• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
perf_callchain_entry__sizeof(void)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 
perf_callchain_kernel(struct perf_callchain_entry_ctx * entry,struct pt_regs * regs)37 __weak void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
38 				  struct pt_regs *regs)
39 {
40 }
41 
perf_callchain_user(struct perf_callchain_entry_ctx * entry,struct pt_regs * regs)42 __weak void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
43 				struct pt_regs *regs)
44 {
45 }
46 
release_callchain_buffers_rcu(struct rcu_head * head)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 
release_callchain_buffers(void)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 
alloc_callchain_buffers(void)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 
get_callchain_buffers(int event_max_stack)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 	/*
121 	 * If requesting per event more than the global cap,
122 	 * return a different error to help userspace figure
123 	 * this out.
124 	 *
125 	 * And also do it here so that we have &callchain_mutex held.
126 	 */
127 	if (event_max_stack > sysctl_perf_event_max_stack) {
128 		err = -EOVERFLOW;
129 		goto exit;
130 	}
131 
132 	if (count > 1) {
133 		/* If the allocation failed, give up */
134 		if (!callchain_cpus_entries)
135 			err = -ENOMEM;
136 		goto exit;
137 	}
138 
139 	err = alloc_callchain_buffers();
140 exit:
141 	if (err)
142 		atomic_dec(&nr_callchain_events);
143 
144 	mutex_unlock(&callchain_mutex);
145 
146 	return err;
147 }
148 
put_callchain_buffers(void)149 void put_callchain_buffers(void)
150 {
151 	if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
152 		release_callchain_buffers();
153 		mutex_unlock(&callchain_mutex);
154 	}
155 }
156 
get_callchain_entry(int * rctx)157 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
158 {
159 	int cpu;
160 	struct callchain_cpus_entries *entries;
161 
162 	*rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
163 	if (*rctx == -1)
164 		return NULL;
165 
166 	entries = rcu_dereference(callchain_cpus_entries);
167 	if (!entries)
168 		return NULL;
169 
170 	cpu = smp_processor_id();
171 
172 	return (((void *)entries->cpu_entries[cpu]) +
173 		(*rctx * perf_callchain_entry__sizeof()));
174 }
175 
176 static void
put_callchain_entry(int rctx)177 put_callchain_entry(int rctx)
178 {
179 	put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
180 }
181 
182 struct perf_callchain_entry *
perf_callchain(struct perf_event * event,struct pt_regs * regs)183 perf_callchain(struct perf_event *event, struct pt_regs *regs)
184 {
185 	bool kernel = !event->attr.exclude_callchain_kernel;
186 	bool user   = !event->attr.exclude_callchain_user;
187 	/* Disallow cross-task user callchains. */
188 	bool crosstask = event->ctx->task && event->ctx->task != current;
189 	const u32 max_stack = event->attr.sample_max_stack;
190 
191 	if (!kernel && !user)
192 		return NULL;
193 
194 	return get_perf_callchain(regs, 0, kernel, user, max_stack, crosstask, true);
195 }
196 
197 struct perf_callchain_entry *
get_perf_callchain(struct pt_regs * regs,u32 init_nr,bool kernel,bool user,u32 max_stack,bool crosstask,bool add_mark)198 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
199 		   u32 max_stack, bool crosstask, bool add_mark)
200 {
201 	struct perf_callchain_entry *entry;
202 	struct perf_callchain_entry_ctx ctx;
203 	int rctx;
204 
205 	entry = get_callchain_entry(&rctx);
206 	if (rctx == -1)
207 		return NULL;
208 
209 	if (!entry)
210 		goto exit_put;
211 
212 	ctx.entry     = entry;
213 	ctx.max_stack = max_stack;
214 	ctx.nr	      = entry->nr = init_nr;
215 	ctx.contexts       = 0;
216 	ctx.contexts_maxed = false;
217 
218 	if (kernel && !user_mode(regs)) {
219 		if (add_mark)
220 			perf_callchain_store_context(&ctx, PERF_CONTEXT_KERNEL);
221 		perf_callchain_kernel(&ctx, regs);
222 	}
223 
224 	if (user) {
225 		if (!user_mode(regs)) {
226 			if  (current->mm)
227 				regs = task_pt_regs(current);
228 			else
229 				regs = NULL;
230 		}
231 
232 		if (regs) {
233 			mm_segment_t fs;
234 
235 			if (crosstask)
236 				goto exit_put;
237 
238 			if (add_mark)
239 				perf_callchain_store_context(&ctx, PERF_CONTEXT_USER);
240 
241 			fs = get_fs();
242 			set_fs(USER_DS);
243 			perf_callchain_user(&ctx, regs);
244 			set_fs(fs);
245 		}
246 	}
247 
248 exit_put:
249 	put_callchain_entry(rctx);
250 
251 	return entry;
252 }
253 
254 /*
255  * Used for sysctl_perf_event_max_stack and
256  * sysctl_perf_event_max_contexts_per_stack.
257  */
perf_event_max_stack_handler(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)258 int perf_event_max_stack_handler(struct ctl_table *table, int write,
259 				 void __user *buffer, size_t *lenp, loff_t *ppos)
260 {
261 	int *value = table->data;
262 	int new_value = *value, ret;
263 	struct ctl_table new_table = *table;
264 
265 	new_table.data = &new_value;
266 	ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
267 	if (ret || !write)
268 		return ret;
269 
270 	mutex_lock(&callchain_mutex);
271 	if (atomic_read(&nr_callchain_events))
272 		ret = -EBUSY;
273 	else
274 		*value = new_value;
275 
276 	mutex_unlock(&callchain_mutex);
277 
278 	return ret;
279 }
280