Loading...
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
16struct callchain_cpus_entries {
17 struct rcu_head rcu_head;
18 struct perf_callchain_entry *cpu_entries[0];
19};
20
21static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
22static atomic_t nr_callchain_events;
23static DEFINE_MUTEX(callchain_mutex);
24static struct callchain_cpus_entries *callchain_cpus_entries;
25
26
27__weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
28 struct pt_regs *regs)
29{
30}
31
32__weak void perf_callchain_user(struct perf_callchain_entry *entry,
33 struct pt_regs *regs)
34{
35}
36
37static 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
50static 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
59static 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
89fail:
90 for_each_possible_cpu(cpu)
91 kfree(entries->cpu_entries[cpu]);
92 kfree(entries);
93
94 return -ENOMEM;
95}
96
97int 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 /* If the allocation failed, give up */
112 if (!callchain_cpus_entries)
113 err = -ENOMEM;
114 goto exit;
115 }
116
117 err = alloc_callchain_buffers();
118exit:
119 if (err)
120 atomic_dec(&nr_callchain_events);
121
122 mutex_unlock(&callchain_mutex);
123
124 return err;
125}
126
127void put_callchain_buffers(void)
128{
129 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
130 release_callchain_buffers();
131 mutex_unlock(&callchain_mutex);
132 }
133}
134
135static struct perf_callchain_entry *get_callchain_entry(int *rctx)
136{
137 int cpu;
138 struct callchain_cpus_entries *entries;
139
140 *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
141 if (*rctx == -1)
142 return NULL;
143
144 entries = rcu_dereference(callchain_cpus_entries);
145 if (!entries)
146 return NULL;
147
148 cpu = smp_processor_id();
149
150 return &entries->cpu_entries[cpu][*rctx];
151}
152
153static void
154put_callchain_entry(int rctx)
155{
156 put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
157}
158
159struct perf_callchain_entry *
160perf_callchain(struct perf_event *event, struct pt_regs *regs)
161{
162 bool kernel = !event->attr.exclude_callchain_kernel;
163 bool user = !event->attr.exclude_callchain_user;
164 /* Disallow cross-task user callchains. */
165 bool crosstask = event->ctx->task && event->ctx->task != current;
166
167 if (!kernel && !user)
168 return NULL;
169
170 return get_perf_callchain(regs, 0, kernel, user, crosstask, true);
171}
172
173struct perf_callchain_entry *
174get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
175 bool crosstask, bool add_mark)
176{
177 struct perf_callchain_entry *entry;
178 int rctx;
179
180 entry = get_callchain_entry(&rctx);
181 if (rctx == -1)
182 return NULL;
183
184 if (!entry)
185 goto exit_put;
186
187 entry->nr = init_nr;
188
189 if (kernel && !user_mode(regs)) {
190 if (add_mark)
191 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
192 perf_callchain_kernel(entry, regs);
193 }
194
195 if (user) {
196 if (!user_mode(regs)) {
197 if (current->mm)
198 regs = task_pt_regs(current);
199 else
200 regs = NULL;
201 }
202
203 if (regs) {
204 if (crosstask)
205 goto exit_put;
206
207 if (add_mark)
208 perf_callchain_store(entry, PERF_CONTEXT_USER);
209 perf_callchain_user(entry, regs);
210 }
211 }
212
213exit_put:
214 put_callchain_entry(rctx);
215
216 return entry;
217}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Performance events callchain code, extracted from core.c:
4 *
5 * Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
6 * Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
7 * Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
8 * Copyright © 2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 */
10
11#include <linux/perf_event.h>
12#include <linux/slab.h>
13#include <linux/sched/task_stack.h>
14
15#include "internal.h"
16
17struct callchain_cpus_entries {
18 struct rcu_head rcu_head;
19 struct perf_callchain_entry *cpu_entries[];
20};
21
22int sysctl_perf_event_max_stack __read_mostly = PERF_MAX_STACK_DEPTH;
23int sysctl_perf_event_max_contexts_per_stack __read_mostly = PERF_MAX_CONTEXTS_PER_STACK;
24
25static inline size_t perf_callchain_entry__sizeof(void)
26{
27 return (sizeof(struct perf_callchain_entry) +
28 sizeof(__u64) * (sysctl_perf_event_max_stack +
29 sysctl_perf_event_max_contexts_per_stack));
30}
31
32static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
33static atomic_t nr_callchain_events;
34static DEFINE_MUTEX(callchain_mutex);
35static struct callchain_cpus_entries *callchain_cpus_entries;
36
37
38__weak void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
39 struct pt_regs *regs)
40{
41}
42
43__weak void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
44 struct pt_regs *regs)
45{
46}
47
48static void release_callchain_buffers_rcu(struct rcu_head *head)
49{
50 struct callchain_cpus_entries *entries;
51 int cpu;
52
53 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
54
55 for_each_possible_cpu(cpu)
56 kfree(entries->cpu_entries[cpu]);
57
58 kfree(entries);
59}
60
61static void release_callchain_buffers(void)
62{
63 struct callchain_cpus_entries *entries;
64
65 entries = callchain_cpus_entries;
66 RCU_INIT_POINTER(callchain_cpus_entries, NULL);
67 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
68}
69
70static int alloc_callchain_buffers(void)
71{
72 int cpu;
73 int size;
74 struct callchain_cpus_entries *entries;
75
76 /*
77 * We can't use the percpu allocation API for data that can be
78 * accessed from NMI. Use a temporary manual per cpu allocation
79 * until that gets sorted out.
80 */
81 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
82
83 entries = kzalloc(size, GFP_KERNEL);
84 if (!entries)
85 return -ENOMEM;
86
87 size = perf_callchain_entry__sizeof() * PERF_NR_CONTEXTS;
88
89 for_each_possible_cpu(cpu) {
90 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
91 cpu_to_node(cpu));
92 if (!entries->cpu_entries[cpu])
93 goto fail;
94 }
95
96 rcu_assign_pointer(callchain_cpus_entries, entries);
97
98 return 0;
99
100fail:
101 for_each_possible_cpu(cpu)
102 kfree(entries->cpu_entries[cpu]);
103 kfree(entries);
104
105 return -ENOMEM;
106}
107
108int get_callchain_buffers(int event_max_stack)
109{
110 int err = 0;
111 int count;
112
113 mutex_lock(&callchain_mutex);
114
115 count = atomic_inc_return(&nr_callchain_events);
116 if (WARN_ON_ONCE(count < 1)) {
117 err = -EINVAL;
118 goto exit;
119 }
120
121 /*
122 * If requesting per event more than the global cap,
123 * return a different error to help userspace figure
124 * this out.
125 *
126 * And also do it here so that we have &callchain_mutex held.
127 */
128 if (event_max_stack > sysctl_perf_event_max_stack) {
129 err = -EOVERFLOW;
130 goto exit;
131 }
132
133 if (count == 1)
134 err = alloc_callchain_buffers();
135exit:
136 if (err)
137 atomic_dec(&nr_callchain_events);
138
139 mutex_unlock(&callchain_mutex);
140
141 return err;
142}
143
144void put_callchain_buffers(void)
145{
146 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
147 release_callchain_buffers();
148 mutex_unlock(&callchain_mutex);
149 }
150}
151
152struct perf_callchain_entry *get_callchain_entry(int *rctx)
153{
154 int cpu;
155 struct callchain_cpus_entries *entries;
156
157 *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
158 if (*rctx == -1)
159 return NULL;
160
161 entries = rcu_dereference(callchain_cpus_entries);
162 if (!entries) {
163 put_recursion_context(this_cpu_ptr(callchain_recursion), *rctx);
164 return NULL;
165 }
166
167 cpu = smp_processor_id();
168
169 return (((void *)entries->cpu_entries[cpu]) +
170 (*rctx * perf_callchain_entry__sizeof()));
171}
172
173void
174put_callchain_entry(int rctx)
175{
176 put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
177}
178
179struct perf_callchain_entry *
180get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
181 u32 max_stack, bool crosstask, bool add_mark)
182{
183 struct perf_callchain_entry *entry;
184 struct perf_callchain_entry_ctx ctx;
185 int rctx;
186
187 entry = get_callchain_entry(&rctx);
188 if (!entry)
189 return NULL;
190
191 ctx.entry = entry;
192 ctx.max_stack = max_stack;
193 ctx.nr = entry->nr = init_nr;
194 ctx.contexts = 0;
195 ctx.contexts_maxed = false;
196
197 if (kernel && !user_mode(regs)) {
198 if (add_mark)
199 perf_callchain_store_context(&ctx, PERF_CONTEXT_KERNEL);
200 perf_callchain_kernel(&ctx, regs);
201 }
202
203 if (user) {
204 if (!user_mode(regs)) {
205 if (current->mm)
206 regs = task_pt_regs(current);
207 else
208 regs = NULL;
209 }
210
211 if (regs) {
212 if (crosstask)
213 goto exit_put;
214
215 if (add_mark)
216 perf_callchain_store_context(&ctx, PERF_CONTEXT_USER);
217
218 perf_callchain_user(&ctx, regs);
219 }
220 }
221
222exit_put:
223 put_callchain_entry(rctx);
224
225 return entry;
226}
227
228/*
229 * Used for sysctl_perf_event_max_stack and
230 * sysctl_perf_event_max_contexts_per_stack.
231 */
232int perf_event_max_stack_handler(struct ctl_table *table, int write,
233 void *buffer, size_t *lenp, loff_t *ppos)
234{
235 int *value = table->data;
236 int new_value = *value, ret;
237 struct ctl_table new_table = *table;
238
239 new_table.data = &new_value;
240 ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
241 if (ret || !write)
242 return ret;
243
244 mutex_lock(&callchain_mutex);
245 if (atomic_read(&nr_callchain_events))
246 ret = -EBUSY;
247 else
248 *value = new_value;
249
250 mutex_unlock(&callchain_mutex);
251
252 return ret;
253}