Linux Audio

Check our new training course

Loading...
v4.10.11
  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
 21int sysctl_perf_event_max_stack __read_mostly = PERF_MAX_STACK_DEPTH;
 22int sysctl_perf_event_max_contexts_per_stack __read_mostly = PERF_MAX_CONTEXTS_PER_STACK;
 23
 24static 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
 31static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
 32static atomic_t nr_callchain_events;
 33static DEFINE_MUTEX(callchain_mutex);
 34static 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
 47static 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
 60static 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
 69static 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
 99fail:
100	for_each_possible_cpu(cpu)
101		kfree(entries->cpu_entries[cpu]);
102	kfree(entries);
103
104	return -ENOMEM;
105}
106
107int 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();
 
137exit:
138	if (err)
139		atomic_dec(&nr_callchain_events);
140
141	mutex_unlock(&callchain_mutex);
142
143	return err;
144}
145
146void 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
154static 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
173static void
174put_callchain_entry(int rctx)
175{
176	put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
177}
178
179struct perf_callchain_entry *
180perf_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
194struct perf_callchain_entry *
195get_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
239exit_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 */
249int 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}
v4.17
  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 <linux/sched/task_stack.h>
 15
 16#include "internal.h"
 17
 18struct callchain_cpus_entries {
 19	struct rcu_head			rcu_head;
 20	struct perf_callchain_entry	*cpu_entries[0];
 21};
 22
 23int sysctl_perf_event_max_stack __read_mostly = PERF_MAX_STACK_DEPTH;
 24int sysctl_perf_event_max_contexts_per_stack __read_mostly = PERF_MAX_CONTEXTS_PER_STACK;
 25
 26static inline size_t perf_callchain_entry__sizeof(void)
 27{
 28	return (sizeof(struct perf_callchain_entry) +
 29		sizeof(__u64) * (sysctl_perf_event_max_stack +
 30				 sysctl_perf_event_max_contexts_per_stack));
 31}
 32
 33static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
 34static atomic_t nr_callchain_events;
 35static DEFINE_MUTEX(callchain_mutex);
 36static struct callchain_cpus_entries *callchain_cpus_entries;
 37
 38
 39__weak void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
 40				  struct pt_regs *regs)
 41{
 42}
 43
 44__weak void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
 45				struct pt_regs *regs)
 46{
 47}
 48
 49static void release_callchain_buffers_rcu(struct rcu_head *head)
 50{
 51	struct callchain_cpus_entries *entries;
 52	int cpu;
 53
 54	entries = container_of(head, struct callchain_cpus_entries, rcu_head);
 55
 56	for_each_possible_cpu(cpu)
 57		kfree(entries->cpu_entries[cpu]);
 58
 59	kfree(entries);
 60}
 61
 62static void release_callchain_buffers(void)
 63{
 64	struct callchain_cpus_entries *entries;
 65
 66	entries = callchain_cpus_entries;
 67	RCU_INIT_POINTER(callchain_cpus_entries, NULL);
 68	call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
 69}
 70
 71static int alloc_callchain_buffers(void)
 72{
 73	int cpu;
 74	int size;
 75	struct callchain_cpus_entries *entries;
 76
 77	/*
 78	 * We can't use the percpu allocation API for data that can be
 79	 * accessed from NMI. Use a temporary manual per cpu allocation
 80	 * until that gets sorted out.
 81	 */
 82	size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
 83
 84	entries = kzalloc(size, GFP_KERNEL);
 85	if (!entries)
 86		return -ENOMEM;
 87
 88	size = perf_callchain_entry__sizeof() * PERF_NR_CONTEXTS;
 89
 90	for_each_possible_cpu(cpu) {
 91		entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
 92							 cpu_to_node(cpu));
 93		if (!entries->cpu_entries[cpu])
 94			goto fail;
 95	}
 96
 97	rcu_assign_pointer(callchain_cpus_entries, entries);
 98
 99	return 0;
100
101fail:
102	for_each_possible_cpu(cpu)
103		kfree(entries->cpu_entries[cpu]);
104	kfree(entries);
105
106	return -ENOMEM;
107}
108
109int get_callchain_buffers(int event_max_stack)
110{
111	int err = 0;
112	int count;
113
114	mutex_lock(&callchain_mutex);
115
116	count = atomic_inc_return(&nr_callchain_events);
117	if (WARN_ON_ONCE(count < 1)) {
118		err = -EINVAL;
119		goto exit;
120	}
121
122	/*
123	 * If requesting per event more than the global cap,
124	 * return a different error to help userspace figure
125	 * this out.
126	 *
127	 * And also do it here so that we have &callchain_mutex held.
128	 */
129	if (event_max_stack > sysctl_perf_event_max_stack) {
130		err = -EOVERFLOW;
 
 
 
 
131		goto exit;
132	}
133
134	if (count == 1)
135		err = alloc_callchain_buffers();
136exit:
137	if (err)
138		atomic_dec(&nr_callchain_events);
139
140	mutex_unlock(&callchain_mutex);
141
142	return err;
143}
144
145void put_callchain_buffers(void)
146{
147	if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
148		release_callchain_buffers();
149		mutex_unlock(&callchain_mutex);
150	}
151}
152
153static struct perf_callchain_entry *get_callchain_entry(int *rctx)
154{
155	int cpu;
156	struct callchain_cpus_entries *entries;
157
158	*rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
159	if (*rctx == -1)
160		return NULL;
161
162	entries = rcu_dereference(callchain_cpus_entries);
163	if (!entries)
164		return NULL;
165
166	cpu = smp_processor_id();
167
168	return (((void *)entries->cpu_entries[cpu]) +
169		(*rctx * perf_callchain_entry__sizeof()));
170}
171
172static void
173put_callchain_entry(int rctx)
174{
175	put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
176}
177
178struct perf_callchain_entry *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
179get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
180		   u32 max_stack, bool crosstask, bool add_mark)
181{
182	struct perf_callchain_entry *entry;
183	struct perf_callchain_entry_ctx ctx;
184	int rctx;
185
186	entry = get_callchain_entry(&rctx);
187	if (rctx == -1)
188		return NULL;
189
190	if (!entry)
191		goto exit_put;
192
193	ctx.entry     = entry;
194	ctx.max_stack = max_stack;
195	ctx.nr	      = entry->nr = init_nr;
196	ctx.contexts       = 0;
197	ctx.contexts_maxed = false;
198
199	if (kernel && !user_mode(regs)) {
200		if (add_mark)
201			perf_callchain_store_context(&ctx, PERF_CONTEXT_KERNEL);
202		perf_callchain_kernel(&ctx, regs);
203	}
204
205	if (user) {
206		if (!user_mode(regs)) {
207			if  (current->mm)
208				regs = task_pt_regs(current);
209			else
210				regs = NULL;
211		}
212
213		if (regs) {
214			mm_segment_t fs;
215
216			if (crosstask)
217				goto exit_put;
218
219			if (add_mark)
220				perf_callchain_store_context(&ctx, PERF_CONTEXT_USER);
221
222			fs = get_fs();
223			set_fs(USER_DS);
224			perf_callchain_user(&ctx, regs);
225			set_fs(fs);
226		}
227	}
228
229exit_put:
230	put_callchain_entry(rctx);
231
232	return entry;
233}
234
235/*
236 * Used for sysctl_perf_event_max_stack and
237 * sysctl_perf_event_max_contexts_per_stack.
238 */
239int perf_event_max_stack_handler(struct ctl_table *table, int write,
240				 void __user *buffer, size_t *lenp, loff_t *ppos)
241{
242	int *value = table->data;
243	int new_value = *value, ret;
244	struct ctl_table new_table = *table;
245
246	new_table.data = &new_value;
247	ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
248	if (ret || !write)
249		return ret;
250
251	mutex_lock(&callchain_mutex);
252	if (atomic_read(&nr_callchain_events))
253		ret = -EBUSY;
254	else
255		*value = new_value;
256
257	mutex_unlock(&callchain_mutex);
258
259	return ret;
260}