perf counters: add support for group counters
authorIngo Molnar <mingo@elte.hu>
Thu, 11 Dec 2008 07:38:42 +0000 (08:38 +0100)
committerIngo Molnar <mingo@elte.hu>
Thu, 11 Dec 2008 14:45:49 +0000 (15:45 +0100)
Impact: add group counters

This patch adds the "counter groups" abstraction.

Groups of counters behave much like normal 'single' counters, with a
few semantic and behavioral extensions on top of that.

A counter group is created by creating a new counter with the open()
syscall's group-leader group_fd file descriptor parameter pointing
to another, already existing counter.

Groups of counters are scheduled in and out in one atomic group, and
they are also roundrobin-scheduled atomically.

Counters that are member of a group can also record events with an
(atomic) extended timestamp that extends to all members of the group,
if the record type is set to PERF_RECORD_GROUP.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86/kernel/cpu/perf_counter.c
include/linux/perf_counter.h
kernel/perf_counter.c

index ef1936a871aa044a25cbe1e830fa3fe3c9fd5c6b..54b4ad0cce6836499c5c67de7d4afdad8062547b 100644 (file)
@@ -346,18 +346,22 @@ static void perf_save_and_restart(struct perf_counter *counter)
 }
 
 static void
-perf_handle_group(struct perf_counter *leader, u64 *status, u64 *overflown)
+perf_handle_group(struct perf_counter *sibling, u64 *status, u64 *overflown)
 {
-       struct perf_counter_context *ctx = leader->ctx;
-       struct perf_counter *counter;
+       struct perf_counter *counter, *group_leader = sibling->group_leader;
        int bit;
 
-       list_for_each_entry(counter, &ctx->counters, list) {
-               if (counter->hw_event.record_type != PERF_RECORD_SIMPLE ||
-                   counter == leader)
-                       continue;
+       /*
+        * Store the counter's own timestamp first:
+        */
+       perf_store_irq_data(sibling, sibling->hw_event.type);
+       perf_store_irq_data(sibling, atomic64_counter_read(sibling));
 
-               if (counter->active) {
+       /*
+        * Then store sibling timestamps (if any):
+        */
+       list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
+               if (!counter->active) {
                        /*
                         * When counter was not in the overflow mask, we have to
                         * read it from hardware. We read it as well, when it
@@ -371,8 +375,8 @@ perf_handle_group(struct perf_counter *leader, u64 *status, u64 *overflown)
                                perf_save_and_restart(counter);
                        }
                }
-               perf_store_irq_data(leader, counter->hw_event.type);
-               perf_store_irq_data(leader, atomic64_counter_read(counter));
+               perf_store_irq_data(sibling, counter->hw_event.type);
+               perf_store_irq_data(sibling, atomic64_counter_read(counter));
        }
 }
 
@@ -416,10 +420,6 @@ again:
                        perf_store_irq_data(counter, instruction_pointer(regs));
                        break;
                case PERF_RECORD_GROUP:
-                       perf_store_irq_data(counter,
-                                           counter->hw_event.type);
-                       perf_store_irq_data(counter,
-                                           atomic64_counter_read(counter));
                        perf_handle_group(counter, &status, &ack);
                        break;
                }
index a2b4852e2d70e1899f01b729b911609b473d656f..7af7d8965460a0a911a7d6f2b77ae6203aa944e6 100644 (file)
@@ -117,7 +117,10 @@ struct perf_data {
  * struct perf_counter - performance counter kernel representation:
  */
 struct perf_counter {
-       struct list_head                list;
+       struct list_head                list_entry;
+       struct list_head                sibling_list;
+       struct perf_counter             *group_leader;
+
        int                             active;
 #if BITS_PER_LONG == 64
        atomic64_t                      count;
@@ -158,7 +161,8 @@ struct perf_counter_context {
         * Protect the list of counters:
         */
        spinlock_t              lock;
-       struct list_head        counters;
+
+       struct list_head        counter_list;
        int                     nr_counters;
        int                     nr_active;
        struct task_struct      *task;
index 0d323ceda3a40d8f93c2c40b0e4cf382a775d96b..fa59fe8c02d5b7d393f6fef36ca97e9a6bd84a74 100644 (file)
@@ -10,6 +10,7 @@
 #include <linux/fs.h>
 #include <linux/cpu.h>
 #include <linux/smp.h>
+#include <linux/file.h>
 #include <linux/poll.h>
 #include <linux/sysfs.h>
 #include <linux/ptrace.h>
@@ -55,7 +56,7 @@ void __weak hw_perf_counter_setup(void) { }
  * Read the cached counter in counter safe against cross CPU / NMI
  * modifications. 64 bit version - no complications.
  */
-static inline u64 perf_read_counter_safe(struct perf_counter *counter)
+static inline u64 perf_counter_read_safe(struct perf_counter *counter)
 {
        return (u64) atomic64_read(&counter->count);
 }
@@ -66,7 +67,7 @@ static inline u64 perf_read_counter_safe(struct perf_counter *counter)
  * Read the cached counter in counter safe against cross CPU / NMI
  * modifications. 32 bit version.
  */
-static u64 perf_read_counter_safe(struct perf_counter *counter)
+static u64 perf_counter_read_safe(struct perf_counter *counter)
 {
        u32 cntl, cnth;
 
@@ -83,13 +84,55 @@ static u64 perf_read_counter_safe(struct perf_counter *counter)
 
 #endif
 
+static void
+list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
+{
+       struct perf_counter *group_leader = counter->group_leader;
+
+       /*
+        * Depending on whether it is a standalone or sibling counter,
+        * add it straight to the context's counter list, or to the group
+        * leader's sibling list:
+        */
+       if (counter->group_leader == counter)
+               list_add_tail(&counter->list_entry, &ctx->counter_list);
+       else
+               list_add_tail(&counter->list_entry, &group_leader->sibling_list);
+}
+
+static void
+list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
+{
+       struct perf_counter *sibling, *tmp;
+
+       list_del_init(&counter->list_entry);
+
+       if (list_empty(&counter->sibling_list))
+               return;
+
+       /*
+        * If this was a group counter with sibling counters then
+        * upgrade the siblings to singleton counters by adding them
+        * to the context list directly:
+        */
+       list_for_each_entry_safe(sibling, tmp,
+                                &counter->sibling_list, list_entry) {
+
+               list_del_init(&sibling->list_entry);
+               list_add_tail(&sibling->list_entry, &ctx->counter_list);
+               WARN_ON_ONCE(!sibling->group_leader);
+               WARN_ON_ONCE(sibling->group_leader == sibling);
+               sibling->group_leader = sibling;
+       }
+}
+
 /*
  * Cross CPU call to remove a performance counter
  *
  * We disable the counter on the hardware level first. After that we
  * remove it from the context list.
  */
-static void __perf_remove_from_context(void *info)
+static void __perf_counter_remove_from_context(void *info)
 {
        struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
        struct perf_counter *counter = info;
@@ -119,7 +162,7 @@ static void __perf_remove_from_context(void *info)
         * counters on a global level. NOP for non NMI based counters.
         */
        hw_perf_disable_all();
-       list_del_init(&counter->list);
+       list_del_counter(counter, ctx);
        hw_perf_enable_all();
 
        if (!ctx->task) {
@@ -144,7 +187,7 @@ static void __perf_remove_from_context(void *info)
  * CPU counters are removed with a smp call. For task counters we only
  * call when the task is on a CPU.
  */
-static void perf_remove_from_context(struct perf_counter *counter)
+static void perf_counter_remove_from_context(struct perf_counter *counter)
 {
        struct perf_counter_context *ctx = counter->ctx;
        struct task_struct *task = ctx->task;
@@ -155,32 +198,32 @@ static void perf_remove_from_context(struct perf_counter *counter)
                 * the removal is always sucessful.
                 */
                smp_call_function_single(counter->cpu,
-                                        __perf_remove_from_context,
+                                        __perf_counter_remove_from_context,
                                         counter, 1);
                return;
        }
 
 retry:
-       task_oncpu_function_call(task, __perf_remove_from_context,
+       task_oncpu_function_call(task, __perf_counter_remove_from_context,
                                 counter);
 
        spin_lock_irq(&ctx->lock);
        /*
         * If the context is active we need to retry the smp call.
         */
-       if (ctx->nr_active && !list_empty(&counter->list)) {
+       if (ctx->nr_active && !list_empty(&counter->list_entry)) {
                spin_unlock_irq(&ctx->lock);
                goto retry;
        }
 
        /*
         * The lock prevents that this context is scheduled in so we
-        * can remove the counter safely, if it the call above did not
+        * can remove the counter safely, if the call above did not
         * succeed.
         */
-       if (!list_empty(&counter->list)) {
+       if (!list_empty(&counter->list_entry)) {
                ctx->nr_counters--;
-               list_del_init(&counter->list);
+               list_del_counter(counter, ctx);
                counter->task = NULL;
        }
        spin_unlock_irq(&ctx->lock);
@@ -211,7 +254,7 @@ static void __perf_install_in_context(void *info)
         * counters on a global level. NOP for non NMI based counters.
         */
        hw_perf_disable_all();
-       list_add_tail(&counter->list, &ctx->counters);
+       list_add_counter(counter, ctx);
        hw_perf_enable_all();
 
        ctx->nr_counters++;
@@ -268,7 +311,7 @@ retry:
         * If the context is active and the counter has not been added
         * we need to retry the smp call.
         */
-       if (ctx->nr_active && list_empty(&counter->list)) {
+       if (ctx->nr_active && list_empty(&counter->list_entry)) {
                spin_unlock_irq(&ctx->lock);
                goto retry;
        }
@@ -278,13 +321,45 @@ retry:
         * can add the counter safely, if it the call above did not
         * succeed.
         */
-       if (list_empty(&counter->list)) {
-               list_add_tail(&counter->list, &ctx->counters);
+       if (list_empty(&counter->list_entry)) {
+               list_add_counter(counter, ctx);
                ctx->nr_counters++;
        }
        spin_unlock_irq(&ctx->lock);
 }
 
+static void
+counter_sched_out(struct perf_counter *counter,
+                 struct perf_cpu_context *cpuctx,
+                 struct perf_counter_context *ctx)
+{
+       if (!counter->active)
+               return;
+
+       hw_perf_counter_disable(counter);
+       counter->active =  0;
+       counter->oncpu  = -1;
+
+       cpuctx->active_oncpu--;
+       ctx->nr_active--;
+}
+
+static void
+group_sched_out(struct perf_counter *group_counter,
+               struct perf_cpu_context *cpuctx,
+               struct perf_counter_context *ctx)
+{
+       struct perf_counter *counter;
+
+       counter_sched_out(group_counter, cpuctx, ctx);
+
+       /*
+        * Schedule out siblings (if any):
+        */
+       list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
+               counter_sched_out(counter, cpuctx, ctx);
+}
+
 /*
  * Called from scheduler to remove the counters of the current task,
  * with interrupts disabled.
@@ -306,21 +381,48 @@ void perf_counter_task_sched_out(struct task_struct *task, int cpu)
                return;
 
        spin_lock(&ctx->lock);
-       list_for_each_entry(counter, &ctx->counters, list) {
-               if (!ctx->nr_active)
-                       break;
-               if (counter->active) {
-                       hw_perf_counter_disable(counter);
-                       counter->active = 0;
-                       counter->oncpu = -1;
-                       ctx->nr_active--;
-                       cpuctx->active_oncpu--;
-               }
+       if (ctx->nr_active) {
+               list_for_each_entry(counter, &ctx->counter_list, list_entry)
+                       group_sched_out(counter, cpuctx, ctx);
        }
        spin_unlock(&ctx->lock);
        cpuctx->task_ctx = NULL;
 }
 
+static void
+counter_sched_in(struct perf_counter *counter,
+                struct perf_cpu_context *cpuctx,
+                struct perf_counter_context *ctx,
+                int cpu)
+{
+       if (!counter->active)
+               return;
+
+       hw_perf_counter_enable(counter);
+       counter->active = 1;
+       counter->oncpu = cpu;   /* TODO: put 'cpu' into cpuctx->cpu */
+
+       cpuctx->active_oncpu++;
+       ctx->nr_active++;
+}
+
+static void
+group_sched_in(struct perf_counter *group_counter,
+              struct perf_cpu_context *cpuctx,
+              struct perf_counter_context *ctx,
+              int cpu)
+{
+       struct perf_counter *counter;
+
+       counter_sched_in(group_counter, cpuctx, ctx, cpu);
+
+       /*
+        * Schedule in siblings as one group (if any):
+        */
+       list_for_each_entry(counter, &group_counter->sibling_list, list_entry)
+               counter_sched_in(counter, cpuctx, ctx, cpu);
+}
+
 /*
  * Called from scheduler to add the counters of the current task
  * with interrupts disabled.
@@ -342,19 +444,21 @@ void perf_counter_task_sched_in(struct task_struct *task, int cpu)
                return;
 
        spin_lock(&ctx->lock);
-       list_for_each_entry(counter, &ctx->counters, list) {
+       list_for_each_entry(counter, &ctx->counter_list, list_entry) {
                if (ctx->nr_active == cpuctx->max_pertask)
                        break;
+
+               /*
+                * Listen to the 'cpu' scheduling filter constraint
+                * of counters:
+                */
                if (counter->cpu != -1 && counter->cpu != cpu)
                        continue;
 
-               hw_perf_counter_enable(counter);
-               counter->active = 1;
-               counter->oncpu = cpu;
-               ctx->nr_active++;
-               cpuctx->active_oncpu++;
+               group_sched_in(counter, cpuctx, ctx, cpu);
        }
        spin_unlock(&ctx->lock);
+
        cpuctx->task_ctx = ctx;
 }
 
@@ -371,12 +475,12 @@ void perf_counter_task_tick(struct task_struct *curr, int cpu)
        spin_lock(&ctx->lock);
 
        /*
-        * Rotate the first entry last:
+        * Rotate the first entry last (works just fine for group counters too):
         */
        hw_perf_disable_all();
-       list_for_each_entry(counter, &ctx->counters, list) {
-               list_del(&counter->list);
-               list_add_tail(&counter->list, &ctx->counters);
+       list_for_each_entry(counter, &ctx->counter_list, list_entry) {
+               list_del(&counter->list_entry);
+               list_add_tail(&counter->list_entry, &ctx->counter_list);
                break;
        }
        hw_perf_enable_all();
@@ -386,17 +490,24 @@ void perf_counter_task_tick(struct task_struct *curr, int cpu)
        perf_counter_task_sched_in(curr, cpu);
 }
 
+/*
+ * Initialize the perf_counter context in a task_struct:
+ */
+static void
+__perf_counter_init_context(struct perf_counter_context *ctx,
+                           struct task_struct *task)
+{
+       spin_lock_init(&ctx->lock);
+       INIT_LIST_HEAD(&ctx->counter_list);
+       ctx->nr_counters        = 0;
+       ctx->task               = task;
+}
 /*
  * Initialize the perf_counter context in task_struct
  */
 void perf_counter_init_task(struct task_struct *task)
 {
-       struct perf_counter_context *ctx = &task->perf_counter_ctx;
-
-       spin_lock_init(&ctx->lock);
-       INIT_LIST_HEAD(&ctx->counters);
-       ctx->nr_counters = 0;
-       ctx->task = task;
+       __perf_counter_init_context(&task->perf_counter_ctx, task);
 }
 
 /*
@@ -407,7 +518,7 @@ static void __hw_perf_counter_read(void *info)
        hw_perf_counter_read(info);
 }
 
-static u64 perf_read_counter(struct perf_counter *counter)
+static u64 perf_counter_read(struct perf_counter *counter)
 {
        /*
         * If counter is enabled and currently active on a CPU, update the
@@ -418,7 +529,7 @@ static u64 perf_read_counter(struct perf_counter *counter)
                                         __hw_perf_counter_read, counter, 1);
        }
 
-       return perf_read_counter_safe(counter);
+       return perf_counter_read_safe(counter);
 }
 
 /*
@@ -555,7 +666,7 @@ static int perf_release(struct inode *inode, struct file *file)
 
        mutex_lock(&counter->mutex);
 
-       perf_remove_from_context(counter);
+       perf_counter_remove_from_context(counter);
        put_context(ctx);
 
        mutex_unlock(&counter->mutex);
@@ -577,7 +688,7 @@ perf_read_hw(struct perf_counter *counter, char __user *buf, size_t count)
                return -EINVAL;
 
        mutex_lock(&counter->mutex);
-       cntval = perf_read_counter(counter);
+       cntval = perf_counter_read(counter);
        mutex_unlock(&counter->mutex);
 
        return put_user(cntval, (u64 __user *) buf) ? -EFAULT : sizeof(cntval);
@@ -707,15 +818,25 @@ static const struct file_operations perf_fops = {
  * Allocate and initialize a counter structure
  */
 static struct perf_counter *
-perf_counter_alloc(struct perf_counter_hw_event *hw_event, int cpu)
+perf_counter_alloc(struct perf_counter_hw_event *hw_event,
+                  int cpu,
+                  struct perf_counter *group_leader)
 {
        struct perf_counter *counter = kzalloc(sizeof(*counter), GFP_KERNEL);
 
        if (!counter)
                return NULL;
 
+       /*
+        * Single counters are their own group leaders, with an
+        * empty sibling list:
+        */
+       if (!group_leader)
+               group_leader = counter;
+
        mutex_init(&counter->mutex);
-       INIT_LIST_HEAD(&counter->list);
+       INIT_LIST_HEAD(&counter->list_entry);
+       INIT_LIST_HEAD(&counter->sibling_list);
        init_waitqueue_head(&counter->waitq);
 
        counter->irqdata                = &counter->data[0];
@@ -723,6 +844,7 @@ perf_counter_alloc(struct perf_counter_hw_event *hw_event, int cpu)
        counter->cpu                    = cpu;
        counter->hw_event               = *hw_event;
        counter->wakeup_pending         = 0;
+       counter->group_leader           = group_leader;
 
        return counter;
 }
@@ -743,20 +865,45 @@ asmlinkage int sys_perf_counter_open(
        int                             group_fd)
 
 {
-       struct perf_counter_context *ctx;
+       struct perf_counter *counter, *group_leader;
        struct perf_counter_hw_event hw_event;
-       struct perf_counter *counter;
+       struct perf_counter_context *ctx;
+       struct file *group_file = NULL;
+       int fput_needed = 0;
        int ret;
 
        if (copy_from_user(&hw_event, hw_event_uptr, sizeof(hw_event)) != 0)
                return -EFAULT;
 
+       /*
+        * Look up the group leader:
+        */
+       group_leader = NULL;
+       if (group_fd != -1) {
+               ret = -EINVAL;
+               group_file = fget_light(group_fd, &fput_needed);
+               if (!group_file)
+                       goto out_fput;
+               if (group_file->f_op != &perf_fops)
+                       goto out_fput;
+
+               group_leader = group_file->private_data;
+               /*
+                * Do not allow a recursive hierarchy:
+                */
+               if (group_leader->group_leader)
+                       goto out_fput;
+       }
+
+       /*
+        * Get the target context (task or percpu):
+        */
        ctx = find_get_context(pid, cpu);
        if (IS_ERR(ctx))
                return PTR_ERR(ctx);
 
        ret = -ENOMEM;
-       counter = perf_counter_alloc(&hw_event, cpu);
+       counter = perf_counter_alloc(&hw_event, cpu, group_leader);
        if (!counter)
                goto err_put_context;
 
@@ -770,11 +917,14 @@ asmlinkage int sys_perf_counter_open(
        if (ret < 0)
                goto err_remove_free_put_context;
 
+out_fput:
+       fput_light(group_file, fput_needed);
+
        return ret;
 
 err_remove_free_put_context:
        mutex_lock(&counter->mutex);
-       perf_remove_from_context(counter);
+       perf_counter_remove_from_context(counter);
        mutex_unlock(&counter->mutex);
 
 err_free_put_context:
@@ -783,40 +933,40 @@ err_free_put_context:
 err_put_context:
        put_context(ctx);
 
-       return ret;
+       goto out_fput;
 }
 
-static void __cpuinit perf_init_cpu(int cpu)
+static void __cpuinit perf_counter_init_cpu(int cpu)
 {
-       struct perf_cpu_context *ctx;
+       struct perf_cpu_context *cpuctx;
 
-       ctx = &per_cpu(perf_cpu_context, cpu);
-       spin_lock_init(&ctx->ctx.lock);
-       INIT_LIST_HEAD(&ctx->ctx.counters);
+       cpuctx = &per_cpu(perf_cpu_context, cpu);
+       __perf_counter_init_context(&cpuctx->ctx, NULL);
 
        mutex_lock(&perf_resource_mutex);
-       ctx->max_pertask = perf_max_counters - perf_reserved_percpu;
+       cpuctx->max_pertask = perf_max_counters - perf_reserved_percpu;
        mutex_unlock(&perf_resource_mutex);
+
        hw_perf_counter_setup();
 }
 
 #ifdef CONFIG_HOTPLUG_CPU
-static void __perf_exit_cpu(void *info)
+static void __perf_counter_exit_cpu(void *info)
 {
        struct perf_cpu_context *cpuctx = &__get_cpu_var(perf_cpu_context);
        struct perf_counter_context *ctx = &cpuctx->ctx;
        struct perf_counter *counter, *tmp;
 
-       list_for_each_entry_safe(counter, tmp, &ctx->counters, list)
-               __perf_remove_from_context(counter);
+       list_for_each_entry_safe(counter, tmp, &ctx->counter_list, list_entry)
+               __perf_counter_remove_from_context(counter);
 
 }
-static void perf_exit_cpu(int cpu)
+static void perf_counter_exit_cpu(int cpu)
 {
-       smp_call_function_single(cpu, __perf_exit_cpu, NULL, 1);
+       smp_call_function_single(cpu, __perf_counter_exit_cpu, NULL, 1);
 }
 #else
-static inline void perf_exit_cpu(int cpu) { }
+static inline void perf_counter_exit_cpu(int cpu) { }
 #endif
 
 static int __cpuinit
@@ -828,12 +978,12 @@ perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
 
        case CPU_UP_PREPARE:
        case CPU_UP_PREPARE_FROZEN:
-               perf_init_cpu(cpu);
+               perf_counter_init_cpu(cpu);
                break;
 
        case CPU_DOWN_PREPARE:
        case CPU_DOWN_PREPARE_FROZEN:
-               perf_exit_cpu(cpu);
+               perf_counter_exit_cpu(cpu);
                break;
 
        default: