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 23 static inline size_t perf_callchain_entry__sizeof(void) 24 { 25 return (sizeof(struct perf_callchain_entry) + 26 sizeof(__u64) * sysctl_perf_event_max_stack); 27 } 28 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; 33 34 35 __weak void perf_callchain_kernel(struct perf_callchain_entry *entry, 36 struct pt_regs *regs) 37 { 38 } 39 40 __weak void perf_callchain_user(struct perf_callchain_entry *entry, 41 struct pt_regs *regs) 42 { 43 } 44 45 static void release_callchain_buffers_rcu(struct rcu_head *head) 46 { 47 struct callchain_cpus_entries *entries; 48 int cpu; 49 50 entries = container_of(head, struct callchain_cpus_entries, rcu_head); 51 52 for_each_possible_cpu(cpu) 53 kfree(entries->cpu_entries[cpu]); 54 55 kfree(entries); 56 } 57 58 static void release_callchain_buffers(void) 59 { 60 struct callchain_cpus_entries *entries; 61 62 entries = callchain_cpus_entries; 63 RCU_INIT_POINTER(callchain_cpus_entries, NULL); 64 call_rcu(&entries->rcu_head, release_callchain_buffers_rcu); 65 } 66 67 static int alloc_callchain_buffers(void) 68 { 69 int cpu; 70 int size; 71 struct callchain_cpus_entries *entries; 72 73 /* 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. 77 */ 78 size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]); 79 80 entries = kzalloc(size, GFP_KERNEL); 81 if (!entries) 82 return -ENOMEM; 83 84 size = perf_callchain_entry__sizeof() * PERF_NR_CONTEXTS; 85 86 for_each_possible_cpu(cpu) { 87 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL, 88 cpu_to_node(cpu)); 89 if (!entries->cpu_entries[cpu]) 90 goto fail; 91 } 92 93 rcu_assign_pointer(callchain_cpus_entries, entries); 94 95 return 0; 96 97 fail: 98 for_each_possible_cpu(cpu) 99 kfree(entries->cpu_entries[cpu]); 100 kfree(entries); 101 102 return -ENOMEM; 103 } 104 105 int get_callchain_buffers(void) 106 { 107 int err = 0; 108 int count; 109 110 mutex_lock(&callchain_mutex); 111 112 count = atomic_inc_return(&nr_callchain_events); 113 if (WARN_ON_ONCE(count < 1)) { 114 err = -EINVAL; 115 goto exit; 116 } 117 118 if (count > 1) { 119 /* If the allocation failed, give up */ 120 if (!callchain_cpus_entries) 121 err = -ENOMEM; 122 goto exit; 123 } 124 125 err = alloc_callchain_buffers(); 126 exit: 127 if (err) 128 atomic_dec(&nr_callchain_events); 129 130 mutex_unlock(&callchain_mutex); 131 132 return err; 133 } 134 135 void put_callchain_buffers(void) 136 { 137 if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) { 138 release_callchain_buffers(); 139 mutex_unlock(&callchain_mutex); 140 } 141 } 142 143 static struct perf_callchain_entry *get_callchain_entry(int *rctx) 144 { 145 int cpu; 146 struct callchain_cpus_entries *entries; 147 148 *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion)); 149 if (*rctx == -1) 150 return NULL; 151 152 entries = rcu_dereference(callchain_cpus_entries); 153 if (!entries) 154 return NULL; 155 156 cpu = smp_processor_id(); 157 158 return (((void *)entries->cpu_entries[cpu]) + 159 (*rctx * perf_callchain_entry__sizeof())); 160 } 161 162 static void 163 put_callchain_entry(int rctx) 164 { 165 put_recursion_context(this_cpu_ptr(callchain_recursion), rctx); 166 } 167 168 struct perf_callchain_entry * 169 perf_callchain(struct perf_event *event, struct pt_regs *regs) 170 { 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; 175 176 if (!kernel && !user) 177 return NULL; 178 179 return get_perf_callchain(regs, 0, kernel, user, crosstask, true); 180 } 181 182 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) 185 { 186 struct perf_callchain_entry *entry; 187 int rctx; 188 189 entry = get_callchain_entry(&rctx); 190 if (rctx == -1) 191 return NULL; 192 193 if (!entry) 194 goto exit_put; 195 196 entry->nr = init_nr; 197 198 if (kernel && !user_mode(regs)) { 199 if (add_mark) 200 perf_callchain_store(entry, PERF_CONTEXT_KERNEL); 201 perf_callchain_kernel(entry, regs); 202 } 203 204 if (user) { 205 if (!user_mode(regs)) { 206 if (current->mm) 207 regs = task_pt_regs(current); 208 else 209 regs = NULL; 210 } 211 212 if (regs) { 213 if (crosstask) 214 goto exit_put; 215 216 if (add_mark) 217 perf_callchain_store(entry, PERF_CONTEXT_USER); 218 perf_callchain_user(entry, regs); 219 } 220 } 221 222 exit_put: 223 put_callchain_entry(rctx); 224 225 return entry; 226 } 227 228 int perf_event_max_stack_handler(struct ctl_table *table, int write, 229 void __user *buffer, size_t *lenp, loff_t *ppos) 230 { 231 int new_value = sysctl_perf_event_max_stack, ret; 232 struct ctl_table new_table = *table; 233 234 new_table.data = &new_value; 235 ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos); 236 if (ret || !write) 237 return ret; 238 239 mutex_lock(&callchain_mutex); 240 if (atomic_read(&nr_callchain_events)) 241 ret = -EBUSY; 242 else 243 sysctl_perf_event_max_stack = new_value; 244 245 mutex_unlock(&callchain_mutex); 246 247 return ret; 248 } 249