xref: /openbmc/linux/kernel/events/callchain.c (revision 80483c3a)
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 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;
23 
24 static inline size_t perf_callchain_entry__sizeof(void)
25 {
26 	return (sizeof(struct perf_callchain_entry) +
27 		sizeof(__u64) * (sysctl_perf_event_max_stack +
28 				 sysctl_perf_event_max_contexts_per_stack));
29 }
30 
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;
35 
36 
37 __weak void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
38 				  struct pt_regs *regs)
39 {
40 }
41 
42 __weak void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
43 				struct pt_regs *regs)
44 {
45 }
46 
47 static void release_callchain_buffers_rcu(struct rcu_head *head)
48 {
49 	struct callchain_cpus_entries *entries;
50 	int cpu;
51 
52 	entries = container_of(head, struct callchain_cpus_entries, rcu_head);
53 
54 	for_each_possible_cpu(cpu)
55 		kfree(entries->cpu_entries[cpu]);
56 
57 	kfree(entries);
58 }
59 
60 static void release_callchain_buffers(void)
61 {
62 	struct callchain_cpus_entries *entries;
63 
64 	entries = callchain_cpus_entries;
65 	RCU_INIT_POINTER(callchain_cpus_entries, NULL);
66 	call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
67 }
68 
69 static int alloc_callchain_buffers(void)
70 {
71 	int cpu;
72 	int size;
73 	struct callchain_cpus_entries *entries;
74 
75 	/*
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.
79 	 */
80 	size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
81 
82 	entries = kzalloc(size, GFP_KERNEL);
83 	if (!entries)
84 		return -ENOMEM;
85 
86 	size = perf_callchain_entry__sizeof() * PERF_NR_CONTEXTS;
87 
88 	for_each_possible_cpu(cpu) {
89 		entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
90 							 cpu_to_node(cpu));
91 		if (!entries->cpu_entries[cpu])
92 			goto fail;
93 	}
94 
95 	rcu_assign_pointer(callchain_cpus_entries, entries);
96 
97 	return 0;
98 
99 fail:
100 	for_each_possible_cpu(cpu)
101 		kfree(entries->cpu_entries[cpu]);
102 	kfree(entries);
103 
104 	return -ENOMEM;
105 }
106 
107 int get_callchain_buffers(int event_max_stack)
108 {
109 	int err = 0;
110 	int count;
111 
112 	mutex_lock(&callchain_mutex);
113 
114 	count = atomic_inc_return(&nr_callchain_events);
115 	if (WARN_ON_ONCE(count < 1)) {
116 		err = -EINVAL;
117 		goto exit;
118 	}
119 
120 	if (count > 1) {
121 		/* If the allocation failed, give up */
122 		if (!callchain_cpus_entries)
123 			err = -ENOMEM;
124 		/*
125 		 * If requesting per event more than the global cap,
126 		 * return a different error to help userspace figure
127 		 * this out.
128 		 *
129 		 * And also do it here so that we have &callchain_mutex held.
130 		 */
131 		if (event_max_stack > sysctl_perf_event_max_stack)
132 			err = -EOVERFLOW;
133 		goto exit;
134 	}
135 
136 	err = alloc_callchain_buffers();
137 exit:
138 	if (err)
139 		atomic_dec(&nr_callchain_events);
140 
141 	mutex_unlock(&callchain_mutex);
142 
143 	return err;
144 }
145 
146 void put_callchain_buffers(void)
147 {
148 	if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
149 		release_callchain_buffers();
150 		mutex_unlock(&callchain_mutex);
151 	}
152 }
153 
154 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
155 {
156 	int cpu;
157 	struct callchain_cpus_entries *entries;
158 
159 	*rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
160 	if (*rctx == -1)
161 		return NULL;
162 
163 	entries = rcu_dereference(callchain_cpus_entries);
164 	if (!entries)
165 		return NULL;
166 
167 	cpu = smp_processor_id();
168 
169 	return (((void *)entries->cpu_entries[cpu]) +
170 		(*rctx * perf_callchain_entry__sizeof()));
171 }
172 
173 static void
174 put_callchain_entry(int rctx)
175 {
176 	put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
177 }
178 
179 struct perf_callchain_entry *
180 perf_callchain(struct perf_event *event, struct pt_regs *regs)
181 {
182 	bool kernel = !event->attr.exclude_callchain_kernel;
183 	bool user   = !event->attr.exclude_callchain_user;
184 	/* Disallow cross-task user callchains. */
185 	bool crosstask = event->ctx->task && event->ctx->task != current;
186 	const u32 max_stack = event->attr.sample_max_stack;
187 
188 	if (!kernel && !user)
189 		return NULL;
190 
191 	return get_perf_callchain(regs, 0, kernel, user, max_stack, crosstask, true);
192 }
193 
194 struct perf_callchain_entry *
195 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
196 		   u32 max_stack, bool crosstask, bool add_mark)
197 {
198 	struct perf_callchain_entry *entry;
199 	struct perf_callchain_entry_ctx ctx;
200 	int rctx;
201 
202 	entry = get_callchain_entry(&rctx);
203 	if (rctx == -1)
204 		return NULL;
205 
206 	if (!entry)
207 		goto exit_put;
208 
209 	ctx.entry     = entry;
210 	ctx.max_stack = max_stack;
211 	ctx.nr	      = entry->nr = init_nr;
212 	ctx.contexts       = 0;
213 	ctx.contexts_maxed = false;
214 
215 	if (kernel && !user_mode(regs)) {
216 		if (add_mark)
217 			perf_callchain_store_context(&ctx, PERF_CONTEXT_KERNEL);
218 		perf_callchain_kernel(&ctx, regs);
219 	}
220 
221 	if (user) {
222 		if (!user_mode(regs)) {
223 			if  (current->mm)
224 				regs = task_pt_regs(current);
225 			else
226 				regs = NULL;
227 		}
228 
229 		if (regs) {
230 			if (crosstask)
231 				goto exit_put;
232 
233 			if (add_mark)
234 				perf_callchain_store_context(&ctx, PERF_CONTEXT_USER);
235 			perf_callchain_user(&ctx, regs);
236 		}
237 	}
238 
239 exit_put:
240 	put_callchain_entry(rctx);
241 
242 	return entry;
243 }
244 
245 /*
246  * Used for sysctl_perf_event_max_stack and
247  * sysctl_perf_event_max_contexts_per_stack.
248  */
249 int perf_event_max_stack_handler(struct ctl_table *table, int write,
250 				 void __user *buffer, size_t *lenp, loff_t *ppos)
251 {
252 	int *value = table->data;
253 	int new_value = *value, ret;
254 	struct ctl_table new_table = *table;
255 
256 	new_table.data = &new_value;
257 	ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
258 	if (ret || !write)
259 		return ret;
260 
261 	mutex_lock(&callchain_mutex);
262 	if (atomic_read(&nr_callchain_events))
263 		ret = -EBUSY;
264 	else
265 		*value = new_value;
266 
267 	mutex_unlock(&callchain_mutex);
268 
269 	return ret;
270 }
271