Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 1 | /* |
Frederic Weisbecker | 97d5a22 | 2010-03-05 05:35:37 +0100 | [diff] [blame] | 2 | * trace event based perf event profiling/tracing |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 3 | * |
| 4 | * Copyright (C) 2009 Red Hat Inc, Peter Zijlstra <pzijlstr@redhat.com> |
Frederic Weisbecker | c530665 | 2010-03-03 07:16:16 +0100 | [diff] [blame] | 5 | * Copyright (C) 2009-2010 Frederic Weisbecker <fweisbec@gmail.com> |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 6 | */ |
| 7 | |
Li Zefan | 558e654 | 2009-08-24 12:19:47 +0800 | [diff] [blame] | 8 | #include <linux/module.h> |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 9 | #include <linux/kprobes.h> |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 10 | #include "trace.h" |
| 11 | |
Namhyung Kim | 6016ee1 | 2010-08-11 12:47:59 +0900 | [diff] [blame] | 12 | static char __percpu *perf_trace_buf[PERF_NR_CONTEXTS]; |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 13 | |
Frederic Weisbecker | eb1e796 | 2010-03-23 00:08:59 +0100 | [diff] [blame] | 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; |
Frederic Weisbecker | ce71b9d | 2009-11-22 05:26:55 +0100 | [diff] [blame] | 20 | |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 21 | /* Count the events in use (per event id, not per instance) */ |
Frederic Weisbecker | 97d5a22 | 2010-03-05 05:35:37 +0100 | [diff] [blame] | 22 | static int total_ref_count; |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 23 | |
Frederic Weisbecker | 61c3265 | 2010-11-18 01:39:17 +0100 | [diff] [blame] | 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 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 47 | static int perf_trace_event_init(struct ftrace_event_call *tp_event, |
| 48 | struct perf_event *p_event) |
Frederic Weisbecker | e5e25cf | 2009-09-18 00:54:43 +0200 | [diff] [blame] | 49 | { |
Namhyung Kim | 6016ee1 | 2010-08-11 12:47:59 +0900 | [diff] [blame] | 50 | struct hlist_head __percpu *list; |
Frederic Weisbecker | 61c3265 | 2010-11-18 01:39:17 +0100 | [diff] [blame] | 51 | int ret; |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 52 | int cpu; |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 53 | |
Frederic Weisbecker | 61c3265 | 2010-11-18 01:39:17 +0100 | [diff] [blame] | 54 | ret = perf_trace_event_perm(tp_event, p_event); |
| 55 | if (ret) |
| 56 | return ret; |
| 57 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 58 | p_event->tp_event = tp_event; |
| 59 | if (tp_event->perf_refcount++ > 0) |
Frederic Weisbecker | e5e25cf | 2009-09-18 00:54:43 +0200 | [diff] [blame] | 60 | return 0; |
| 61 | |
Frederic Weisbecker | 61c3265 | 2010-11-18 01:39:17 +0100 | [diff] [blame] | 62 | ret = -ENOMEM; |
| 63 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 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)); |
| 70 | |
| 71 | tp_event->perf_events = list; |
Frederic Weisbecker | e5e25cf | 2009-09-18 00:54:43 +0200 | [diff] [blame] | 72 | |
Frederic Weisbecker | 97d5a22 | 2010-03-05 05:35:37 +0100 | [diff] [blame] | 73 | if (!total_ref_count) { |
Namhyung Kim | 6016ee1 | 2010-08-11 12:47:59 +0900 | [diff] [blame] | 74 | char __percpu *buf; |
Peter Zijlstra | b7e2ece | 2010-05-19 10:52:27 +0200 | [diff] [blame] | 75 | int i; |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 76 | |
Frederic Weisbecker | 7ae07ea | 2010-08-14 20:45:13 +0200 | [diff] [blame] | 77 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
Namhyung Kim | 6016ee1 | 2010-08-11 12:47:59 +0900 | [diff] [blame] | 78 | buf = (char __percpu *)alloc_percpu(perf_trace_t); |
Peter Zijlstra | b7e2ece | 2010-05-19 10:52:27 +0200 | [diff] [blame] | 79 | if (!buf) |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 80 | goto fail; |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 81 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 82 | perf_trace_buf[i] = buf; |
Peter Zijlstra | b7e2ece | 2010-05-19 10:52:27 +0200 | [diff] [blame] | 83 | } |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 84 | } |
| 85 | |
Steven Rostedt | a1d0ce8 | 2010-06-08 11:22:06 -0400 | [diff] [blame] | 86 | ret = tp_event->class->reg(tp_event, TRACE_REG_PERF_REGISTER); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 87 | if (ret) |
| 88 | goto fail; |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 89 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 90 | total_ref_count++; |
| 91 | return 0; |
| 92 | |
| 93 | fail: |
Peter Zijlstra | b7e2ece | 2010-05-19 10:52:27 +0200 | [diff] [blame] | 94 | if (!total_ref_count) { |
| 95 | int i; |
| 96 | |
Frederic Weisbecker | 7ae07ea | 2010-08-14 20:45:13 +0200 | [diff] [blame] | 97 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
Peter Zijlstra | b7e2ece | 2010-05-19 10:52:27 +0200 | [diff] [blame] | 98 | free_percpu(perf_trace_buf[i]); |
| 99 | perf_trace_buf[i] = NULL; |
| 100 | } |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 101 | } |
| 102 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 103 | if (!--tp_event->perf_refcount) { |
| 104 | free_percpu(tp_event->perf_events); |
| 105 | tp_event->perf_events = NULL; |
Frederic Weisbecker | fe8e5b5 | 2009-10-03 14:55:18 +0200 | [diff] [blame] | 106 | } |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 107 | |
| 108 | return ret; |
Frederic Weisbecker | e5e25cf | 2009-09-18 00:54:43 +0200 | [diff] [blame] | 109 | } |
| 110 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 111 | int perf_trace_init(struct perf_event *p_event) |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 112 | { |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 113 | struct ftrace_event_call *tp_event; |
| 114 | int event_id = p_event->attr.config; |
Li Zefan | 20c8928 | 2009-05-06 10:33:45 +0800 | [diff] [blame] | 115 | int ret = -EINVAL; |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 116 | |
Li Zefan | 20c8928 | 2009-05-06 10:33:45 +0800 | [diff] [blame] | 117 | mutex_lock(&event_mutex); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 118 | list_for_each_entry(tp_event, &ftrace_events, list) { |
Steven Rostedt | ff5f149 | 2010-05-21 11:49:57 -0400 | [diff] [blame] | 119 | if (tp_event->event.type == event_id && |
Steven Rostedt | a1d0ce8 | 2010-06-08 11:22:06 -0400 | [diff] [blame] | 120 | tp_event->class && tp_event->class->reg && |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 121 | try_module_get(tp_event->mod)) { |
| 122 | ret = perf_trace_event_init(tp_event, p_event); |
Li Zefan | 9cb627d | 2010-09-01 12:58:43 +0200 | [diff] [blame] | 123 | if (ret) |
| 124 | module_put(tp_event->mod); |
Li Zefan | 20c8928 | 2009-05-06 10:33:45 +0800 | [diff] [blame] | 125 | break; |
| 126 | } |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 127 | } |
Li Zefan | 20c8928 | 2009-05-06 10:33:45 +0800 | [diff] [blame] | 128 | mutex_unlock(&event_mutex); |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 129 | |
Li Zefan | 20c8928 | 2009-05-06 10:33:45 +0800 | [diff] [blame] | 130 | return ret; |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 131 | } |
| 132 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 133 | int perf_trace_add(struct perf_event *p_event, int flags) |
Frederic Weisbecker | e5e25cf | 2009-09-18 00:54:43 +0200 | [diff] [blame] | 134 | { |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 135 | struct ftrace_event_call *tp_event = p_event->tp_event; |
Namhyung Kim | 6016ee1 | 2010-08-11 12:47:59 +0900 | [diff] [blame] | 136 | struct hlist_head __percpu *pcpu_list; |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 137 | struct hlist_head *list; |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 138 | |
Namhyung Kim | 6016ee1 | 2010-08-11 12:47:59 +0900 | [diff] [blame] | 139 | pcpu_list = tp_event->perf_events; |
| 140 | if (WARN_ON_ONCE(!pcpu_list)) |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 141 | return -EINVAL; |
Frederic Weisbecker | e5e25cf | 2009-09-18 00:54:43 +0200 | [diff] [blame] | 142 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 143 | if (!(flags & PERF_EF_START)) |
| 144 | p_event->hw.state = PERF_HES_STOPPED; |
| 145 | |
Namhyung Kim | 6016ee1 | 2010-08-11 12:47:59 +0900 | [diff] [blame] | 146 | list = this_cpu_ptr(pcpu_list); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 147 | hlist_add_head_rcu(&p_event->hlist_entry, list); |
Frederic Weisbecker | 20ab4425 | 2009-09-18 06:10:28 +0200 | [diff] [blame] | 148 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 149 | return 0; |
Frederic Weisbecker | e5e25cf | 2009-09-18 00:54:43 +0200 | [diff] [blame] | 150 | } |
| 151 | |
Peter Zijlstra | a4eaf7f | 2010-06-16 14:37:10 +0200 | [diff] [blame] | 152 | void perf_trace_del(struct perf_event *p_event, int flags) |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 153 | { |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 154 | hlist_del_rcu(&p_event->hlist_entry); |
| 155 | } |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 156 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 157 | void perf_trace_destroy(struct perf_event *p_event) |
| 158 | { |
| 159 | struct ftrace_event_call *tp_event = p_event->tp_event; |
| 160 | int i; |
| 161 | |
Peter Zijlstra | 2e97942 | 2010-05-21 16:22:33 +0200 | [diff] [blame] | 162 | mutex_lock(&event_mutex); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 163 | if (--tp_event->perf_refcount > 0) |
Peter Zijlstra | 2e97942 | 2010-05-21 16:22:33 +0200 | [diff] [blame] | 164 | goto out; |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 165 | |
Steven Rostedt | a1d0ce8 | 2010-06-08 11:22:06 -0400 | [diff] [blame] | 166 | tp_event->class->reg(tp_event, TRACE_REG_PERF_UNREGISTER); |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 167 | |
Peter Zijlstra | 3771f07 | 2010-05-21 12:31:09 +0200 | [diff] [blame] | 168 | /* |
Frederic Weisbecker | 669336e | 2010-07-20 17:29:54 +0200 | [diff] [blame] | 169 | * Ensure our callback won't be called anymore. The buffers |
| 170 | * will be freed after that. |
Peter Zijlstra | 3771f07 | 2010-05-21 12:31:09 +0200 | [diff] [blame] | 171 | */ |
Frederic Weisbecker | 669336e | 2010-07-20 17:29:54 +0200 | [diff] [blame] | 172 | tracepoint_synchronize_unregister(); |
Peter Zijlstra | 3771f07 | 2010-05-21 12:31:09 +0200 | [diff] [blame] | 173 | |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 174 | free_percpu(tp_event->perf_events); |
| 175 | tp_event->perf_events = NULL; |
| 176 | |
| 177 | if (!--total_ref_count) { |
Frederic Weisbecker | 7ae07ea | 2010-08-14 20:45:13 +0200 | [diff] [blame] | 178 | for (i = 0; i < PERF_NR_CONTEXTS; i++) { |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 179 | free_percpu(perf_trace_buf[i]); |
| 180 | perf_trace_buf[i] = NULL; |
Li Zefan | 20c8928 | 2009-05-06 10:33:45 +0800 | [diff] [blame] | 181 | } |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 182 | } |
Peter Zijlstra | 2e97942 | 2010-05-21 16:22:33 +0200 | [diff] [blame] | 183 | out: |
Li Zefan | 9cb627d | 2010-09-01 12:58:43 +0200 | [diff] [blame] | 184 | module_put(tp_event->mod); |
Peter Zijlstra | 2e97942 | 2010-05-21 16:22:33 +0200 | [diff] [blame] | 185 | mutex_unlock(&event_mutex); |
Peter Zijlstra | ac199db | 2009-03-19 20:26:15 +0100 | [diff] [blame] | 186 | } |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 187 | |
Frederic Weisbecker | 97d5a22 | 2010-03-05 05:35:37 +0100 | [diff] [blame] | 188 | __kprobes void *perf_trace_buf_prepare(int size, unsigned short type, |
Peter Zijlstra | b7e2ece | 2010-05-19 10:52:27 +0200 | [diff] [blame] | 189 | struct pt_regs *regs, int *rctxp) |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 190 | { |
| 191 | struct trace_entry *entry; |
Peter Zijlstra | 87f44bb | 2010-05-25 11:02:55 +0200 | [diff] [blame] | 192 | unsigned long flags; |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 193 | char *raw_data; |
Peter Zijlstra | b7e2ece | 2010-05-19 10:52:27 +0200 | [diff] [blame] | 194 | int pc; |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 195 | |
Frederic Weisbecker | eb1e796 | 2010-03-23 00:08:59 +0100 | [diff] [blame] | 196 | BUILD_BUG_ON(PERF_MAX_TRACE_SIZE % sizeof(unsigned long)); |
| 197 | |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 198 | pc = preempt_count(); |
| 199 | |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 200 | *rctxp = perf_swevent_get_recursion_context(); |
| 201 | if (*rctxp < 0) |
Peter Zijlstra | 1c024eca | 2010-05-19 14:02:22 +0200 | [diff] [blame] | 202 | return NULL; |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 203 | |
Peter Zijlstra | 3771f07 | 2010-05-21 12:31:09 +0200 | [diff] [blame] | 204 | raw_data = this_cpu_ptr(perf_trace_buf[*rctxp]); |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 205 | |
| 206 | /* zero the dead bytes from align to not leak stack to user */ |
Frederic Weisbecker | eb1e796 | 2010-03-23 00:08:59 +0100 | [diff] [blame] | 207 | memset(&raw_data[size - sizeof(u64)], 0, sizeof(u64)); |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 208 | |
| 209 | entry = (struct trace_entry *)raw_data; |
Peter Zijlstra | 87f44bb | 2010-05-25 11:02:55 +0200 | [diff] [blame] | 210 | local_save_flags(flags); |
| 211 | tracing_generic_entry_update(entry, flags, pc); |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 212 | entry->type = type; |
| 213 | |
| 214 | return raw_data; |
Xiao Guangrong | 430ad5a | 2010-01-28 09:32:29 +0800 | [diff] [blame] | 215 | } |
Frederic Weisbecker | 97d5a22 | 2010-03-05 05:35:37 +0100 | [diff] [blame] | 216 | EXPORT_SYMBOL_GPL(perf_trace_buf_prepare); |