tracing: add function profiler
authorSteven Rostedt <srostedt@redhat.com>
Fri, 20 Mar 2009 16:50:56 +0000 (12:50 -0400)
committerSteven Rostedt <srostedt@redhat.com>
Wed, 25 Mar 2009 03:40:00 +0000 (23:40 -0400)
Impact: new profiling feature

This patch adds a function profiler. In debugfs/tracing/ two new
files are created.

  function_profile_enabled  - to enable or disable profiling

  trace_stat/functions   - the profiled functions.

For example:

  echo 1 > /debugfs/tracing/function_profile_enabled
  ./hackbench 50
  echo 0 > /debugfs/tracing/function_profile_enabled

yields:

  cat /debugfs/tracing/trace_stat/functions

  Function                               Hit
  --------                               ---
  _spin_lock                        10106442
  _spin_unlock                      10097492
  kfree                              6013704
  _spin_unlock_irqrestore            4423941
  _spin_lock_irqsave                 4406825
  __phys_addr                        4181686
  __slab_free                        4038222
  dput                               4030130
  path_put                           4023387
  unroll_tree_refs                   4019532
[...]

The most hit functions are listed first. Functions that are not
hit are not listed.

This feature depends on and uses dynamic function tracing. When the
function profiling is disabled, no overhead occurs. But it still
takes up around 300KB to hold the data, thus it is not recomended
to keep it enabled for systems low on memory.

When a '1' is echoed into the function_profile_enabled file, the
counters for is function is reset back to zero. Thus you can see what
functions are hit most by different programs.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>
include/linux/ftrace.h
kernel/trace/Kconfig
kernel/trace/ftrace.c

index 015a3d22cf7434deb500233623ae14b778d05e5b..0456c3a51c6697c7eb8f49bbdf0f9faa74c78fb4 100644 (file)
@@ -153,6 +153,10 @@ struct dyn_ftrace {
                unsigned long           flags;
                struct dyn_ftrace       *newlist;
        };
+#ifdef CONFIG_FUNCTION_PROFILER
+       unsigned long                   counter;
+       struct hlist_node               node;
+#endif
        struct dyn_arch_ftrace          arch;
 };
 
index 8a4d72931042a5092186f9e1c7b6da782f43e706..95e9ad5735d9c29aa62f5694c2927335198f275e 100644 (file)
@@ -105,6 +105,7 @@ config FUNCTION_GRAPH_TRACER
          This is done by setting the current return address on the current
          task structure into a stack of calls.
 
+
 config IRQSOFF_TRACER
        bool "Interrupts-off Latency Tracer"
        default n
@@ -376,6 +377,24 @@ config DYNAMIC_FTRACE
         were made. If so, it runs stop_machine (stops all CPUS)
         and modifies the code to jump over the call to ftrace.
 
+config FUNCTION_PROFILER
+       bool "Kernel function profiler"
+       depends on DYNAMIC_FTRACE
+       default n
+       help
+        This option enables the kernel function profiler. When the dynamic
+        function tracing is enabled, a counter is added into the function
+        records used by the dynamic function tracer. A file is created in
+        debugfs called function_profile_enabled which defaults to zero.
+        When a 1 is echoed into this file profiling begins, and when a
+        zero is entered, profiling stops. A file in the trace_stats
+        directory called functions, that show the list of functions that
+        have been hit and their counters.
+
+        This takes up around 320K more memory.
+
+        If in doubt, say N
+
 config FTRACE_MCOUNT_RECORD
        def_bool y
        depends on DYNAMIC_FTRACE
index 7b8722baf1531b3f82df2dde923d343332bd7d6e..11f364c776d549676f273e75aeefab2360109def 100644 (file)
@@ -34,6 +34,7 @@
 #include <asm/ftrace.h>
 
 #include "trace.h"
+#include "trace_stat.h"
 
 #define FTRACE_WARN_ON(cond)                   \
        do {                                    \
@@ -261,7 +262,6 @@ struct ftrace_func_probe {
        struct rcu_head         rcu;
 };
 
-
 enum {
        FTRACE_ENABLE_CALLS             = (1 << 0),
        FTRACE_DISABLE_CALLS            = (1 << 1),
@@ -309,6 +309,307 @@ static struct dyn_ftrace *ftrace_free_records;
                }                               \
        }
 
+#ifdef CONFIG_FUNCTION_PROFILER
+static struct hlist_head *ftrace_profile_hash;
+static int ftrace_profile_bits;
+static int ftrace_profile_enabled;
+static DEFINE_MUTEX(ftrace_profile_lock);
+
+static void *
+function_stat_next(void *v, int idx)
+{
+       struct dyn_ftrace *rec = v;
+       struct ftrace_page *pg;
+
+       pg = (struct ftrace_page *)((unsigned long)rec & PAGE_MASK);
+
+ again:
+       rec++;
+       if ((void *)rec >= (void *)&pg->records[pg->index]) {
+               pg = pg->next;
+               if (!pg)
+                       return NULL;
+               rec = &pg->records[0];
+       }
+
+       if (rec->flags & FTRACE_FL_FREE ||
+           rec->flags & FTRACE_FL_FAILED ||
+           !(rec->flags & FTRACE_FL_CONVERTED) ||
+           /* ignore non hit functions */
+           !rec->counter)
+               goto again;
+
+       return rec;
+}
+
+static void *function_stat_start(struct tracer_stat *trace)
+{
+       return function_stat_next(&ftrace_pages_start->records[0], 0);
+}
+
+static int function_stat_cmp(void *p1, void *p2)
+{
+       struct dyn_ftrace *a = p1;
+       struct dyn_ftrace *b = p2;
+
+       if (a->counter < b->counter)
+               return -1;
+       if (a->counter > b->counter)
+               return 1;
+       else
+               return 0;
+}
+
+static int function_stat_headers(struct seq_file *m)
+{
+       seq_printf(m, "  Function                               Hit\n"
+                     "  --------                               ---\n");
+       return 0;
+}
+
+static int function_stat_show(struct seq_file *m, void *v)
+{
+       struct dyn_ftrace *rec = v;
+       char str[KSYM_SYMBOL_LEN];
+
+       kallsyms_lookup(rec->ip, NULL, NULL, NULL, str);
+
+       seq_printf(m, "  %-30.30s  %10lu\n", str, rec->counter);
+       return 0;
+}
+
+static struct tracer_stat function_stats = {
+       .name = "functions",
+       .stat_start = function_stat_start,
+       .stat_next = function_stat_next,
+       .stat_cmp = function_stat_cmp,
+       .stat_headers = function_stat_headers,
+       .stat_show = function_stat_show
+};
+
+static void ftrace_profile_init(int nr_funcs)
+{
+       unsigned long addr;
+       int order;
+       int size;
+
+       /*
+        * We are profiling all functions, lets make it 1/4th of the
+        * number of functions that are in core kernel. So we have to
+        * iterate 4 times.
+        */
+       order = (sizeof(struct hlist_head) * nr_funcs) / 4;
+       order = get_order(order);
+       size = 1 << (PAGE_SHIFT + order);
+
+       pr_info("Allocating %d KB for profiler hash\n", size >> 10);
+
+       addr = __get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
+       if (!addr) {
+               pr_warning("Could not allocate function profiler hash\n");
+               return;
+       }
+
+       ftrace_profile_hash = (void *)addr;
+
+       /*
+        * struct hlist_head should be a pointer of 4 or 8 bytes.
+        * And a simple bit manipulation can be done, but if for
+        * some reason struct hlist_head is not a mulitple of 2,
+        * then we play it safe, and simply count. This function
+        * is done once at boot up, so it is not that critical in
+        * performance.
+        */
+
+       size--;
+       size /= sizeof(struct hlist_head);
+
+       for (; size; size >>= 1)
+               ftrace_profile_bits++;
+
+       pr_info("Function profiler has %d hash buckets\n",
+               1 << ftrace_profile_bits);
+
+       return;
+}
+
+static ssize_t
+ftrace_profile_read(struct file *filp, char __user *ubuf,
+                    size_t cnt, loff_t *ppos)
+{
+       char buf[64];
+       int r;
+
+       r = sprintf(buf, "%u\n", ftrace_profile_enabled);
+       return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
+}
+
+static void ftrace_profile_reset(void)
+{
+       struct dyn_ftrace *rec;
+       struct ftrace_page *pg;
+
+       do_for_each_ftrace_rec(pg, rec) {
+               rec->counter = 0;
+       } while_for_each_ftrace_rec();
+}
+
+static struct dyn_ftrace *ftrace_find_profiled_func(unsigned long ip)
+{
+       struct dyn_ftrace *rec;
+       struct hlist_head *hhd;
+       struct hlist_node *n;
+       unsigned long flags;
+       unsigned long key;
+
+       if (!ftrace_profile_hash)
+               return NULL;
+
+       key = hash_long(ip, ftrace_profile_bits);
+       hhd = &ftrace_profile_hash[key];
+
+       if (hlist_empty(hhd))
+               return NULL;
+
+       local_irq_save(flags);
+       hlist_for_each_entry_rcu(rec, n, hhd, node) {
+               if (rec->ip == ip)
+                       goto out;
+       }
+       rec = NULL;
+ out:
+       local_irq_restore(flags);
+
+       return rec;
+}
+
+static void
+function_profile_call(unsigned long ip, unsigned long parent_ip)
+{
+       struct dyn_ftrace *rec;
+       unsigned long flags;
+
+       if (!ftrace_profile_enabled)
+               return;
+
+       local_irq_save(flags);
+       rec = ftrace_find_profiled_func(ip);
+       if (!rec)
+               goto out;
+
+       rec->counter++;
+ out:
+       local_irq_restore(flags);
+}
+
+static struct ftrace_ops ftrace_profile_ops __read_mostly =
+{
+       .func = function_profile_call,
+};
+
+static ssize_t
+ftrace_profile_write(struct file *filp, const char __user *ubuf,
+                    size_t cnt, loff_t *ppos)
+{
+       unsigned long val;
+       char buf[64];
+       int ret;
+
+       if (!ftrace_profile_hash) {
+               pr_info("Can not enable hash due to earlier problems\n");
+               return -ENODEV;
+       }
+
+       if (cnt >= sizeof(buf))
+               return -EINVAL;
+
+       if (copy_from_user(&buf, ubuf, cnt))
+               return -EFAULT;
+
+       buf[cnt] = 0;
+
+       ret = strict_strtoul(buf, 10, &val);
+       if (ret < 0)
+               return ret;
+
+       val = !!val;
+
+       mutex_lock(&ftrace_profile_lock);
+       if (ftrace_profile_enabled ^ val) {
+               if (val) {
+                       ftrace_profile_reset();
+                       register_ftrace_function(&ftrace_profile_ops);
+                       ftrace_profile_enabled = 1;
+               } else {
+                       ftrace_profile_enabled = 0;
+                       unregister_ftrace_function(&ftrace_profile_ops);
+               }
+       }
+       mutex_unlock(&ftrace_profile_lock);
+
+       filp->f_pos += cnt;
+
+       return cnt;
+}
+
+static const struct file_operations ftrace_profile_fops = {
+       .open           = tracing_open_generic,
+       .read           = ftrace_profile_read,
+       .write          = ftrace_profile_write,
+};
+
+static void ftrace_profile_debugfs(struct dentry *d_tracer)
+{
+       struct dentry *entry;
+       int ret;
+
+       ret = register_stat_tracer(&function_stats);
+       if (ret) {
+               pr_warning("Warning: could not register "
+                          "function stats\n");
+               return;
+       }
+
+       entry = debugfs_create_file("function_profile_enabled", 0644,
+                                   d_tracer, NULL, &ftrace_profile_fops);
+       if (!entry)
+               pr_warning("Could not create debugfs "
+                          "'function_profile_enabled' entry\n");
+}
+
+static void ftrace_add_profile(struct dyn_ftrace *rec)
+{
+       unsigned long key;
+
+       if (!ftrace_profile_hash)
+               return;
+
+       key = hash_long(rec->ip, ftrace_profile_bits);
+       hlist_add_head_rcu(&rec->node, &ftrace_profile_hash[key]);
+}
+
+static void ftrace_profile_release(struct dyn_ftrace *rec)
+{
+       mutex_lock(&ftrace_profile_lock);
+       hlist_del(&rec->node);
+       mutex_unlock(&ftrace_profile_lock);
+}
+
+#else /* CONFIG_FUNCTION_PROFILER */
+static void ftrace_profile_init(int nr_funcs)
+{
+}
+static void ftrace_add_profile(struct dyn_ftrace *rec)
+{
+}
+static void ftrace_profile_debugfs(struct dentry *d_tracer)
+{
+}
+static void ftrace_profile_release(struct dyn_ftrace *rec)
+{
+}
+#endif /* CONFIG_FUNCTION_PROFILER */
+
 #ifdef CONFIG_KPROBES
 
 static int frozen_record_count;
@@ -359,8 +660,10 @@ void ftrace_release(void *start, unsigned long size)
        mutex_lock(&ftrace_lock);
        do_for_each_ftrace_rec(pg, rec) {
                if ((rec->ip >= s) && (rec->ip < e) &&
-                   !(rec->flags & FTRACE_FL_FREE))
+                   !(rec->flags & FTRACE_FL_FREE)) {
                        ftrace_free_rec(rec);
+                       ftrace_profile_release(rec);
+               }
        } while_for_each_ftrace_rec();
        mutex_unlock(&ftrace_lock);
 }
@@ -414,6 +717,8 @@ ftrace_record_ip(unsigned long ip)
        rec->newlist = ftrace_new_addrs;
        ftrace_new_addrs = rec;
 
+       ftrace_add_profile(rec);
+
        return rec;
 }
 
@@ -2157,6 +2462,8 @@ static __init int ftrace_init_dyn_debugfs(struct dentry *d_tracer)
                           "'set_graph_function' entry\n");
 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
 
+       ftrace_profile_debugfs(d_tracer);
+
        return 0;
 }
 
@@ -2225,6 +2532,8 @@ void __init ftrace_init(void)
        if (ret)
                goto failed;
 
+       ftrace_profile_init(count);
+
        last_ftrace_enabled = ftrace_enabled = 1;
 
        ret = ftrace_convert_nops(NULL,