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;
22 int sysctl_perf_event_max_contexts_per_stack __read_mostly = PERF_MAX_CONTEXTS_PER_STACK;
24 static inline size_t perf_callchain_entry__sizeof(void)
26 return (sizeof(struct perf_callchain_entry) +
27 sizeof(__u64) * (sysctl_perf_event_max_stack +
28 sysctl_perf_event_max_contexts_per_stack));
31 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
32 static atomic_t nr_callchain_events;
33 static DEFINE_MUTEX(callchain_mutex);
34 static struct callchain_cpus_entries *callchain_cpus_entries;
37 __weak void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
42 __weak void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
47 static void release_callchain_buffers_rcu(struct rcu_head *head)
49 struct callchain_cpus_entries *entries;
52 entries = container_of(head, struct callchain_cpus_entries, rcu_head);
54 for_each_possible_cpu(cpu)
55 kfree(entries->cpu_entries[cpu]);
60 static void release_callchain_buffers(void)
62 struct callchain_cpus_entries *entries;
64 entries = callchain_cpus_entries;
65 RCU_INIT_POINTER(callchain_cpus_entries, NULL);
66 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
69 static int alloc_callchain_buffers(void)
73 struct callchain_cpus_entries *entries;
76 * We can't use the percpu allocation API for data that can be
77 * accessed from NMI. Use a temporary manual per cpu allocation
78 * until that gets sorted out.
80 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
82 entries = kzalloc(size, GFP_KERNEL);
86 size = perf_callchain_entry__sizeof() * PERF_NR_CONTEXTS;
88 for_each_possible_cpu(cpu) {
89 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
91 if (!entries->cpu_entries[cpu])
95 rcu_assign_pointer(callchain_cpus_entries, entries);
100 for_each_possible_cpu(cpu)
101 kfree(entries->cpu_entries[cpu]);
107 int get_callchain_buffers(void)
112 mutex_lock(&callchain_mutex);
114 count = atomic_inc_return(&nr_callchain_events);
115 if (WARN_ON_ONCE(count < 1)) {
121 /* If the allocation failed, give up */
122 if (!callchain_cpus_entries)
127 err = alloc_callchain_buffers();
130 atomic_dec(&nr_callchain_events);
132 mutex_unlock(&callchain_mutex);
137 void put_callchain_buffers(void)
139 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
140 release_callchain_buffers();
141 mutex_unlock(&callchain_mutex);
145 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
148 struct callchain_cpus_entries *entries;
150 *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
154 entries = rcu_dereference(callchain_cpus_entries);
158 cpu = smp_processor_id();
160 return (((void *)entries->cpu_entries[cpu]) +
161 (*rctx * perf_callchain_entry__sizeof()));
165 put_callchain_entry(int rctx)
167 put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
170 struct perf_callchain_entry *
171 perf_callchain(struct perf_event *event, struct pt_regs *regs)
173 bool kernel = !event->attr.exclude_callchain_kernel;
174 bool user = !event->attr.exclude_callchain_user;
175 /* Disallow cross-task user callchains. */
176 bool crosstask = event->ctx->task && event->ctx->task != current;
178 if (!kernel && !user)
181 return get_perf_callchain(regs, 0, kernel, user, sysctl_perf_event_max_stack, crosstask, true);
184 struct perf_callchain_entry *
185 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
186 u32 max_stack, bool crosstask, bool add_mark)
188 struct perf_callchain_entry *entry;
189 struct perf_callchain_entry_ctx ctx;
192 entry = get_callchain_entry(&rctx);
200 ctx.max_stack = max_stack;
201 ctx.nr = entry->nr = init_nr;
203 ctx.contexts_maxed = false;
205 if (kernel && !user_mode(regs)) {
207 perf_callchain_store_context(&ctx, PERF_CONTEXT_KERNEL);
208 perf_callchain_kernel(&ctx, regs);
212 if (!user_mode(regs)) {
214 regs = task_pt_regs(current);
224 perf_callchain_store_context(&ctx, PERF_CONTEXT_USER);
225 perf_callchain_user(&ctx, regs);
230 put_callchain_entry(rctx);
236 * Used for sysctl_perf_event_max_stack and
237 * sysctl_perf_event_max_contexts_per_stack.
239 int perf_event_max_stack_handler(struct ctl_table *table, int write,
240 void __user *buffer, size_t *lenp, loff_t *ppos)
242 int *value = table->data;
243 int new_value = *value, ret;
244 struct ctl_table new_table = *table;
246 new_table.data = &new_value;
247 ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
251 mutex_lock(&callchain_mutex);
252 if (atomic_read(&nr_callchain_events))
257 mutex_unlock(&callchain_mutex);