kernel/trace/trace_workqueue.c

   1 /*
   2  * Workqueue statistical tracer.
   3  *
   4  * Copyright (C) 2008 Frederic Weisbecker <fweisbec@gmail.com>
   5  *
   6  */
   7
   8
   9 #include <trace/events/workqueue.h>
  10 #include <linux/list.h>
  11 #include <linux/percpu.h>
  12 #include <linux/kref.h>
  13 #include "trace_stat.h"
  14 #include "trace.h"
  15
  16
  17 /* A cpu workqueue thread */
  18 struct cpu_workqueue_stats {
  19         struct list_head            list;
  20         struct kref                 kref;
  21         int                         cpu;
  22         pid_t                       pid;
  23 /* Can be inserted from interrupt or user context, need to be atomic */
  24         atomic_t                    inserted;
  25 /*
  26  *  Don't need to be atomic, works are serialized in a single workqueue thread
  27  *  on a single CPU.
  28  */
  29         unsigned int                executed;
  30 };
  31
  32 /* List of workqueue threads on one cpu */
  33 struct workqueue_global_stats {
  34         struct list_head        list;
  35         spinlock_t              lock;
  36 };
  37
  38 /* Don't need a global lock because allocated before the workqueues, and
  39  * never freed.
  40  */
  41 static DEFINE_PER_CPU(struct workqueue_global_stats, all_workqueue_stat);
  42 #define workqueue_cpu_stat(cpu) (&per_cpu(all_workqueue_stat, cpu))
  43
  44 static void cpu_workqueue_stat_free(struct kref *kref)
  45 {
  46         kfree(container_of(kref, struct cpu_workqueue_stats, kref));
  47 }
  48
  49 /* Insertion of a work */
  50 static void
  51 probe_workqueue_insertion(struct task_struct *wq_thread,
  52                           struct work_struct *work)
  53 {
  54         int cpu = cpumask_first(&wq_thread->cpus_allowed);
  55         struct cpu_workqueue_stats *node;
  56         unsigned long flags;
  57
  58         spin_lock_irqsave(&workqueue_cpu_stat(cpu)->lock, flags);
  59         list_for_each_entry(node, &workqueue_cpu_stat(cpu)->list, list) {
  60                 if (node->pid == wq_thread->pid) {
  61                         atomic_inc(&node->inserted);
  62                         goto found;
  63                 }
  64         }
  65         pr_debug("trace_workqueue: entry not found\n");
  66 found:
  67         spin_unlock_irqrestore(&workqueue_cpu_stat(cpu)->lock, flags);
  68 }
  69
  70 /* Execution of a work */
  71 static void
  72 probe_workqueue_execution(struct task_struct *wq_thread,
  73                           struct work_struct *work)
  74 {
  75         int cpu = cpumask_first(&wq_thread->cpus_allowed);
  76         struct cpu_workqueue_stats *node;
  77         unsigned long flags;
  78
  79         spin_lock_irqsave(&workqueue_cpu_stat(cpu)->lock, flags);
  80         list_for_each_entry(node, &workqueue_cpu_stat(cpu)->list, list) {
  81                 if (node->pid == wq_thread->pid) {
  82                         node->executed++;
  83                         goto found;
  84                 }
  85         }
  86         pr_debug("trace_workqueue: entry not found\n");
  87 found:
  88         spin_unlock_irqrestore(&workqueue_cpu_stat(cpu)->lock, flags);
  89 }
  90
  91 /* Creation of a cpu workqueue thread */
  92 static void probe_workqueue_creation(struct task_struct *wq_thread, int cpu)
  93 {
  94         struct cpu_workqueue_stats *cws;
  95         unsigned long flags;
  96
  97         WARN_ON(cpu < 0);
  98
  99         /* Workqueues are sometimes created in atomic context */
 100         cws = kzalloc(sizeof(struct cpu_workqueue_stats), GFP_ATOMIC);
 101         if (!cws) {
 102                 pr_warning("trace_workqueue: not enough memory\n");
 103                 return;
 104         }
 105         INIT_LIST_HEAD(&cws->list);
 106         kref_init(&cws->kref);
 107         cws->cpu = cpu;
 108         cws->pid = wq_thread->pid;
 109
 110         spin_lock_irqsave(&workqueue_cpu_stat(cpu)->lock, flags);
 111         list_add_tail(&cws->list, &workqueue_cpu_stat(cpu)->list);
 112         spin_unlock_irqrestore(&workqueue_cpu_stat(cpu)->lock, flags);
 113 }
 114
 115 /* Destruction of a cpu workqueue thread */
 116 static void probe_workqueue_destruction(struct task_struct *wq_thread)
 117 {
 118         /* Workqueue only execute on one cpu */
 119         int cpu = cpumask_first(&wq_thread->cpus_allowed);
 120         struct cpu_workqueue_stats *node, *next;
 121         unsigned long flags;
 122
 123         spin_lock_irqsave(&workqueue_cpu_stat(cpu)->lock, flags);
 124         list_for_each_entry_safe(node, next, &workqueue_cpu_stat(cpu)->list,
 125                                                         list) {
 126                 if (node->pid == wq_thread->pid) {
 127                         list_del(&node->list);
 128                         kref_put(&node->kref, cpu_workqueue_stat_free);
 129                         goto found;
 130                 }
 131         }
 132
 133         pr_debug("trace_workqueue: don't find workqueue to destroy\n");
 134 found:
 135         spin_unlock_irqrestore(&workqueue_cpu_stat(cpu)->lock, flags);
 136
 137 }
 138
 139 static struct cpu_workqueue_stats *workqueue_stat_start_cpu(int cpu)
 140 {
 141         unsigned long flags;
 142         struct cpu_workqueue_stats *ret = NULL;
 143
 144
 145         spin_lock_irqsave(&workqueue_cpu_stat(cpu)->lock, flags);
 146
 147         if (!list_empty(&workqueue_cpu_stat(cpu)->list)) {
 148                 ret = list_entry(workqueue_cpu_stat(cpu)->list.next,
 149                                  struct cpu_workqueue_stats, list);
 150                 kref_get(&ret->kref);
 151         }
 152
 153         spin_unlock_irqrestore(&workqueue_cpu_stat(cpu)->lock, flags);
 154
 155         return ret;
 156 }
 157
 158 static void *workqueue_stat_start(struct tracer_stat *trace)
 159 {
 160         int cpu;
 161         void *ret = NULL;
 162
 163         for_each_possible_cpu(cpu) {
 164                 ret = workqueue_stat_start_cpu(cpu);
 165                 if (ret)
 166                         return ret;
 167         }
 168         return NULL;
 169 }
 170
 171 static void *workqueue_stat_next(void *prev, int idx)
 172 {
 173         struct cpu_workqueue_stats *prev_cws = prev;
 174         struct cpu_workqueue_stats *ret;
 175         int cpu = prev_cws->cpu;
 176         unsigned long flags;
 177
 178         spin_lock_irqsave(&workqueue_cpu_stat(cpu)->lock, flags);
 179         if (list_is_last(&prev_cws->list, &workqueue_cpu_stat(cpu)->list)) {
 180                 spin_unlock_irqrestore(&workqueue_cpu_stat(cpu)->lock, flags);
 181                 do {
 182                         cpu = cpumask_next(cpu, cpu_possible_mask);
 183                         if (cpu >= nr_cpu_ids)
 184                                 return NULL;
 185                 } while (!(ret = workqueue_stat_start_cpu(cpu)));
 186                 return ret;
 187         } else {
 188                 ret = list_entry(prev_cws->list.next,
 189                                  struct cpu_workqueue_stats, list);
 190                 kref_get(&ret->kref);
 191         }
 192         spin_unlock_irqrestore(&workqueue_cpu_stat(cpu)->lock, flags);
 193
 194         return ret;
 195 }
 196
 197 static int workqueue_stat_show(struct seq_file *s, void *p)
 198 {
 199         struct cpu_workqueue_stats *cws = p;
 200         struct pid *pid;
 201         struct task_struct *tsk;
 202
 203         pid = find_get_pid(cws->pid);
 204         if (pid) {
 205                 tsk = get_pid_task(pid, PIDTYPE_PID);
 206                 if (tsk) {
 207                         seq_printf(s, "%3d %6d     %6u       %s\n", cws->cpu,
 208                                    atomic_read(&cws->inserted), cws->executed,
 209                                    tsk->comm);
 210                         put_task_struct(tsk);
 211                 }
 212                 put_pid(pid);
 213         }
 214
 215         return 0;
 216 }
 217
 218 static void workqueue_stat_release(void *stat)
 219 {
 220         struct cpu_workqueue_stats *node = stat;
 221
 222         kref_put(&node->kref, cpu_workqueue_stat_free);
 223 }
 224
 225 static int workqueue_stat_headers(struct seq_file *s)
 226 {
 227         seq_printf(s, "# CPU  INSERTED  EXECUTED   NAME\n");
 228         seq_printf(s, "# |      |         |          |\n");
 229         return 0;
 230 }
 231
 232 struct tracer_stat workqueue_stats __read_mostly = {
 233         .name = "workqueues",
 234         .stat_start = workqueue_stat_start,
 235         .stat_next = workqueue_stat_next,
 236         .stat_show = workqueue_stat_show,
 237         .stat_release = workqueue_stat_release,
 238         .stat_headers = workqueue_stat_headers
 239 };
 240
 241
 242 int __init stat_workqueue_init(void)
 243 {
 244         if (register_stat_tracer(&workqueue_stats)) {
 245                 pr_warning("Unable to register workqueue stat tracer\n");
 246                 return 1;
 247         }
 248
 249         return 0;
 250 }
 251 fs_initcall(stat_workqueue_init);
 252
 253 /*
 254  * Workqueues are created very early, just after pre-smp initcalls.
 255  * So we must register our tracepoints at this stage.
 256  */
 257 int __init trace_workqueue_early_init(void)
 258 {
 259         int ret, cpu;
 260
 261         ret = register_trace_workqueue_insertion(probe_workqueue_insertion);
 262         if (ret)
 263                 goto out;
 264
 265         ret = register_trace_workqueue_execution(probe_workqueue_execution);
 266         if (ret)
 267                 goto no_insertion;
 268
 269         ret = register_trace_workqueue_creation(probe_workqueue_creation);
 270         if (ret)
 271                 goto no_execution;
 272
 273         ret = register_trace_workqueue_destruction(probe_workqueue_destruction);
 274         if (ret)
 275                 goto no_creation;
 276
 277         for_each_possible_cpu(cpu) {
 278                 spin_lock_init(&workqueue_cpu_stat(cpu)->lock);
 279                 INIT_LIST_HEAD(&workqueue_cpu_stat(cpu)->list);
 280         }
 281
 282         return 0;
 283
 284 no_creation:
 285         unregister_trace_workqueue_creation(probe_workqueue_creation);
 286 no_execution:
 287         unregister_trace_workqueue_execution(probe_workqueue_execution);
 288 no_insertion:
 289         unregister_trace_workqueue_insertion(probe_workqueue_insertion);
 290 out:
 291         pr_warning("trace_workqueue: unable to trace workqueues\n");
 292
 293         return 1;
 294 }
 295 early_initcall(trace_workqueue_early_init);