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 18 struct callchain_cpus_entries { 19 struct rcu_head rcu_head; 20 struct perf_callchain_entry *cpu_entries[0]; 21 }; 22 23 int sysctl_perf_event_max_stack __read_mostly = PERF_MAX_STACK_DEPTH; 24 int sysctl_perf_event_max_contexts_per_stack __read_mostly = PERF_MAX_CONTEXTS_PER_STACK; 25 26 static 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 33 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]); 34 static atomic_t nr_callchain_events; 35 static DEFINE_MUTEX(callchain_mutex); 36 static 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 49 static 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 62 static 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 71 static 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; 100 101 fail: 102 for_each_possible_cpu(cpu) 103 kfree(entries->cpu_entries[cpu]); 104 kfree(entries); 105 106 return -ENOMEM; 107 } 108 109 int 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(); 136 exit: 137 if (err) 138 atomic_dec(&nr_callchain_events); 139 140 mutex_unlock(&callchain_mutex); 141 142 return err; 143 } 144 145 void 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 } 152 153 static 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 } 171 172 static void 173 put_callchain_entry(int rctx) 174 { 175 put_recursion_context(this_cpu_ptr(callchain_recursion), rctx); 176 } 177 178 struct perf_callchain_entry * 179 get_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 } 228 229 exit_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 */ 239 int 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 } 261