]>
Commit | Line | Data |
---|---|---|
ac199db0 | 1 | /* |
97d5a220 | 2 | * trace event based perf event profiling/tracing |
ac199db0 PZ |
3 | * |
4 | * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra <pzijlstr@redhat.com> | |
c530665c | 5 | * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com> |
ac199db0 PZ |
6 | */ |
7 | ||
558e6547 | 8 | #include <linux/module.h> |
430ad5a6 | 9 | #include <linux/kprobes.h> |
ac199db0 PZ |
10 | #include "trace.h" |
11 | ||
6016ee13 | 12 | static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS]; |
20ab4425 | 13 | |
eb1e7961 FW |
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; | |
ce71b9df | 20 | |
20ab4425 | 21 | /* Count the events in use (per event id, not per instance) */ |
97d5a220 | 22 | static int total_ref_count; |
20ab4425 | 23 | |
61c32659 FW |
24 | static int perf_trace_event_perm(struct ftrace_event_call *tp_event, |
25 | struct perf_event *p_event) | |
26 | { | |
27 | /* No tracing, just counting, so no obvious leak */ | |
28 | if (!(p_event->attr.sample_type & PERF_SAMPLE_RAW)) | |
29 | return 0; | |
30 | ||
31 | /* Some events are ok to be traced by non-root users... */ | |
32 | if (p_event->attach_state == PERF_ATTACH_TASK) { | |
33 | if (tp_event->flags & TRACE_EVENT_FL_CAP_ANY) | |
34 | return 0; | |
35 | } | |
36 | ||
37 | /* | |
38 | * ...otherwise raw tracepoint data can be a severe data leak, | |
39 | * only allow root to have these. | |
40 | */ | |
41 | if (perf_paranoid_tracepoint_raw() && !capable(CAP_SYS_ADMIN)) | |
42 | return -EPERM; | |
43 | ||
44 | return 0; | |
45 | } | |
46 | ||
1c024eca PZ |
47 | static int perf_trace_event_init(struct ftrace_event_call *tp_event, |
48 | struct perf_event *p_event) | |
e5e25cf4 | 49 | { |
6016ee13 | 50 | struct hlist_head __percpu *list; |
61c32659 | 51 | int ret; |
1c024eca | 52 | int cpu; |
20ab4425 | 53 | |
61c32659 FW |
54 | ret = perf_trace_event_perm(tp_event, p_event); |
55 | if (ret) | |
56 | return ret; | |
57 | ||
1c024eca PZ |
58 | p_event->tp_event = tp_event; |
59 | if (tp_event->perf_refcount++ > 0) | |
e5e25cf4 FW |
60 | return 0; |
61 | ||
61c32659 FW |
62 | ret = -ENOMEM; |
63 | ||
1c024eca PZ |
64 | list = alloc_percpu(struct hlist_head); |
65 | if (!list) | |
66 | goto fail; | |
67 | ||
68 | for_each_possible_cpu(cpu) | |
69 | INIT_HLIST_HEAD(per_cpu_ptr(list, cpu)); | |
20ab4425 | 70 | |
1c024eca | 71 | tp_event->perf_events = list; |
e5e25cf4 | 72 | |
97d5a220 | 73 | if (!total_ref_count) { |
6016ee13 | 74 | char __percpu *buf; |
b7e2ecef | 75 | int i; |
20ab4425 | 76 | |
7ae07ea3 | 77 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
6016ee13 | 78 | buf = (char __percpu *)alloc_percpu(perf_trace_t); |
b7e2ecef | 79 | if (!buf) |
1c024eca | 80 | goto fail; |
20ab4425 | 81 | |
1c024eca | 82 | perf_trace_buf[i] = buf; |
b7e2ecef | 83 | } |
20ab4425 FW |
84 | } |
85 | ||
a1d0ce82 | 86 | ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER); |
1c024eca PZ |
87 | if (ret) |
88 | goto fail; | |
20ab4425 | 89 | |
1c024eca PZ |
90 | total_ref_count++; |
91 | return 0; | |
92 | ||
93 | fail: | |
97d5a220 | 94 | if (!total_ref_count) { |
b7e2ecef PZ |
95 | int i; |
96 | ||
7ae07ea3 | 97 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
b7e2ecef PZ |
98 | free_percpu(perf_trace_buf[i]); |
99 | perf_trace_buf[i] = NULL; | |
100 | } | |
fe8e5b5a | 101 | } |
1c024eca PZ |
102 | |
103 | if (!--tp_event->perf_refcount) { | |
104 | free_percpu(tp_event->perf_events); | |
105 | tp_event->perf_events = NULL; | |
fe8e5b5a | 106 | } |
20ab4425 FW |
107 | |
108 | return ret; | |
e5e25cf4 FW |
109 | } |
110 | ||
1c024eca | 111 | int perf_trace_init(struct perf_event *p_event) |
ac199db0 | 112 | { |
1c024eca PZ |
113 | struct ftrace_event_call *tp_event; |
114 | int event_id = p_event->attr.config; | |
20c8928a | 115 | int ret = -EINVAL; |
ac199db0 | 116 | |
20c8928a | 117 | mutex_lock(&event_mutex); |
1c024eca | 118 | list_for_each_entry(tp_event, &ftrace_events, list) { |
ff5f149b | 119 | if (tp_event->event.type == event_id && |
a1d0ce82 | 120 | tp_event->class && tp_event->class->reg && |
1c024eca PZ |
121 | try_module_get(tp_event->mod)) { |
122 | ret = perf_trace_event_init(tp_event, p_event); | |
9cb627d5 LZ |
123 | if (ret) |
124 | module_put(tp_event->mod); | |
20c8928a LZ |
125 | break; |
126 | } | |
ac199db0 | 127 | } |
20c8928a | 128 | mutex_unlock(&event_mutex); |
ac199db0 | 129 | |
20c8928a | 130 | return ret; |
ac199db0 PZ |
131 | } |
132 | ||
a4eaf7f1 | 133 | int perf_trace_add(struct perf_event *p_event, int flags) |
e5e25cf4 | 134 | { |
1c024eca | 135 | struct ftrace_event_call *tp_event = p_event->tp_event; |
6016ee13 | 136 | struct hlist_head __percpu *pcpu_list; |
1c024eca | 137 | struct hlist_head *list; |
20ab4425 | 138 | |
6016ee13 NK |
139 | pcpu_list = tp_event->perf_events; |
140 | if (WARN_ON_ONCE(!pcpu_list)) | |
1c024eca | 141 | return -EINVAL; |
20ab4425 | 142 | |
a4eaf7f1 PZ |
143 | if (!(flags & PERF_EF_START)) |
144 | p_event->hw.state = PERF_HES_STOPPED; | |
145 | ||
6016ee13 | 146 | list = this_cpu_ptr(pcpu_list); |
1c024eca | 147 | hlist_add_head_rcu(&p_event->hlist_entry, list); |
20ab4425 | 148 | |
1c024eca PZ |
149 | return 0; |
150 | } | |
20ab4425 | 151 | |
a4eaf7f1 | 152 | void perf_trace_del(struct perf_event *p_event, int flags) |
1c024eca PZ |
153 | { |
154 | hlist_del_rcu(&p_event->hlist_entry); | |
e5e25cf4 FW |
155 | } |
156 | ||
1c024eca | 157 | void perf_trace_destroy(struct perf_event *p_event) |
ac199db0 | 158 | { |
1c024eca PZ |
159 | struct ftrace_event_call *tp_event = p_event->tp_event; |
160 | int i; | |
ac199db0 | 161 | |
2e97942f | 162 | mutex_lock(&event_mutex); |
1c024eca | 163 | if (--tp_event->perf_refcount > 0) |
2e97942f | 164 | goto out; |
1c024eca | 165 | |
a1d0ce82 | 166 | tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER); |
1c024eca | 167 | |
3771f077 | 168 | /* |
669336e4 FW |
169 | * Ensure our callback won't be called anymore. The buffers |
170 | * will be freed after that. | |
3771f077 | 171 | */ |
669336e4 | 172 | tracepoint_synchronize_unregister(); |
3771f077 | 173 | |
1c024eca PZ |
174 | free_percpu(tp_event->perf_events); |
175 | tp_event->perf_events = NULL; | |
176 | ||
177 | if (!--total_ref_count) { | |
7ae07ea3 | 178 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
1c024eca PZ |
179 | free_percpu(perf_trace_buf[i]); |
180 | perf_trace_buf[i] = NULL; | |
20c8928a | 181 | } |
ac199db0 | 182 | } |
2e97942f | 183 | out: |
9cb627d5 | 184 | module_put(tp_event->mod); |
2e97942f | 185 | mutex_unlock(&event_mutex); |
ac199db0 | 186 | } |
430ad5a6 | 187 | |
97d5a220 | 188 | __kprobes void *perf_trace_buf_prepare(int size, unsigned short type, |
b7e2ecef | 189 | struct pt_regs *regs, int *rctxp) |
430ad5a6 XG |
190 | { |
191 | struct trace_entry *entry; | |
87f44bbc | 192 | unsigned long flags; |
1c024eca | 193 | char *raw_data; |
b7e2ecef | 194 | int pc; |
430ad5a6 | 195 | |
eb1e7961 FW |
196 | BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long)); |
197 | ||
430ad5a6 XG |
198 | pc = preempt_count(); |
199 | ||
430ad5a6 XG |
200 | *rctxp = perf_swevent_get_recursion_context(); |
201 | if (*rctxp < 0) | |
1c024eca | 202 | return NULL; |
430ad5a6 | 203 | |
3771f077 | 204 | raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]); |
430ad5a6 XG |
205 | |
206 | /* zero the dead bytes from align to not leak stack to user */ | |
eb1e7961 | 207 | memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64)); |
430ad5a6 XG |
208 | |
209 | entry = (struct trace_entry *)raw_data; | |
87f44bbc PZ |
210 | local_save_flags(flags); |
211 | tracing_generic_entry_update(entry, flags, pc); | |
430ad5a6 XG |
212 | entry->type = type; |
213 | ||
214 | return raw_data; | |
430ad5a6 | 215 | } |
97d5a220 | 216 | EXPORT_SYMBOL_GPL(perf_trace_buf_prepare); |