Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/gerg/m68knommu
[safe/jmp/linux-2.6] / include / linux / perf_event.h
index 7b18b4f..3fd5c82 100644 (file)
@@ -203,8 +203,19 @@ struct perf_event_attr {
                                enable_on_exec :  1, /* next exec enables     */
                                task           :  1, /* trace fork/exit       */
                                watermark      :  1, /* wakeup_watermark      */
-
-                               __reserved_1   : 49;
+                               /*
+                                * precise_ip:
+                                *
+                                *  0 - SAMPLE_IP can have arbitrary skid
+                                *  1 - SAMPLE_IP must have constant skid
+                                *  2 - SAMPLE_IP requested to have 0 skid
+                                *  3 - SAMPLE_IP must have 0 skid
+                                *
+                                *  See also PERF_RECORD_MISC_EXACT_IP
+                                */
+                               precise_ip     :  2, /* skid constraint       */
+
+                               __reserved_1   : 47;
 
        union {
                __u32           wakeup_events;    /* wakeup every n events */
@@ -287,11 +298,24 @@ struct perf_event_mmap_page {
        __u64   data_tail;              /* user-space written tail */
 };
 
-#define PERF_RECORD_MISC_CPUMODE_MASK          (3 << 0)
+#define PERF_RECORD_MISC_CPUMODE_MASK          (7 << 0)
 #define PERF_RECORD_MISC_CPUMODE_UNKNOWN       (0 << 0)
 #define PERF_RECORD_MISC_KERNEL                        (1 << 0)
 #define PERF_RECORD_MISC_USER                  (2 << 0)
 #define PERF_RECORD_MISC_HYPERVISOR            (3 << 0)
+#define PERF_RECORD_MISC_GUEST_KERNEL          (4 << 0)
+#define PERF_RECORD_MISC_GUEST_USER            (5 << 0)
+
+/*
+ * Indicates that the content of PERF_SAMPLE_IP points to
+ * the actual instruction that triggered the event. See also
+ * perf_event_attr::precise_ip.
+ */
+#define PERF_RECORD_MISC_EXACT_IP              (1 << 14)
+/*
+ * Reserve the last bit to indicate some extended misc field
+ */
+#define PERF_RECORD_MISC_EXT_RESERVED          (1 << 15)
 
 struct perf_event_header {
        __u32   type;
@@ -439,6 +463,12 @@ enum perf_callchain_context {
 # include <asm/perf_event.h>
 #endif
 
+struct perf_guest_info_callbacks {
+       int (*is_in_guest) (void);
+       int (*is_user_mode) (void);
+       unsigned long (*get_guest_ip) (void);
+};
+
 #ifdef CONFIG_HAVE_HW_BREAKPOINT
 #include <asm/hw_breakpoint.h>
 #endif
@@ -452,6 +482,8 @@ enum perf_callchain_context {
 #include <linux/fs.h>
 #include <linux/pid_namespace.h>
 #include <linux/workqueue.h>
+#include <linux/ftrace.h>
+#include <linux/cpu.h>
 #include <asm/atomic.h>
 
 #define PERF_MAX_STACK_DEPTH           255
@@ -466,6 +498,17 @@ struct perf_raw_record {
        void                            *data;
 };
 
+struct perf_branch_entry {
+       __u64                           from;
+       __u64                           to;
+       __u64                           flags;
+};
+
+struct perf_branch_stack {
+       __u64                           nr;
+       struct perf_branch_entry        entries[0];
+};
+
 struct task_struct;
 
 /**
@@ -487,9 +530,8 @@ struct hw_perf_event {
                        struct hrtimer  hrtimer;
                };
 #ifdef CONFIG_HAVE_HW_BREAKPOINT
-               union { /* breakpoint */
-                       struct arch_hw_breakpoint       info;
-               };
+               /* breakpoint */
+               struct arch_hw_breakpoint       info;
 #endif
        };
        atomic64_t                      prev_count;
@@ -505,6 +547,8 @@ struct hw_perf_event {
 
 struct perf_event;
 
+#define PERF_EVENT_TXN_STARTED 1
+
 /**
  * struct pmu - generic performance monitoring unit
  */
@@ -515,6 +559,16 @@ struct pmu {
        void (*stop)                    (struct perf_event *event);
        void (*read)                    (struct perf_event *event);
        void (*unthrottle)              (struct perf_event *event);
+
+       /*
+        * group events scheduling is treated as a transaction,
+        * add group events as a whole and perform one schedulability test.
+        * If test fails, roll back the whole group
+        */
+
+       void (*start_txn)       (const struct pmu *pmu);
+       void (*cancel_txn)      (const struct pmu *pmu);
+       int  (*commit_txn)      (const struct pmu *pmu);
 };
 
 /**
@@ -570,6 +624,14 @@ enum perf_group_flag {
        PERF_GROUP_SOFTWARE = 0x1,
 };
 
+#define SWEVENT_HLIST_BITS     8
+#define SWEVENT_HLIST_SIZE     (1 << SWEVENT_HLIST_BITS)
+
+struct swevent_hlist {
+       struct hlist_head       heads[SWEVENT_HLIST_SIZE];
+       struct rcu_head         rcu_head;
+};
+
 /**
  * struct perf_event - performance event kernel representation:
  */
@@ -578,6 +640,7 @@ struct perf_event {
        struct list_head                group_entry;
        struct list_head                event_entry;
        struct list_head                sibling_list;
+       struct hlist_node               hlist_entry;
        int                             nr_siblings;
        int                             group_flags;
        struct perf_event               *group_leader;
@@ -725,6 +788,9 @@ struct perf_cpu_context {
        int                             active_oncpu;
        int                             max_pertask;
        int                             exclusive;
+       struct swevent_hlist            *swevent_hlist;
+       struct mutex                    hlist_mutex;
+       int                             hlist_refcount;
 
        /*
         * Recursion avoidance:
@@ -768,9 +834,6 @@ extern void perf_disable(void);
 extern void perf_enable(void);
 extern int perf_event_task_disable(void);
 extern int perf_event_task_enable(void);
-extern int hw_perf_group_sched_in(struct perf_event *group_leader,
-              struct perf_cpu_context *cpuctx,
-              struct perf_event_context *ctx);
 extern void perf_event_update_userpage(struct perf_event *event);
 extern int perf_event_release_kernel(struct perf_event *event);
 extern struct perf_event *
@@ -802,6 +865,13 @@ struct perf_sample_data {
        struct perf_raw_record          *raw;
 };
 
+static inline
+void perf_sample_data_init(struct perf_sample_data *data, u64 addr)
+{
+       data->addr = addr;
+       data->raw  = NULL;
+}
+
 extern void perf_output_sample(struct perf_output_handle *handle,
                               struct perf_event_header *header,
                               struct perf_sample_data *data,
@@ -834,11 +904,56 @@ extern atomic_t perf_swevent_enabled[PERF_COUNT_SW_MAX];
 
 extern void __perf_sw_event(u32, u64, int, struct pt_regs *, u64);
 
+extern void
+perf_arch_fetch_caller_regs(struct pt_regs *regs, unsigned long ip, int skip);
+
+/*
+ * Take a snapshot of the regs. Skip ip and frame pointer to
+ * the nth caller. We only need a few of the regs:
+ * - ip for PERF_SAMPLE_IP
+ * - cs for user_mode() tests
+ * - bp for callchains
+ * - eflags, for future purposes, just in case
+ */
+static inline void perf_fetch_caller_regs(struct pt_regs *regs, int skip)
+{
+       unsigned long ip;
+
+       memset(regs, 0, sizeof(*regs));
+
+       switch (skip) {
+       case 1 :
+               ip = CALLER_ADDR0;
+               break;
+       case 2 :
+               ip = CALLER_ADDR1;
+               break;
+       case 3 :
+               ip = CALLER_ADDR2;
+               break;
+       case 4:
+               ip = CALLER_ADDR3;
+               break;
+       /* No need to support further for now */
+       default:
+               ip = 0;
+       }
+
+       return perf_arch_fetch_caller_regs(regs, ip, skip);
+}
+
 static inline void
 perf_sw_event(u32 event_id, u64 nr, int nmi, struct pt_regs *regs, u64 addr)
 {
-       if (atomic_read(&perf_swevent_enabled[event_id]))
+       if (atomic_read(&perf_swevent_enabled[event_id])) {
+               struct pt_regs hot_regs;
+
+               if (!regs) {
+                       perf_fetch_caller_regs(&hot_regs, 1);
+                       regs = &hot_regs;
+               }
                __perf_sw_event(event_id, nr, nmi, regs, addr);
+       }
 }
 
 extern void __perf_event_mmap(struct vm_area_struct *vma);
@@ -849,6 +964,10 @@ static inline void perf_event_mmap(struct vm_area_struct *vma)
                __perf_event_mmap(vma);
 }
 
+extern struct perf_guest_info_callbacks *perf_guest_cbs;
+extern int perf_register_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
+extern int perf_unregister_guest_info_callbacks(struct perf_guest_info_callbacks *callbacks);
+
 extern void perf_event_comm(struct task_struct *tsk);
 extern void perf_event_fork(struct task_struct *tsk);
 
@@ -858,8 +977,24 @@ extern int sysctl_perf_event_paranoid;
 extern int sysctl_perf_event_mlock;
 extern int sysctl_perf_event_sample_rate;
 
+static inline bool perf_paranoid_tracepoint_raw(void)
+{
+       return sysctl_perf_event_paranoid > -1;
+}
+
+static inline bool perf_paranoid_cpu(void)
+{
+       return sysctl_perf_event_paranoid > 0;
+}
+
+static inline bool perf_paranoid_kernel(void)
+{
+       return sysctl_perf_event_paranoid > 1;
+}
+
 extern void perf_event_init(void);
-extern void perf_tp_event(int event_id, u64 addr, u64 count, void *record, int entry_size);
+extern void perf_tp_event(int event_id, u64 addr, u64 count, void *record,
+                         int entry_size, struct pt_regs *regs);
 extern void perf_bp_event(struct perf_event *event, void *data);
 
 #ifndef perf_misc_flags
@@ -902,6 +1037,11 @@ perf_sw_event(u32 event_id, u64 nr, int nmi,
 static inline void
 perf_bp_event(struct perf_event *event, void *data)                    { }
 
+static inline int perf_register_guest_info_callbacks
+(struct perf_guest_info_callbacks *callbacks) { return 0; }
+static inline int perf_unregister_guest_info_callbacks
+(struct perf_guest_info_callbacks *callbacks) { return 0; }
+
 static inline void perf_event_mmap(struct vm_area_struct *vma)         { }
 static inline void perf_event_comm(struct task_struct *tsk)            { }
 static inline void perf_event_fork(struct task_struct *tsk)            { }
@@ -915,5 +1055,21 @@ static inline void perf_event_disable(struct perf_event *event)           { }
 #define perf_output_put(handle, x) \
        perf_output_copy((handle), &(x), sizeof(x))
 
+/*
+ * This has to have a higher priority than migration_notifier in sched.c.
+ */
+#define perf_cpu_notifier(fn)                                  \
+do {                                                           \
+       static struct notifier_block fn##_nb __cpuinitdata =    \
+               { .notifier_call = fn, .priority = 20 };        \
+       fn(&fn##_nb, (unsigned long)CPU_UP_PREPARE,             \
+               (void *)(unsigned long)smp_processor_id());     \
+       fn(&fn##_nb, (unsigned long)CPU_STARTING,               \
+               (void *)(unsigned long)smp_processor_id());     \
+       fn(&fn##_nb, (unsigned long)CPU_ONLINE,                 \
+               (void *)(unsigned long)smp_processor_id());     \
+       register_cpu_notifier(&fn##_nb);                        \
+} while (0)
+
 #endif /* __KERNEL__ */
 #endif /* _LINUX_PERF_EVENT_H */