,24 +650,6 @@ hw_perf_counter_init(struct perf_counter *counter)
 }
 
 /*
- * Handle wakeups.
- */
-void perf_counter_do_pending(void)
-{
-        int i;
-        struct cpu_hw_counters *cpuhw = &__get_cpu_var(cpu_hw_counters);
-        struct perf_counter *counter;
-
-        for (i = 0; i < cpuhw->n_counters; ++i) {
-                counter = cpuhw->counter[i];
-                if (counter && counter->wakeup_pending) {
-                        counter->wakeup_pending = 0;
-                        wake_up(&counter->waitq);
-                }
-        }
-}
-
-/*
  * A counter has overflowed; update its count and record
  * things if requested.  Note that interrupts are hard-disabled
  * here so there is no possibility of being interrupted.
@@ -720,7 +702,7 @@ static void perf_counter_interrupt(struct pt_regs *regs)
         struct cpu_hw_counters *cpuhw = &__get_cpu_var(cpu_hw_counters);
         struct perf_counter *counter;
         long val;
-        int need_wakeup = 0, found = 0;
+        int found = 0;
 
         for (i = 0; i < cpuhw->n_counters; ++i) {
                 counter = cpuhw->counter[i];
@@ -761,7 +743,7 @@ static void perf_counter_interrupt(struct pt_regs *regs)
          * immediately; otherwise we'll have do the wakeup when interrupts
          * get soft-enabled.
          */
-        if (get_perf_counter_pending() && regs->softe) {
+        if (test_perf_counter_pending() && regs->softe) {
                 irq_enter();
                 clear_perf_counter_pending();
                 perf_counter_do_pending();
diff --git a/arch/x86/include/asm/perf_counter.h b/arch/x86/include/asm/perf_counter.h
index 1662043b340f..e2b0e66b2353 100644
--- a/arch/x86/include/asm/perf_counter.h
+++ b/arch/x86/include/asm/perf_counter.h
@@ -84,8 +84,9 @@ union cpuid10_edx {
 #define MSR_ARCH_PERFMON_FIXED_CTR2                     0x30b
 #define X86_PMC_IDX_FIXED_BUS_CYCLES                    (X86_PMC_IDX_FIXED + 2)
 
-#define set_perf_counter_pending()      \
-                set_tsk_thread_flag(current, TIF_PERF_COUNTERS);
+#define set_perf_counter_pending()      do { } while (0)
+#define clear_perf_counter_pending()    do { } while (0)
+#define test_perf_counter_pending()     (0)
 
 #ifdef CONFIG_PERF_COUNTERS
 extern void init_hw_perf_counters(void);
diff --git a/arch/x86/include/asm/thread_info.h b/arch/x86/include/asm/thread_info.h
index 3ffd5d2a3676..8820a73ae090 100644
--- a/arch/x86/include/asm/thread_info.h
+++ b/arch/x86/include/asm/thread_info.h
@@ -83,7 +83,6 @@ struct thread_info {
 #define TIF_SYSCALL_AUDIT       7       /* syscall auditing active */
 #define TIF_SECCOMP             8       /* secure computing */
 #define TIF_MCE_NOTIFY          10      /* notify userspace of an MCE */
-#define TIF_PERF_COUNTERS       11      /* notify perf counter work */
 #define TIF_NOTSC               16      /* TSC is not accessible in userland */
 #define TIF_IA32                17      /* 32bit process */
 #define TIF_FORK                18      /* ret_from_fork */
@@ -107,7 +106,6 @@ struct thread_info {
 #define _TIF_SYSCALL_AUDIT      (1 << TIF_SYSCALL_AUDIT)
 #define _TIF_SECCOMP            (1 << TIF_SECCOMP)
 #define _TIF_MCE_NOTIFY         (1 << TIF_MCE_NOTIFY)
-#define _TIF_PERF_COUNTERS      (1 << TIF_PERF_COUNTERS)
 #define _TIF_NOTSC              (1 << TIF_NOTSC)
 #define _TIF_IA32               (1 << TIF_IA32)
 #define _TIF_FORK               (1 << TIF_FORK)
@@ -141,7 +139,7 @@ struct thread_info {
 
 /* Only used for 64 bit */
 #define _TIF_DO_NOTIFY_MASK                                             \
-        (_TIF_SIGPENDING|_TIF_MCE_NOTIFY|_TIF_PERF_COUNTERS|_TIF_NOTIFY_RESUME)
+        (_TIF_SIGPENDING|_TIF_MCE_NOTIFY|_TIF_NOTIFY_RESUME)
 
 /* flags to check in __switch_to() */
 #define _TIF_WORK_CTXSW                                                 \
diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index 3f95b0cdc550..7aab177fb566 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -227,7 +227,6 @@ static int __hw_perf_counter_init(struct perf_counter *counter)
                  */
                 hwc->config |= pmc_ops->event_map(perf_event_id(hw_event));
         }
-        counter->wakeup_pending = 0;
 
         return 0;
 }
@@ -773,34 +772,6 @@ void smp_perf_counter_interrupt(struct pt_regs *regs)
         irq_exit();
 }
 
-/*
- * This handler is triggered by NMI contexts:
- */
-void perf_counter_notify(struct pt_regs *regs)
-{
-        struct cpu_hw_counters *cpuc;
-        unsigned long flags;
-        int bit, cpu;
-
-        local_irq_save(flags);
-        cpu = smp_processor_id();
-        cpuc = &per_cpu(cpu_hw_counters, cpu);
-
-        for_each_bit(bit, cpuc->used, X86_PMC_IDX_MAX) {
-                struct perf_counter *counter = cpuc->counters[bit];
-
-                if (!counter)
-                        continue;
-
-                if (counter->wakeup_pending) {
-                        counter->wakeup_pending = 0;
-                        wake_up(&counter->waitq);
-                }
-        }
-
-        local_irq_restore(flags);
-}
-
 void perf_counters_lapic_init(int nmi)
 {
         u32 apic_val;
diff --git a/arch/x86/kernel/signal.c b/arch/x86/kernel/signal.c
index 611615a92c90..0a813b17b172 100644
--- a/arch/x86/kernel/signal.c
+++ b/arch/x86/kernel/signal.c
@@ -6,7 +6,6 @@
  *  2000-06-20  Pentium III FXSR, SSE support by Gareth Hughes
  *  2000-2002   x86-64 support by Andi Kleen
  */
-#include <linux/perf_counter.h>
 #include <linux/sched.h>
 #include <linux/mm.h>
 #include <linux/smp.h>
@@ -872,11 +871,6 @@ do_notify_resume(struct pt_regs *regs, void *unused, __u32 thread_info_flags)
                 tracehook_notify_resume(regs);
         }
 
-        if (thread_info_flags & _TIF_PERF_COUNTERS) {
-                clear_thread_flag(TIF_PERF_COUNTERS);
-                perf_counter_notify(regs);
-        }
-
 #ifdef CONFIG_X86_32
         clear_thread_flag(TIF_IRET);
 #endif /* CONFIG_X86_32 */
diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index 6bf67ce17625..0d833228eee5 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -275,6 +275,10 @@ struct perf_mmap_data {
         void                            *data_pages[0];
 };
 
+struct perf_wakeup_entry {
+        struct perf_wakeup_entry *next;
+};
+
 /**
  * struct perf_counter - performance counter kernel representation:
  */
@@ -350,7 +354,7 @@ struct perf_counter {
         /* poll related */
         wait_queue_head_t               waitq;
         /* optional: for NMIs */
-        int                             wakeup_pending;
+        struct perf_wakeup_entry        wakeup;
 
         void (*destroy)(struct perf_counter *);
         struct rcu_head                 rcu_head;
@@ -427,7 +431,7 @@ extern void perf_counter_task_sched_out(struct task_struct *task, int cpu);
 extern void perf_counter_task_tick(struct task_struct *task, int cpu);
 extern void perf_counter_init_task(struct task_struct *child);
 extern void perf_counter_exit_task(struct task_struct *child);
-extern void perf_counter_notify(struct pt_regs *regs);
+extern void perf_counter_do_pending(void);
 extern void perf_counter_print_debug(void);
 extern void perf_counter_unthrottle(void);
 extern u64 hw_perf_save_disable(void);
@@ -461,7 +465,7 @@ static inline void
 perf_counter_task_tick(struct task_struct *task, int cpu)               { }
 static inline void perf_counter_init_task(struct task_struct *child)    { }
 static inline void perf_counter_exit_task(struct task_struct *child)    { }
-static inline void perf_counter_notify(struct pt_regs *regs)            { }
+static inline void perf_counter_do_pending(void)                        { }
 static inline void perf_counter_print_debug(void)                       { }
 static inline void perf_counter_unthrottle(void)                        { }
 static inline void hw_perf_restore(u64 ctrl)                            { }
@@ -469,8 +473,9 @@ static inline u64 hw_perf_save_disable(void)		      { return 0; }
 static inline int perf_counter_task_disable(void)       { return -EINVAL; }
 static inline int perf_counter_task_enable(void)        { return -EINVAL; }
 
-static inline void perf_swcounter_event(u32 event, u64 nr,
-                                        int nmi, struct pt_regs *regs)  { }
+static inline void
+perf_swcounter_event(u32 event, u64 nr, int nmi, struct pt_regs *regs)  { }
+
 #endif
 
 #endif /* __KERNEL__ */
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 3b862a7988cd..f70ff80e79d7 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1197,8 +1197,12 @@ static void free_counter_rcu(struct rcu_head *head)
         kfree(counter);
 }
 
+static void perf_pending_sync(struct perf_counter *counter);
+
 static void free_counter(struct perf_counter *counter)
 {
+        perf_pending_sync(counter);
+
         if (counter->destroy)
                 counter->destroy(counter);
 
@@ -1529,6 +1533,118 @@ static const struct file_operations perf_fops = {
 };
 
 /*
+ * Perf counter wakeup
+ *
+ * If there's data, ensure we set the poll() state and publish everything
+ * to user-space before waking everybody up.
+ */
+
+void perf_counter_wakeup(struct perf_counter *counter)
+{
+        struct perf_mmap_data *data;
+
+        rcu_read_lock();
+        data = rcu_dereference(counter->data);
+        if (data) {
+                (void)atomic_xchg(&data->wakeup, POLL_IN);
+                __perf_counter_update_userpage(counter, data);
+        }
+        rcu_read_unlock();
+
+        wake_up_all(&counter->waitq);
+}
+
+/*
+ * Pending wakeups
+ *
+ * Handle the case where we need to wakeup up from NMI (or rq->lock) context.
+ *
+ * The NMI bit means we cannot possibly take locks. Therefore, maintain a
+ * single linked list and use cmpxchg() to add entries lockless.
+ */
+
+#define PENDING_TAIL ((struct perf_wakeup_entry *)-1UL)
+
+static DEFINE_PER_CPU(struct perf_wakeup_entry *, perf_wakeup_head) = {
+        PENDING_TAIL,
+};
+
+static void perf_pending_queue(struct perf_counter *counter)
+{
+        struct perf_wakeup_entry **head;
+        struct perf_wakeup_entry *prev, *next;
+
+        if (cmpxchg(&counter->wakeup.next, NULL, PENDING_TAIL) != NULL)
+                return;
+
+        head = &get_cpu_var(perf_wakeup_head);
+
+        do {
+                prev = counter->wakeup.next = *head;
+                next = &counter->wakeup;
+        } while (cmpxchg(head, prev, next) != prev);
+
+        set_perf_counter_pending();
+
+        put_cpu_var(perf_wakeup_head);
+}
+
+static int __perf_pending_run(void)
+{
+        struct perf_wakeup_entry *list;
+        int nr = 0;
+
+        list = xchg(&__get_cpu_var(perf_wakeup_head), PENDING_TAIL);
+        while (list != PENDING_TAIL) {
+                struct perf_counter *counter = container_of(list,
+                                struct perf_counter, wakeup);
+
+                list = list->next;
+
+                counter->wakeup.next = NULL;
+                /*
+                 * Ensure we observe the unqueue before we issue the wakeup,
+                 * so that we won't be waiting forever.
+                 * -- see perf_not_pending().
+                 */
+                smp_wmb();
+
+                perf_counter_wakeup(counter);
+                nr++;
+        }
+
+        return nr;
+}
+
+static inline int perf_not_pending(struct perf_counter *counter)
+{
+        /*
+         * If we flush on whatever cpu we run, there is a chance we don't
+         * need to wait.
+         */
+        get_cpu();
+        __perf_pending_run();
+        put_cpu();
+
+        /*
+         * Ensure we see the proper queue state before going to sleep
+         * so that we do not miss the wakeup. -- see perf_pending_handle()
+         */
+        smp_rmb();
+        return counter->wakeup.next == NULL;
+}
+
+static void perf_pending_sync(struct perf_counter *counter)
+{
+        wait_event(counter->waitq, perf_not_pending(counter));
+}
+
+void perf_counter_do_pending(void)
+{
+        __perf_pending_run();
+}
+
+/*
  * Output
  */
 
@@ -1611,13 +1727,10 @@ static void perf_output_copy(struct perf_output_handle *handle,
 static void perf_output_end(struct perf_output_handle *handle, int nmi)
 {
         if (handle->wakeup) {
-                (void)atomic_xchg(&handle->data->wakeup, POLL_IN);
-                __perf_counter_update_userpage(handle->counter, handle->data);
-                if (nmi) {
-                        handle->counter->wakeup_pending = 1;
-                        set_perf_counter_pending();
-                } else
-                        wake_up(&handle->counter->waitq);
+                if (nmi)
+                        perf_pending_queue(handle->counter);
+                else
+                        perf_counter_wakeup(handle->counter);
         }
         rcu_read_unlock();
 }
@@ -2211,7 +2324,6 @@ perf_counter_alloc(struct perf_counter_hw_event *hw_event,
 
         counter->cpu                    = cpu;
         counter->hw_event               = *hw_event;
-        counter->wakeup_pending         = 0;
         counter->group_leader           = group_leader;
         counter->hw_ops                 = NULL;
         counter->ctx                    = ctx;
diff --git a/kernel/timer.c b/kernel/timer.c
index b4555568b4e4..672ca25fbc43 100644
--- a/kernel/timer.c
+++ b/kernel/timer.c
@@ -37,6 +37,7 @@
 #include <linux/delay.h>
 #include <linux/tick.h>
 #include <linux/kallsyms.h>
+#include <linux/perf_counter.h>
 
 #include <asm/uaccess.h>
 #include <asm/unistd.h>
@@ -1167,6 +1168,8 @@ static void run_timer_softirq(struct softirq_action *h)
 {
         struct tvec_base *base = __get_cpu_var(tvec_bases);
 
+        perf_counter_do_pending();
+
         hrtimer_run_pending();
 
         if (time_after_eq(jiffies, base->timer_jiffies))