1 /*
2 * Performance events callchain code, extracted from core.c:
3 *
4 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
5 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
6 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
7 * Copyright � 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
8 *
9 * For licensing details see kernel-base/COPYING
10 */
11
12 #include <linux/perf_event.h>
13 #include <linux/slab.h>
14 #include "internal.h"
15
16 struct callchain_cpus_entries {
17 struct rcu_head rcu_head;
18 struct perf_callchain_entry *cpu_entries[0];
19 };
20
21 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
22 static atomic_t nr_callchain_events;
23 static DEFINE_MUTEX(callchain_mutex);
24 static struct callchain_cpus_entries *callchain_cpus_entries;
25
26
perf_callchain_kernel(struct perf_callchain_entry * entry,struct pt_regs * regs)27 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
28 struct pt_regs *regs)
29 {
30 }
31
perf_callchain_user(struct perf_callchain_entry * entry,struct pt_regs * regs)32 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
33 struct pt_regs *regs)
34 {
35 }
36
release_callchain_buffers_rcu(struct rcu_head * head)37 static void release_callchain_buffers_rcu(struct rcu_head *head)
38 {
39 struct callchain_cpus_entries *entries;
40 int cpu;
41
42 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
43
44 for_each_possible_cpu(cpu)
45 kfree(entries->cpu_entries[cpu]);
46
47 kfree(entries);
48 }
49
release_callchain_buffers(void)50 static void release_callchain_buffers(void)
51 {
52 struct callchain_cpus_entries *entries;
53
54 entries = callchain_cpus_entries;
55 RCU_INIT_POINTER(callchain_cpus_entries, NULL);
56 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
57 }
58
alloc_callchain_buffers(void)59 static int alloc_callchain_buffers(void)
60 {
61 int cpu;
62 int size;
63 struct callchain_cpus_entries *entries;
64
65 /*
66 * We can't use the percpu allocation API for data that can be
67 * accessed from NMI. Use a temporary manual per cpu allocation
68 * until that gets sorted out.
69 */
70 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
71
72 entries = kzalloc(size, GFP_KERNEL);
73 if (!entries)
74 return -ENOMEM;
75
76 size = sizeof(struct perf_callchain_entry) * PERF_NR_CONTEXTS;
77
78 for_each_possible_cpu(cpu) {
79 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
80 cpu_to_node(cpu));
81 if (!entries->cpu_entries[cpu])
82 goto fail;
83 }
84
85 rcu_assign_pointer(callchain_cpus_entries, entries);
86
87 return 0;
88
89 fail:
90 for_each_possible_cpu(cpu)
91 kfree(entries->cpu_entries[cpu]);
92 kfree(entries);
93
94 return -ENOMEM;
95 }
96
get_callchain_buffers(void)97 int get_callchain_buffers(void)
98 {
99 int err = 0;
100 int count;
101
102 mutex_lock(&callchain_mutex);
103
104 count = atomic_inc_return(&nr_callchain_events);
105 if (WARN_ON_ONCE(count < 1)) {
106 err = -EINVAL;
107 goto exit;
108 }
109
110 if (count == 1)
111 err = alloc_callchain_buffers();
112 exit:
113 if (err)
114 atomic_dec(&nr_callchain_events);
115
116 mutex_unlock(&callchain_mutex);
117
118 return err;
119 }
120
put_callchain_buffers(void)121 void put_callchain_buffers(void)
122 {
123 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
124 release_callchain_buffers();
125 mutex_unlock(&callchain_mutex);
126 }
127 }
128
get_callchain_entry(int * rctx)129 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
130 {
131 int cpu;
132 struct callchain_cpus_entries *entries;
133
134 *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
135 if (*rctx == -1)
136 return NULL;
137
138 entries = rcu_dereference(callchain_cpus_entries);
139 if (!entries)
140 return NULL;
141
142 cpu = smp_processor_id();
143
144 return &entries->cpu_entries[cpu][*rctx];
145 }
146
147 static void
put_callchain_entry(int rctx)148 put_callchain_entry(int rctx)
149 {
150 put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
151 }
152
153 struct perf_callchain_entry *
perf_callchain(struct perf_event * event,struct pt_regs * regs)154 perf_callchain(struct perf_event *event, struct pt_regs *regs)
155 {
156 int rctx;
157 struct perf_callchain_entry *entry;
158
159 int kernel = !event->attr.exclude_callchain_kernel;
160 int user = !event->attr.exclude_callchain_user;
161
162 if (!kernel && !user)
163 return NULL;
164
165 entry = get_callchain_entry(&rctx);
166 if (rctx == -1)
167 return NULL;
168
169 if (!entry)
170 goto exit_put;
171
172 entry->nr = 0;
173
174 if (kernel && !user_mode(regs)) {
175 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
176 perf_callchain_kernel(entry, regs);
177 }
178
179 if (user) {
180 if (!user_mode(regs)) {
181 if (current->mm)
182 regs = task_pt_regs(current);
183 else
184 regs = NULL;
185 }
186
187 if (regs) {
188 /*
189 * Disallow cross-task user callchains.
190 */
191 if (event->ctx->task && event->ctx->task != current)
192 goto exit_put;
193
194 perf_callchain_store(entry, PERF_CONTEXT_USER);
195 perf_callchain_user(entry, regs);
196 }
197 }
198
199 exit_put:
200 put_callchain_entry(rctx);
201
202 return entry;
203 }
204