]> git.karo-electronics.de Git - mv-sheeva.git/blob - kernel/trace/trace_event_perf.c
perf: Rework the PMU methods
[mv-sheeva.git] / kernel / trace / trace_event_perf.c
1 /*
2  * trace event based perf event profiling/tracing
3  *
4  * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra <pzijlstr@redhat.com>
5  * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com>
6  */
7
8 #include <linux/module.h>
9 #include <linux/kprobes.h>
10 #include "trace.h"
11
12 static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS];
13
14 /*
15  * Force it to be aligned to unsigned long to avoid misaligned accesses
16  * suprises
17  */
18 typedef typeof(unsigned long [PERF_MAX_TRACE_SIZE / sizeof(unsigned long)])
19         perf_trace_t;
20
21 /* Count the events in use (per event id, not per instance) */
22 static int      total_ref_count;
23
24 static int perf_trace_event_init(struct ftrace_event_call *tp_event,
25                                  struct perf_event *p_event)
26 {
27         struct hlist_head __percpu *list;
28         int ret = -ENOMEM;
29         int cpu;
30
31         p_event->tp_event = tp_event;
32         if (tp_event->perf_refcount++ > 0)
33                 return 0;
34
35         list = alloc_percpu(struct hlist_head);
36         if (!list)
37                 goto fail;
38
39         for_each_possible_cpu(cpu)
40                 INIT_HLIST_HEAD(per_cpu_ptr(list, cpu));
41
42         tp_event->perf_events = list;
43
44         if (!total_ref_count) {
45                 char __percpu *buf;
46                 int i;
47
48                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
49                         buf = (char __percpu *)alloc_percpu(perf_trace_t);
50                         if (!buf)
51                                 goto fail;
52
53                         perf_trace_buf[i] = buf;
54                 }
55         }
56
57         ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER);
58         if (ret)
59                 goto fail;
60
61         total_ref_count++;
62         return 0;
63
64 fail:
65         if (!total_ref_count) {
66                 int i;
67
68                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
69                         free_percpu(perf_trace_buf[i]);
70                         perf_trace_buf[i] = NULL;
71                 }
72         }
73
74         if (!--tp_event->perf_refcount) {
75                 free_percpu(tp_event->perf_events);
76                 tp_event->perf_events = NULL;
77         }
78
79         return ret;
80 }
81
82 int perf_trace_init(struct perf_event *p_event)
83 {
84         struct ftrace_event_call *tp_event;
85         int event_id = p_event->attr.config;
86         int ret = -EINVAL;
87
88         mutex_lock(&event_mutex);
89         list_for_each_entry(tp_event, &ftrace_events, list) {
90                 if (tp_event->event.type == event_id &&
91                     tp_event->class && tp_event->class->reg &&
92                     try_module_get(tp_event->mod)) {
93                         ret = perf_trace_event_init(tp_event, p_event);
94                         if (ret)
95                                 module_put(tp_event->mod);
96                         break;
97                 }
98         }
99         mutex_unlock(&event_mutex);
100
101         return ret;
102 }
103
104 int perf_trace_add(struct perf_event *p_event, int flags)
105 {
106         struct ftrace_event_call *tp_event = p_event->tp_event;
107         struct hlist_head __percpu *pcpu_list;
108         struct hlist_head *list;
109
110         pcpu_list = tp_event->perf_events;
111         if (WARN_ON_ONCE(!pcpu_list))
112                 return -EINVAL;
113
114         if (!(flags & PERF_EF_START))
115                 p_event->hw.state = PERF_HES_STOPPED;
116
117         list = this_cpu_ptr(pcpu_list);
118         hlist_add_head_rcu(&p_event->hlist_entry, list);
119
120         return 0;
121 }
122
123 void perf_trace_del(struct perf_event *p_event, int flags)
124 {
125         hlist_del_rcu(&p_event->hlist_entry);
126 }
127
128 void perf_trace_destroy(struct perf_event *p_event)
129 {
130         struct ftrace_event_call *tp_event = p_event->tp_event;
131         int i;
132
133         mutex_lock(&event_mutex);
134         if (--tp_event->perf_refcount > 0)
135                 goto out;
136
137         tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER);
138
139         /*
140          * Ensure our callback won't be called anymore. The buffers
141          * will be freed after that.
142          */
143         tracepoint_synchronize_unregister();
144
145         free_percpu(tp_event->perf_events);
146         tp_event->perf_events = NULL;
147
148         if (!--total_ref_count) {
149                 for (i = 0; i < PERF_NR_CONTEXTS; i++) {
150                         free_percpu(perf_trace_buf[i]);
151                         perf_trace_buf[i] = NULL;
152                 }
153         }
154 out:
155         module_put(tp_event->mod);
156         mutex_unlock(&event_mutex);
157 }
158
159 __kprobes void *perf_trace_buf_prepare(int size, unsigned short type,
160                                        struct pt_regs *regs, int *rctxp)
161 {
162         struct trace_entry *entry;
163         unsigned long flags;
164         char *raw_data;
165         int pc;
166
167         BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long));
168
169         pc = preempt_count();
170
171         *rctxp = perf_swevent_get_recursion_context();
172         if (*rctxp < 0)
173                 return NULL;
174
175         raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]);
176
177         /* zero the dead bytes from align to not leak stack to user */
178         memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64));
179
180         entry = (struct trace_entry *)raw_data;
181         local_save_flags(flags);
182         tracing_generic_entry_update(entry, flags, pc);
183         entry->type = type;
184
185         return raw_data;
186 }
187 EXPORT_SYMBOL_GPL(perf_trace_buf_prepare);