kernel/events/callchain.c

   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         /*
 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                 err = alloc_callchain_buffers();
 134 exit:
 135         if (err)
 136                 atomic_dec(&nr_callchain_events);
 137
 138         mutex_unlock(&callchain_mutex);
 139
 140         return err;
 141 }
 142
 143 void put_callchain_buffers(void)
 144 {
 145         if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
 146                 release_callchain_buffers();
 147                 mutex_unlock(&callchain_mutex);
 148         }
 149 }
 150
 151 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
 152 {
 153         int cpu;
 154         struct callchain_cpus_entries *entries;
 155
 156         *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
 157         if (*rctx == -1)
 158                 return NULL;
 159
 160         entries = rcu_dereference(callchain_cpus_entries);
 161         if (!entries)
 162                 return NULL;
 163
 164         cpu = smp_processor_id();
 165
 166         return (((void *)entries->cpu_entries[cpu]) +
 167                 (*rctx * perf_callchain_entry__sizeof()));
 168 }
 169
 170 static void
 171 put_callchain_entry(int rctx)
 172 {
 173         put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
 174 }
 175
 176 struct perf_callchain_entry *
 177 perf_callchain(struct perf_event *event, struct pt_regs *regs)
 178 {
 179         bool kernel = !event->attr.exclude_callchain_kernel;
 180         bool user   = !event->attr.exclude_callchain_user;
 181         /* Disallow cross-task user callchains. */
 182         bool crosstask = event->ctx->task && event->ctx->task != current;
 183         const u32 max_stack = event->attr.sample_max_stack;
 184
 185         if (!kernel && !user)
 186                 return NULL;
 187
 188         return get_perf_callchain(regs, 0, kernel, user, max_stack, crosstask, true);
 189 }
 190
 191 struct perf_callchain_entry *
 192 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
 193                    u32 max_stack, bool crosstask, bool add_mark)
 194 {
 195         struct perf_callchain_entry *entry;
 196         struct perf_callchain_entry_ctx ctx;
 197         int rctx;
 198
 199         entry = get_callchain_entry(&rctx);
 200         if (rctx == -1)
 201                 return NULL;
 202
 203         if (!entry)
 204                 goto exit_put;
 205
 206         ctx.entry     = entry;
 207         ctx.max_stack = max_stack;
 208         ctx.nr        = entry->nr = init_nr;
 209         ctx.contexts       = 0;
 210         ctx.contexts_maxed = false;
 211
 212         if (kernel && !user_mode(regs)) {
 213                 if (add_mark)
 214                         perf_callchain_store_context(&ctx, PERF_CONTEXT_KERNEL);
 215                 perf_callchain_kernel(&ctx, regs);
 216         }
 217
 218         if (user) {
 219                 if (!user_mode(regs)) {
 220                         if  (current->mm)
 221                                 regs = task_pt_regs(current);
 222                         else
 223                                 regs = NULL;
 224                 }
 225
 226                 if (regs) {
 227                         mm_segment_t fs;
 228
 229                         if (crosstask)
 230                                 goto exit_put;
 231
 232                         if (add_mark)
 233                                 perf_callchain_store_context(&ctx, PERF_CONTEXT_USER);
 234
 235                         fs = get_fs();
 236                         set_fs(USER_DS);
 237                         perf_callchain_user(&ctx, regs);
 238                         set_fs(fs);
 239                 }
 240         }
 241
 242 exit_put:
 243         put_callchain_entry(rctx);
 244
 245         return entry;
 246 }
 247
 248 /*
 249  * Used for sysctl_perf_event_max_stack and
 250  * sysctl_perf_event_max_contexts_per_stack.
 251  */
 252 int perf_event_max_stack_handler(struct ctl_table *table, int write,
 253                                  void __user *buffer, size_t *lenp, loff_t *ppos)
 254 {
 255         int *value = table->data;
 256         int new_value = *value, ret;
 257         struct ctl_table new_table = *table;
 258
 259         new_table.data = &new_value;
 260         ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
 261         if (ret || !write)
 262                 return ret;
 263
 264         mutex_lock(&callchain_mutex);
 265         if (atomic_read(&nr_callchain_events))
 266                 ret = -EBUSY;
 267         else
 268                 *value = new_value;
 269
 270         mutex_unlock(&callchain_mutex);
 271
 272         return ret;
 273 }