kernel/events/callchain.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Performance events callchain code, extracted from core.c:
   4  *
   5  *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
   6  *  Copyright (C) 2008-2011 Red Hat, Inc., Ingo Molnar
   7  *  Copyright (C) 2008-2011 Red Hat, Inc., Peter Zijlstra
   8  *  Copyright  ©  2009 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   9  */
  10
  11 #include <linux/perf_event.h>
  12 #include <linux/slab.h>
  13 #include <linux/sched/task_stack.h>
  14
  15 #include "internal.h"
  16
  17 struct callchain_cpus_entries {
  18         struct rcu_head                 rcu_head;
  19         struct perf_callchain_entry     *cpu_entries[0];
  20 };
  21
  22 int sysctl_perf_event_max_stack __read_mostly = PERF_MAX_STACK_DEPTH;
  23 int sysctl_perf_event_max_contexts_per_stack __read_mostly = PERF_MAX_CONTEXTS_PER_STACK;
  24
  25 static inline size_t perf_callchain_entry__sizeof(void)
  26 {
  27         return (sizeof(struct perf_callchain_entry) +
  28                 sizeof(__u64) * (sysctl_perf_event_max_stack +
  29                                  sysctl_perf_event_max_contexts_per_stack));
  30 }
  31
  32 static DEFINE_PER_CPU(int, callchain_recursion[PERF_NR_CONTEXTS]);
  33 static atomic_t nr_callchain_events;
  34 static DEFINE_MUTEX(callchain_mutex);
  35 static struct callchain_cpus_entries *callchain_cpus_entries;
  36
  37
  38 __weak void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
  39                                   struct pt_regs *regs)
  40 {
  41 }
  42
  43 __weak void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
  44                                 struct pt_regs *regs)
  45 {
  46 }
  47
  48 static void release_callchain_buffers_rcu(struct rcu_head *head)
  49 {
  50         struct callchain_cpus_entries *entries;
  51         int cpu;
  52
  53         entries = container_of(head, struct callchain_cpus_entries, rcu_head);
  54
  55         for_each_possible_cpu(cpu)
  56                 kfree(entries->cpu_entries[cpu]);
  57
  58         kfree(entries);
  59 }
  60
  61 static void release_callchain_buffers(void)
  62 {
  63         struct callchain_cpus_entries *entries;
  64
  65         entries = callchain_cpus_entries;
  66         RCU_INIT_POINTER(callchain_cpus_entries, NULL);
  67         call_rcu(&entries->rcu_head, release_callchain_buffers_rcu);
  68 }
  69
  70 static int alloc_callchain_buffers(void)
  71 {
  72         int cpu;
  73         int size;
  74         struct callchain_cpus_entries *entries;
  75
  76         /*
  77          * We can't use the percpu allocation API for data that can be
  78          * accessed from NMI. Use a temporary manual per cpu allocation
  79          * until that gets sorted out.
  80          */
  81         size = offsetof(struct callchain_cpus_entries, cpu_entries[nr_cpu_ids]);
  82
  83         entries = kzalloc(size, GFP_KERNEL);
  84         if (!entries)
  85                 return -ENOMEM;
  86
  87         size = perf_callchain_entry__sizeof() * PERF_NR_CONTEXTS;
  88
  89         for_each_possible_cpu(cpu) {
  90                 entries->cpu_entries[cpu] = kmalloc_node(size, GFP_KERNEL,
  91                                                          cpu_to_node(cpu));
  92                 if (!entries->cpu_entries[cpu])
  93                         goto fail;
  94         }
  95
  96         rcu_assign_pointer(callchain_cpus_entries, entries);
  97
  98         return 0;
  99
 100 fail:
 101         for_each_possible_cpu(cpu)
 102                 kfree(entries->cpu_entries[cpu]);
 103         kfree(entries);
 104
 105         return -ENOMEM;
 106 }
 107
 108 int get_callchain_buffers(int event_max_stack)
 109 {
 110         int err = 0;
 111         int count;
 112
 113         mutex_lock(&callchain_mutex);
 114
 115         count = atomic_inc_return(&nr_callchain_events);
 116         if (WARN_ON_ONCE(count < 1)) {
 117                 err = -EINVAL;
 118                 goto exit;
 119         }
 120
 121         /*
 122          * If requesting per event more than the global cap,
 123          * return a different error to help userspace figure
 124          * this out.
 125          *
 126          * And also do it here so that we have &callchain_mutex held.
 127          */
 128         if (event_max_stack > sysctl_perf_event_max_stack) {
 129                 err = -EOVERFLOW;
 130                 goto exit;
 131         }
 132
 133         if (count == 1)
 134                 err = alloc_callchain_buffers();
 135 exit:
 136         if (err)
 137                 atomic_dec(&nr_callchain_events);
 138
 139         mutex_unlock(&callchain_mutex);
 140
 141         return err;
 142 }
 143
 144 void put_callchain_buffers(void)
 145 {
 146         if (atomic_dec_and_mutex_lock(&nr_callchain_events, &callchain_mutex)) {
 147                 release_callchain_buffers();
 148                 mutex_unlock(&callchain_mutex);
 149         }
 150 }
 151
 152 static struct perf_callchain_entry *get_callchain_entry(int *rctx)
 153 {
 154         int cpu;
 155         struct callchain_cpus_entries *entries;
 156
 157         *rctx = get_recursion_context(this_cpu_ptr(callchain_recursion));
 158         if (*rctx == -1)
 159                 return NULL;
 160
 161         entries = rcu_dereference(callchain_cpus_entries);
 162         if (!entries)
 163                 return NULL;
 164
 165         cpu = smp_processor_id();
 166
 167         return (((void *)entries->cpu_entries[cpu]) +
 168                 (*rctx * perf_callchain_entry__sizeof()));
 169 }
 170
 171 static void
 172 put_callchain_entry(int rctx)
 173 {
 174         put_recursion_context(this_cpu_ptr(callchain_recursion), rctx);
 175 }
 176
 177 struct perf_callchain_entry *
 178 get_perf_callchain(struct pt_regs *regs, u32 init_nr, bool kernel, bool user,
 179                    u32 max_stack, bool crosstask, bool add_mark)
 180 {
 181         struct perf_callchain_entry *entry;
 182         struct perf_callchain_entry_ctx ctx;
 183         int rctx;
 184
 185         entry = get_callchain_entry(&rctx);
 186         if (rctx == -1)
 187                 return NULL;
 188
 189         if (!entry)
 190                 goto exit_put;
 191
 192         ctx.entry     = entry;
 193         ctx.max_stack = max_stack;
 194         ctx.nr        = entry->nr = init_nr;
 195         ctx.contexts       = 0;
 196         ctx.contexts_maxed = false;
 197
 198         if (kernel && !user_mode(regs)) {
 199                 if (add_mark)
 200                         perf_callchain_store_context(&ctx, PERF_CONTEXT_KERNEL);
 201                 perf_callchain_kernel(&ctx, 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                         mm_segment_t fs;
 214
 215                         if (crosstask)
 216                                 goto exit_put;
 217
 218                         if (add_mark)
 219                                 perf_callchain_store_context(&ctx, PERF_CONTEXT_USER);
 220
 221                         fs = get_fs();
 222                         set_fs(USER_DS);
 223                         perf_callchain_user(&ctx, regs);
 224                         set_fs(fs);
 225                 }
 226         }
 227
 228 exit_put:
 229         put_callchain_entry(rctx);
 230
 231         return entry;
 232 }
 233
 234 /*
 235  * Used for sysctl_perf_event_max_stack and
 236  * sysctl_perf_event_max_contexts_per_stack.
 237  */
 238 int perf_event_max_stack_handler(struct ctl_table *table, int write,
 239                                  void __user *buffer, size_t *lenp, loff_t *ppos)
 240 {
 241         int *value = table->data;
 242         int new_value = *value, ret;
 243         struct ctl_table new_table = *table;
 244
 245         new_table.data = &new_value;
 246         ret = proc_dointvec_minmax(&new_table, write, buffer, lenp, ppos);
 247         if (ret || !write)
 248                 return ret;
 249
 250         mutex_lock(&callchain_mutex);
 251         if (atomic_read(&nr_callchain_events))
 252                 ret = -EBUSY;
 253         else
 254                 *value = new_value;
 255
 256         mutex_unlock(&callchain_mutex);
 257
 258         return ret;
 259 }