2 * Performance events callchain code, extracted from core.c:
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>
9 * For licensing details see kernel-base/COPYING
12 #include <linux/perf_event.h>
13 #include <linux/slab.h>
16 struct callchain_cpus_entries {
17 struct rcu_head rcu_head;
18 struct perf_callchain_entry *cpu_entries[0];
21 int sysctl_perf_event_max_stack __read_mostly = PERF_MAX_STACK_DEPTH;
23 static inline size_t perf_callchain_entry__sizeof(void)
25 return (sizeof(struct perf_callchain_entry) +
26 sizeof(__u64) * sysctl_perf_event_max_stack);
29 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
30 static atomic_t nr_callchain_events;
31 static DEFINE_MUTEX(callchain_mutex);
32 static struct callchain_cpus_entries *callchain_cpus_entries;
35 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry,
40 __weak void perf_callchain_user(struct perf_callchain_entry *entry,
45 static void release_callchain_buffers_rcu(struct rcu_head *head)
47 struct callchain_cpus_entries *entries;
50 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
52 for_each_possible_cpu(cpu)
53 kfree(entries->cpu_entries[cpu]);
58 static void release_callchain_buffers(void)
60 struct callchain_cpus_entries *entries;
62 entries = callchain_cpus_entries;
63 RCU_INIT_POINTER(callchain_cpus_entries, NULL);
64 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
67 static int alloc_callchain_buffers(void)
71 struct callchain_cpus_entries *entries;
74 * We can't use the percpu allocation API for data that can be
75 * accessed from NMI. Use a temporary manual per cpu allocation
76 * until that gets sorted out.
78 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
80 entries = kzalloc(size, GFP_KERNEL);
84 size = perf_callchain_entry__sizeof() * PERF_NR_CONTEXTS;
86 for_each_possible_cpu(cpu) {
87 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
89 if (!entries->cpu_entries[cpu])
93 rcu_assign_pointer(callchain_cpus_entries, entries);
98 for_each_possible_cpu(cpu)
99 kfree(entries->cpu_entries[cpu]);
105 int get_callchain_buffers(void)
110 mutex_lock(&callchain_mutex);
112 count = atomic_inc_return(&nr_callchain_events);
113 if (WARN_ON_ONCE(count < 1)) {
119 /* If the allocation failed, give up */
120 if (!callchain_cpus_entries)
125 err = alloc_callchain_buffers();
128 atomic_dec(&nr_callchain_events);
130 mutex_unlock(&callchain_mutex);
135 void put_callchain_buffers(void)
137 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
138 release_callchain_buffers();
139 mutex_unlock(&callchain_mutex);
143 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
146 struct callchain_cpus_entries *entries;
148 *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
152 entries = rcu_dereference(callchain_cpus_entries);
156 cpu = smp_processor_id();
158 return (((void *)entries->cpu_entries[cpu]) +
159 (*rctx * perf_callchain_entry__sizeof()));
163 put_callchain_entry(int rctx)
165 put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
168 struct perf_callchain_entry *
169 perf_callchain(struct perf_event *event, struct pt_regs *regs)
171 bool kernel = !event->attr.exclude_callchain_kernel;
172 bool user = !event->attr.exclude_callchain_user;
173 /* Disallow cross-task user callchains. */
174 bool crosstask = event->ctx->task && event->ctx->task != current;
176 if (!kernel && !user)
179 return get_perf_callchain(regs, 0, kernel, user, crosstask, true);
182 struct perf_callchain_entry *
183 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
184 bool crosstask, bool add_mark)
186 struct perf_callchain_entry *entry;
189 entry = get_callchain_entry(&rctx);
198 if (kernel && !user_mode(regs)) {
200 perf_callchain_store(entry, PERF_CONTEXT_KERNEL);
201 perf_callchain_kernel(entry, regs);
205 if (!user_mode(regs)) {
207 regs = task_pt_regs(current);
217 perf_callchain_store(entry, PERF_CONTEXT_USER);
218 perf_callchain_user(entry, regs);
223 put_callchain_entry(rctx);
228 int perf_event_max_stack_handler(struct ctl_table *table, int write,
229 void __user *buffer, size_t *lenp, loff_t *ppos)
231 int new_value = sysctl_perf_event_max_stack, ret;
232 struct ctl_table new_table = *table;
234 new_table.data = &new_value;
235 ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
239 mutex_lock(&callchain_mutex);
240 if (atomic_read(&nr_callchain_events))
243 sysctl_perf_event_max_stack = new_value;
245 mutex_unlock(&callchain_mutex);