tracing: Kernel Tracepoints
Mathieu Desnoyers [Fri, 18 Jul 2008 16:16:16 +0000 (12:16 -0400)]
Implementation of kernel tracepoints. Inspired from the Linux Kernel
Markers. Allows complete typing verification by declaring both tracing
statement inline functions and probe registration/unregistration static
inline functions within the same macro "DEFINE_TRACE". No format string
is required. See the tracepoint Documentation and Samples patches for
usage examples.

Taken from the documentation patch :

"A tracepoint placed in code provides a hook to call a function (probe)
that you can provide at runtime. A tracepoint can be "on" (a probe is
connected to it) or "off" (no probe is attached). When a tracepoint is
"off" it has no effect, except for adding a tiny time penalty (checking
a condition for a branch) and space penalty (adding a few bytes for the
function call at the end of the instrumented function and adds a data
structure in a separate section).  When a tracepoint is "on", the
function you provide is called each time the tracepoint is executed, in
the execution context of the caller. When the function provided ends its
execution, it returns to the caller (continuing from the tracepoint
site).

You can put tracepoints at important locations in the code. They are
lightweight hooks that can pass an arbitrary number of parameters, which
prototypes are described in a tracepoint declaration placed in a header
file."

Addition and removal of tracepoints is synchronized by RCU using the
scheduler (and preempt_disable) as guarantees to find a quiescent state
(this is really RCU "classic"). The update side uses rcu_barrier_sched()
with call_rcu_sched() and the read/execute side uses
"preempt_disable()/preempt_enable()".

We make sure the previous array containing probes, which has been
scheduled for deletion by the rcu callback, is indeed freed before we
proceed to the next update. It therefore limits the rate of modification
of a single tracepoint to one update per RCU period. The objective here
is to permit fast batch add/removal of probes on _different_
tracepoints.

Changelog :
- Use #name ":" #proto as string to identify the tracepoint in the
  tracepoint table. This will make sure not type mismatch happens due to
  connexion of a probe with the wrong type to a tracepoint declared with
  the same name in a different header.
- Add tracepoint_entry_free_old.
- Change __TO_TRACE to get rid of the 'i' iterator.

Masami Hiramatsu <mhiramat@redhat.com> :
Tested on x86-64.

Performance impact of a tracepoint : same as markers, except that it
adds about 70 bytes of instructions in an unlikely branch of each
instrumented function (the for loop, the stack setup and the function
call). It currently adds a memory read, a test and a conditional branch
at the instrumentation site (in the hot path). Immediate values will
eventually change this into a load immediate, test and branch, which
removes the memory read which will make the i-cache impact smaller
(changing the memory read for a load immediate removes 3-4 bytes per
site on x86_32 (depending on mov prefixes), or 7-8 bytes on x86_64, it
also saves the d-cache hit).

About the performance impact of tracepoints (which is comparable to
markers), even without immediate values optimizations, tests done by
Hideo Aoki on ia64 show no regression. His test case was using hackbench
on a kernel where scheduler instrumentation (about 5 events in code
scheduler code) was added.

Quoting Hideo Aoki about Markers :

I evaluated overhead of kernel marker using linux-2.6-sched-fixes git
tree, which includes several markers for LTTng, using an ia64 server.

While the immediate trace mark feature isn't implemented on ia64, there
is no major performance regression. So, I think that we don't have any
issues to propose merging marker point patches into Linus's tree from
the viewpoint of performance impact.

I prepared two kernels to evaluate. The first one was compiled without
CONFIG_MARKERS. The second one was enabled CONFIG_MARKERS.

I downloaded the original hackbench from the following URL:
http://devresources.linux-foundation.org/craiger/hackbench/src/hackbench.c

I ran hackbench 5 times in each condition and calculated the average and
difference between the kernels.

    The parameter of hackbench: every 50 from 50 to 800
    The number of CPUs of the server: 2, 4, and 8

Below is the results. As you can see, major performance regression
wasn't found in any case. Even if number of processes increases,
differences between marker-enabled kernel and marker- disabled kernel
doesn't increase. Moreover, if number of CPUs increases, the differences
doesn't increase either.

Curiously, marker-enabled kernel is better than marker-disabled kernel
in more than half cases, although I guess it comes from the difference
of memory access pattern.

* 2 CPUs

Number of | without      | with         | diff     | diff    |
processes | Marker [Sec] | Marker [Sec] |   [Sec]  |   [%]   |
--------------------------------------------------------------
       50 |      4.811   |       4.872  |  +0.061  |  +1.27  |
      100 |      9.854   |      10.309  |  +0.454  |  +4.61  |
      150 |     15.602   |      15.040  |  -0.562  |  -3.6   |
      200 |     20.489   |      20.380  |  -0.109  |  -0.53  |
      250 |     25.798   |      25.652  |  -0.146  |  -0.56  |
      300 |     31.260   |      30.797  |  -0.463  |  -1.48  |
      350 |     36.121   |      35.770  |  -0.351  |  -0.97  |
      400 |     42.288   |      42.102  |  -0.186  |  -0.44  |
      450 |     47.778   |      47.253  |  -0.526  |  -1.1   |
      500 |     51.953   |      52.278  |  +0.325  |  +0.63  |
      550 |     58.401   |      57.700  |  -0.701  |  -1.2   |
      600 |     63.334   |      63.222  |  -0.112  |  -0.18  |
      650 |     68.816   |      68.511  |  -0.306  |  -0.44  |
      700 |     74.667   |      74.088  |  -0.579  |  -0.78  |
      750 |     78.612   |      79.582  |  +0.970  |  +1.23  |
      800 |     85.431   |      85.263  |  -0.168  |  -0.2   |
--------------------------------------------------------------

* 4 CPUs

Number of | without      | with         | diff     | diff    |
processes | Marker [Sec] | Marker [Sec] |   [Sec]  |   [%]   |
--------------------------------------------------------------
       50 |      2.586   |       2.584  |  -0.003  |  -0.1   |
      100 |      5.254   |       5.283  |  +0.030  |  +0.56  |
      150 |      8.012   |       8.074  |  +0.061  |  +0.76  |
      200 |     11.172   |      11.000  |  -0.172  |  -1.54  |
      250 |     13.917   |      14.036  |  +0.119  |  +0.86  |
      300 |     16.905   |      16.543  |  -0.362  |  -2.14  |
      350 |     19.901   |      20.036  |  +0.135  |  +0.68  |
      400 |     22.908   |      23.094  |  +0.186  |  +0.81  |
      450 |     26.273   |      26.101  |  -0.172  |  -0.66  |
      500 |     29.554   |      29.092  |  -0.461  |  -1.56  |
      550 |     32.377   |      32.274  |  -0.103  |  -0.32  |
      600 |     35.855   |      35.322  |  -0.533  |  -1.49  |
      650 |     39.192   |      38.388  |  -0.804  |  -2.05  |
      700 |     41.744   |      41.719  |  -0.025  |  -0.06  |
      750 |     45.016   |      44.496  |  -0.520  |  -1.16  |
      800 |     48.212   |      47.603  |  -0.609  |  -1.26  |
--------------------------------------------------------------

* 8 CPUs

Number of | without      | with         | diff     | diff    |
processes | Marker [Sec] | Marker [Sec] |   [Sec]  |   [%]   |
--------------------------------------------------------------
       50 |      2.094   |       2.072  |  -0.022  |  -1.07  |
      100 |      4.162   |       4.273  |  +0.111  |  +2.66  |
      150 |      6.485   |       6.540  |  +0.055  |  +0.84  |
      200 |      8.556   |       8.478  |  -0.078  |  -0.91  |
      250 |     10.458   |      10.258  |  -0.200  |  -1.91  |
      300 |     12.425   |      12.750  |  +0.325  |  +2.62  |
      350 |     14.807   |      14.839  |  +0.032  |  +0.22  |
      400 |     16.801   |      16.959  |  +0.158  |  +0.94  |
      450 |     19.478   |      19.009  |  -0.470  |  -2.41  |
      500 |     21.296   |      21.504  |  +0.208  |  +0.98  |
      550 |     23.842   |      23.979  |  +0.137  |  +0.57  |
      600 |     26.309   |      26.111  |  -0.198  |  -0.75  |
      650 |     28.705   |      28.446  |  -0.259  |  -0.9   |
      700 |     31.233   |      31.394  |  +0.161  |  +0.52  |
      750 |     34.064   |      33.720  |  -0.344  |  -1.01  |
      800 |     36.320   |      36.114  |  -0.206  |  -0.57  |
--------------------------------------------------------------

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
Acked-by: Masami Hiramatsu <mhiramat@redhat.com>
Acked-by: 'Peter Zijlstra' <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

include/asm-generic/vmlinux.lds.h
include/linux/module.h
include/linux/tracepoint.h [new file with mode: 0644]
init/Kconfig
kernel/Makefile
kernel/module.c
kernel/tracepoint.c [new file with mode: 0644]

index 7440a0d..3d8e472 100644 (file)
        . = ALIGN(8);                                                   \
        VMLINUX_SYMBOL(__start___markers) = .;                          \
        *(__markers)                                                    \
-       VMLINUX_SYMBOL(__stop___markers) = .;
+       VMLINUX_SYMBOL(__stop___markers) = .;                           \
+       VMLINUX_SYMBOL(__start___tracepoints) = .;                      \
+       *(__tracepoints)                                                \
+       VMLINUX_SYMBOL(__stop___tracepoints) = .;
 
 #define RO_DATA(align)                                                 \
        . = ALIGN((align));                                             \
@@ -61,6 +64,7 @@
                *(.rodata) *(.rodata.*)                                 \
                *(__vermagic)           /* Kernel version magic */      \
                *(__markers_strings)    /* Markers: strings */          \
+               *(__tracepoints_strings)/* Tracepoints: strings */      \
        }                                                               \
                                                                        \
        .rodata1          : AT(ADDR(.rodata1) - LOAD_OFFSET) {          \
index 68e0955..8b61135 100644 (file)
@@ -16,6 +16,7 @@
 #include <linux/kobject.h>
 #include <linux/moduleparam.h>
 #include <linux/marker.h>
+#include <linux/tracepoint.h>
 #include <asm/local.h>
 
 #include <asm/module.h>
@@ -331,6 +332,10 @@ struct module
        struct marker *markers;
        unsigned int num_markers;
 #endif
+#ifdef CONFIG_TRACEPOINTS
+       struct tracepoint *tracepoints;
+       unsigned int num_tracepoints;
+#endif
 
 #ifdef CONFIG_MODULE_UNLOAD
        /* What modules depend on me? */
@@ -454,6 +459,9 @@ extern void print_modules(void);
 
 extern void module_update_markers(void);
 
+extern void module_update_tracepoints(void);
+extern int module_get_iter_tracepoints(struct tracepoint_iter *iter);
+
 #else /* !CONFIG_MODULES... */
 #define EXPORT_SYMBOL(sym)
 #define EXPORT_SYMBOL_GPL(sym)
@@ -558,6 +566,15 @@ static inline void module_update_markers(void)
 {
 }
 
+static inline void module_update_tracepoints(void)
+{
+}
+
+static inline int module_get_iter_tracepoints(struct tracepoint_iter *iter)
+{
+       return 0;
+}
+
 #endif /* CONFIG_MODULES */
 
 struct device_driver;
diff --git a/include/linux/tracepoint.h b/include/linux/tracepoint.h
new file mode 100644 (file)
index 0000000..e623a6f
--- /dev/null
@@ -0,0 +1,127 @@
+#ifndef _LINUX_TRACEPOINT_H
+#define _LINUX_TRACEPOINT_H
+
+/*
+ * Kernel Tracepoint API.
+ *
+ * See Documentation/tracepoint.txt.
+ *
+ * (C) Copyright 2008 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ *
+ * Heavily inspired from the Linux Kernel Markers.
+ *
+ * This file is released under the GPLv2.
+ * See the file COPYING for more details.
+ */
+
+#include <linux/types.h>
+#include <linux/rcupdate.h>
+
+struct module;
+struct tracepoint;
+
+struct tracepoint {
+       const char *name;               /* Tracepoint name */
+       int state;                      /* State. */
+       void **funcs;
+} __attribute__((aligned(8)));
+
+
+#define TPPROTO(args...)       args
+#define TPARGS(args...)                args
+
+#ifdef CONFIG_TRACEPOINTS
+
+/*
+ * it_func[0] is never NULL because there is at least one element in the array
+ * when the array itself is non NULL.
+ */
+#define __DO_TRACE(tp, proto, args)                                    \
+       do {                                                            \
+               void **it_func;                                         \
+                                                                       \
+               rcu_read_lock_sched();                                  \
+               it_func = rcu_dereference((tp)->funcs);                 \
+               if (it_func) {                                          \
+                       do {                                            \
+                               ((void(*)(proto))(*it_func))(args);     \
+                       } while (*(++it_func));                         \
+               }                                                       \
+               rcu_read_unlock_sched();                                \
+       } while (0)
+
+/*
+ * Make sure the alignment of the structure in the __tracepoints section will
+ * not add unwanted padding between the beginning of the section and the
+ * structure. Force alignment to the same alignment as the section start.
+ */
+#define DEFINE_TRACE(name, proto, args)                                        \
+       static inline void trace_##name(proto)                          \
+       {                                                               \
+               static const char __tpstrtab_##name[]                   \
+               __attribute__((section("__tracepoints_strings")))       \
+               = #name ":" #proto;                                     \
+               static struct tracepoint __tracepoint_##name            \
+               __attribute__((section("__tracepoints"), aligned(8))) = \
+               { __tpstrtab_##name, 0, NULL };                         \
+               if (unlikely(__tracepoint_##name.state))                \
+                       __DO_TRACE(&__tracepoint_##name,                \
+                               TPPROTO(proto), TPARGS(args));          \
+       }                                                               \
+       static inline int register_trace_##name(void (*probe)(proto))   \
+       {                                                               \
+               return tracepoint_probe_register(#name ":" #proto,      \
+                       (void *)probe);                                 \
+       }                                                               \
+       static inline void unregister_trace_##name(void (*probe)(proto))\
+       {                                                               \
+               tracepoint_probe_unregister(#name ":" #proto,           \
+                       (void *)probe);                                 \
+       }
+
+extern void tracepoint_update_probe_range(struct tracepoint *begin,
+       struct tracepoint *end);
+
+#else /* !CONFIG_TRACEPOINTS */
+#define DEFINE_TRACE(name, proto, args)                        \
+       static inline void _do_trace_##name(struct tracepoint *tp, proto) \
+       { }                                                             \
+       static inline void trace_##name(proto)                          \
+       { }                                                             \
+       static inline int register_trace_##name(void (*probe)(proto))   \
+       {                                                               \
+               return -ENOSYS;                                         \
+       }                                                               \
+       static inline void unregister_trace_##name(void (*probe)(proto))\
+       { }
+
+static inline void tracepoint_update_probe_range(struct tracepoint *begin,
+       struct tracepoint *end)
+{ }
+#endif /* CONFIG_TRACEPOINTS */
+
+/*
+ * Connect a probe to a tracepoint.
+ * Internal API, should not be used directly.
+ */
+extern int tracepoint_probe_register(const char *name, void *probe);
+
+/*
+ * Disconnect a probe from a tracepoint.
+ * Internal API, should not be used directly.
+ */
+extern int tracepoint_probe_unregister(const char *name, void *probe);
+
+struct tracepoint_iter {
+       struct module *module;
+       struct tracepoint *tracepoint;
+};
+
+extern void tracepoint_iter_start(struct tracepoint_iter *iter);
+extern void tracepoint_iter_next(struct tracepoint_iter *iter);
+extern void tracepoint_iter_stop(struct tracepoint_iter *iter);
+extern void tracepoint_iter_reset(struct tracepoint_iter *iter);
+extern int tracepoint_get_iter_range(struct tracepoint **tracepoint,
+       struct tracepoint *begin, struct tracepoint *end);
+
+#endif
index c11da38..7008267 100644 (file)
@@ -771,6 +771,13 @@ config PROFILING
          Say Y here to enable the extended profiling support mechanisms used
          by profilers such as OProfile.
 
+config TRACEPOINTS
+       bool "Activate tracepoints"
+       default y
+       help
+         Place an empty function call at each tracepoint site. Can be
+         dynamically changed for a probe function.
+
 config MARKERS
        bool "Activate markers"
        help
index 4e1d7df..8f9ce7e 100644 (file)
@@ -83,6 +83,7 @@ obj-$(CONFIG_SYSCTL) += utsname_sysctl.o
 obj-$(CONFIG_TASK_DELAY_ACCT) += delayacct.o
 obj-$(CONFIG_TASKSTATS) += taskstats.o tsacct.o
 obj-$(CONFIG_MARKERS) += marker.o
+obj-$(CONFIG_TRACEPOINTS) += tracepoint.o
 obj-$(CONFIG_LATENCYTOP) += latencytop.o
 obj-$(CONFIG_HAVE_GENERIC_DMA_COHERENT) += dma-coherent.o
 obj-$(CONFIG_FTRACE) += trace/
index 9db1191..661d73d 100644 (file)
@@ -46,6 +46,7 @@
 #include <asm/cacheflush.h>
 #include <linux/license.h>
 #include <asm/sections.h>
+#include <linux/tracepoint.h>
 
 #if 0
 #define DEBUGP printk
@@ -1831,6 +1832,8 @@ static noinline struct module *load_module(void __user *umod,
 #endif
        unsigned int markersindex;
        unsigned int markersstringsindex;
+       unsigned int tracepointsindex;
+       unsigned int tracepointsstringsindex;
        struct module *mod;
        long err = 0;
        void *percpu = NULL, *ptr = NULL; /* Stops spurious gcc warning */
@@ -2117,6 +2120,9 @@ static noinline struct module *load_module(void __user *umod,
        markersindex = find_sec(hdr, sechdrs, secstrings, "__markers");
        markersstringsindex = find_sec(hdr, sechdrs, secstrings,
                                        "__markers_strings");
+       tracepointsindex = find_sec(hdr, sechdrs, secstrings, "__tracepoints");
+       tracepointsstringsindex = find_sec(hdr, sechdrs, secstrings,
+                                       "__tracepoints_strings");
 
        /* Now do relocations. */
        for (i = 1; i < hdr->e_shnum; i++) {
@@ -2144,6 +2150,12 @@ static noinline struct module *load_module(void __user *umod,
        mod->num_markers =
                sechdrs[markersindex].sh_size / sizeof(*mod->markers);
 #endif
+#ifdef CONFIG_TRACEPOINTS
+       mod->tracepoints = (void *)sechdrs[tracepointsindex].sh_addr;
+       mod->num_tracepoints =
+               sechdrs[tracepointsindex].sh_size / sizeof(*mod->tracepoints);
+#endif
+
 
         /* Find duplicate symbols */
        err = verify_export_symbols(mod);
@@ -2162,11 +2174,16 @@ static noinline struct module *load_module(void __user *umod,
 
        add_kallsyms(mod, sechdrs, symindex, strindex, secstrings);
 
+       if (!mod->taints) {
 #ifdef CONFIG_MARKERS
-       if (!mod->taints)
                marker_update_probe_range(mod->markers,
                        mod->markers + mod->num_markers);
 #endif
+#ifdef CONFIG_TRACEPOINTS
+               tracepoint_update_probe_range(mod->tracepoints,
+                       mod->tracepoints + mod->num_tracepoints);
+#endif
+       }
        err = module_finalize(hdr, sechdrs, mod);
        if (err < 0)
                goto cleanup;
@@ -2717,3 +2734,50 @@ void module_update_markers(void)
        mutex_unlock(&module_mutex);
 }
 #endif
+
+#ifdef CONFIG_TRACEPOINTS
+void module_update_tracepoints(void)
+{
+       struct module *mod;
+
+       mutex_lock(&module_mutex);
+       list_for_each_entry(mod, &modules, list)
+               if (!mod->taints)
+                       tracepoint_update_probe_range(mod->tracepoints,
+                               mod->tracepoints + mod->num_tracepoints);
+       mutex_unlock(&module_mutex);
+}
+
+/*
+ * Returns 0 if current not found.
+ * Returns 1 if current found.
+ */
+int module_get_iter_tracepoints(struct tracepoint_iter *iter)
+{
+       struct module *iter_mod;
+       int found = 0;
+
+       mutex_lock(&module_mutex);
+       list_for_each_entry(iter_mod, &modules, list) {
+               if (!iter_mod->taints) {
+                       /*
+                        * Sorted module list
+                        */
+                       if (iter_mod < iter->module)
+                               continue;
+                       else if (iter_mod > iter->module)
+                               iter->tracepoint = NULL;
+                       found = tracepoint_get_iter_range(&iter->tracepoint,
+                               iter_mod->tracepoints,
+                               iter_mod->tracepoints
+                                       + iter_mod->num_tracepoints);
+                       if (found) {
+                               iter->module = iter_mod;
+                               break;
+                       }
+               }
+       }
+       mutex_unlock(&module_mutex);
+       return found;
+}
+#endif
diff --git a/kernel/tracepoint.c b/kernel/tracepoint.c
new file mode 100644 (file)
index 0000000..42e86dd
--- /dev/null
@@ -0,0 +1,476 @@
+/*
+ * Copyright (C) 2008 Mathieu Desnoyers
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
+ */
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/types.h>
+#include <linux/jhash.h>
+#include <linux/list.h>
+#include <linux/rcupdate.h>
+#include <linux/tracepoint.h>
+#include <linux/err.h>
+#include <linux/slab.h>
+
+extern struct tracepoint __start___tracepoints[];
+extern struct tracepoint __stop___tracepoints[];
+
+/* Set to 1 to enable tracepoint debug output */
+static const int tracepoint_debug;
+
+/*
+ * tracepoints_mutex nests inside module_mutex. Tracepoints mutex protects the
+ * builtin and module tracepoints and the hash table.
+ */
+static DEFINE_MUTEX(tracepoints_mutex);
+
+/*
+ * Tracepoint hash table, containing the active tracepoints.
+ * Protected by tracepoints_mutex.
+ */
+#define TRACEPOINT_HASH_BITS 6
+#define TRACEPOINT_TABLE_SIZE (1 << TRACEPOINT_HASH_BITS)
+
+/*
+ * Note about RCU :
+ * It is used to to delay the free of multiple probes array until a quiescent
+ * state is reached.
+ * Tracepoint entries modifications are protected by the tracepoints_mutex.
+ */
+struct tracepoint_entry {
+       struct hlist_node hlist;
+       void **funcs;
+       int refcount;   /* Number of times armed. 0 if disarmed. */
+       struct rcu_head rcu;
+       void *oldptr;
+       unsigned char rcu_pending:1;
+       char name[0];
+};
+
+static struct hlist_head tracepoint_table[TRACEPOINT_TABLE_SIZE];
+
+static void free_old_closure(struct rcu_head *head)
+{
+       struct tracepoint_entry *entry = container_of(head,
+               struct tracepoint_entry, rcu);
+       kfree(entry->oldptr);
+       /* Make sure we free the data before setting the pending flag to 0 */
+       smp_wmb();
+       entry->rcu_pending = 0;
+}
+
+static void tracepoint_entry_free_old(struct tracepoint_entry *entry, void *old)
+{
+       if (!old)
+               return;
+       entry->oldptr = old;
+       entry->rcu_pending = 1;
+       /* write rcu_pending before calling the RCU callback */
+       smp_wmb();
+#ifdef CONFIG_PREEMPT_RCU
+       synchronize_sched();    /* Until we have the call_rcu_sched() */
+#endif
+       call_rcu(&entry->rcu, free_old_closure);
+}
+
+static void debug_print_probes(struct tracepoint_entry *entry)
+{
+       int i;
+
+       if (!tracepoint_debug)
+               return;
+
+       for (i = 0; entry->funcs[i]; i++)
+               printk(KERN_DEBUG "Probe %d : %p\n", i, entry->funcs[i]);
+}
+
+static void *
+tracepoint_entry_add_probe(struct tracepoint_entry *entry, void *probe)
+{
+       int nr_probes = 0;
+       void **old, **new;
+
+       WARN_ON(!probe);
+
+       debug_print_probes(entry);
+       old = entry->funcs;
+       if (old) {
+               /* (N -> N+1), (N != 0, 1) probes */
+               for (nr_probes = 0; old[nr_probes]; nr_probes++)
+                       if (old[nr_probes] == probe)
+                               return ERR_PTR(-EEXIST);
+       }
+       /* + 2 : one for new probe, one for NULL func */
+       new = kzalloc((nr_probes + 2) * sizeof(void *), GFP_KERNEL);
+       if (new == NULL)
+               return ERR_PTR(-ENOMEM);
+       if (old)
+               memcpy(new, old, nr_probes * sizeof(void *));
+       new[nr_probes] = probe;
+       entry->refcount = nr_probes + 1;
+       entry->funcs = new;
+       debug_print_probes(entry);
+       return old;
+}
+
+static void *
+tracepoint_entry_remove_probe(struct tracepoint_entry *entry, void *probe)
+{
+       int nr_probes = 0, nr_del = 0, i;
+       void **old, **new;
+
+       old = entry->funcs;
+
+       debug_print_probes(entry);
+       /* (N -> M), (N > 1, M >= 0) probes */
+       for (nr_probes = 0; old[nr_probes]; nr_probes++) {
+               if ((!probe || old[nr_probes] == probe))
+                       nr_del++;
+       }
+
+       if (nr_probes - nr_del == 0) {
+               /* N -> 0, (N > 1) */
+               entry->funcs = NULL;
+               entry->refcount = 0;
+               debug_print_probes(entry);
+               return old;
+       } else {
+               int j = 0;
+               /* N -> M, (N > 1, M > 0) */
+               /* + 1 for NULL */
+               new = kzalloc((nr_probes - nr_del + 1)
+                       * sizeof(void *), GFP_KERNEL);
+               if (new == NULL)
+                       return ERR_PTR(-ENOMEM);
+               for (i = 0; old[i]; i++)
+                       if ((probe && old[i] != probe))
+                               new[j++] = old[i];
+               entry->refcount = nr_probes - nr_del;
+               entry->funcs = new;
+       }
+       debug_print_probes(entry);
+       return old;
+}
+
+/*
+ * Get tracepoint if the tracepoint is present in the tracepoint hash table.
+ * Must be called with tracepoints_mutex held.
+ * Returns NULL if not present.
+ */
+static struct tracepoint_entry *get_tracepoint(const char *name)
+{
+       struct hlist_head *head;
+       struct hlist_node *node;
+       struct tracepoint_entry *e;
+       u32 hash = jhash(name, strlen(name), 0);
+
+       head = &tracepoint_table[hash & ((1 << TRACEPOINT_HASH_BITS)-1)];
+       hlist_for_each_entry(e, node, head, hlist) {
+               if (!strcmp(name, e->name))
+                       return e;
+       }
+       return NULL;
+}
+
+/*
+ * Add the tracepoint to the tracepoint hash table. Must be called with
+ * tracepoints_mutex held.
+ */
+static struct tracepoint_entry *add_tracepoint(const char *name)
+{
+       struct hlist_head *head;
+       struct hlist_node *node;
+       struct tracepoint_entry *e;
+       size_t name_len = strlen(name) + 1;
+       u32 hash = jhash(name, name_len-1, 0);
+
+       head = &tracepoint_table[hash & ((1 << TRACEPOINT_HASH_BITS)-1)];
+       hlist_for_each_entry(e, node, head, hlist) {
+               if (!strcmp(name, e->name)) {
+                       printk(KERN_NOTICE
+                               "tracepoint %s busy\n", name);
+                       return ERR_PTR(-EEXIST);        /* Already there */
+               }
+       }
+       /*
+        * Using kmalloc here to allocate a variable length element. Could
+        * cause some memory fragmentation if overused.
+        */
+       e = kmalloc(sizeof(struct tracepoint_entry) + name_len, GFP_KERNEL);
+       if (!e)
+               return ERR_PTR(-ENOMEM);
+       memcpy(&e->name[0], name, name_len);
+       e->funcs = NULL;
+       e->refcount = 0;
+       e->rcu_pending = 0;
+       hlist_add_head(&e->hlist, head);
+       return e;
+}
+
+/*
+ * Remove the tracepoint from the tracepoint hash table. Must be called with
+ * mutex_lock held.
+ */
+static int remove_tracepoint(const char *name)
+{
+       struct hlist_head *head;
+       struct hlist_node *node;
+       struct tracepoint_entry *e;
+       int found = 0;
+       size_t len = strlen(name) + 1;
+       u32 hash = jhash(name, len-1, 0);
+
+       head = &tracepoint_table[hash & ((1 << TRACEPOINT_HASH_BITS)-1)];
+       hlist_for_each_entry(e, node, head, hlist) {
+               if (!strcmp(name, e->name)) {
+                       found = 1;
+                       break;
+               }
+       }
+       if (!found)
+               return -ENOENT;
+       if (e->refcount)
+               return -EBUSY;
+       hlist_del(&e->hlist);
+       /* Make sure the call_rcu has been executed */
+       if (e->rcu_pending)
+               rcu_barrier();
+       kfree(e);
+       return 0;
+}
+
+/*
+ * Sets the probe callback corresponding to one tracepoint.
+ */
+static void set_tracepoint(struct tracepoint_entry **entry,
+       struct tracepoint *elem, int active)
+{
+       WARN_ON(strcmp((*entry)->name, elem->name) != 0);
+
+       /*
+        * rcu_assign_pointer has a smp_wmb() which makes sure that the new
+        * probe callbacks array is consistent before setting a pointer to it.
+        * This array is referenced by __DO_TRACE from
+        * include/linux/tracepoints.h. A matching smp_read_barrier_depends()
+        * is used.
+        */
+       rcu_assign_pointer(elem->funcs, (*entry)->funcs);
+       elem->state = active;
+}
+
+/*
+ * Disable a tracepoint and its probe callback.
+ * Note: only waiting an RCU period after setting elem->call to the empty
+ * function insures that the original callback is not used anymore. This insured
+ * by preempt_disable around the call site.
+ */
+static void disable_tracepoint(struct tracepoint *elem)
+{
+       elem->state = 0;
+}
+
+/**
+ * tracepoint_update_probe_range - Update a probe range
+ * @begin: beginning of the range
+ * @end: end of the range
+ *
+ * Updates the probe callback corresponding to a range of tracepoints.
+ */
+void tracepoint_update_probe_range(struct tracepoint *begin,
+       struct tracepoint *end)
+{
+       struct tracepoint *iter;
+       struct tracepoint_entry *mark_entry;
+
+       mutex_lock(&tracepoints_mutex);
+       for (iter = begin; iter < end; iter++) {
+               mark_entry = get_tracepoint(iter->name);
+               if (mark_entry) {
+                       set_tracepoint(&mark_entry, iter,
+                                       !!mark_entry->refcount);
+               } else {
+                       disable_tracepoint(iter);
+               }
+       }
+       mutex_unlock(&tracepoints_mutex);
+}
+
+/*
+ * Update probes, removing the faulty probes.
+ */
+static void tracepoint_update_probes(void)
+{
+       /* Core kernel tracepoints */
+       tracepoint_update_probe_range(__start___tracepoints,
+               __stop___tracepoints);
+       /* tracepoints in modules. */
+       module_update_tracepoints();
+}
+
+/**
+ * tracepoint_probe_register -  Connect a probe to a tracepoint
+ * @name: tracepoint name
+ * @probe: probe handler
+ *
+ * Returns 0 if ok, error value on error.
+ * The probe address must at least be aligned on the architecture pointer size.
+ */
+int tracepoint_probe_register(const char *name, void *probe)
+{
+       struct tracepoint_entry *entry;
+       int ret = 0;
+       void *old;
+
+       mutex_lock(&tracepoints_mutex);
+       entry = get_tracepoint(name);
+       if (!entry) {
+               entry = add_tracepoint(name);
+               if (IS_ERR(entry)) {
+                       ret = PTR_ERR(entry);
+                       goto end;
+               }
+       }
+       /*
+        * If we detect that a call_rcu is pending for this tracepoint,
+        * make sure it's executed now.
+        */
+       if (entry->rcu_pending)
+               rcu_barrier();
+       old = tracepoint_entry_add_probe(entry, probe);
+       if (IS_ERR(old)) {
+               ret = PTR_ERR(old);
+               goto end;
+       }
+       mutex_unlock(&tracepoints_mutex);
+       tracepoint_update_probes();             /* may update entry */
+       mutex_lock(&tracepoints_mutex);
+       entry = get_tracepoint(name);
+       WARN_ON(!entry);
+       tracepoint_entry_free_old(entry, old);
+end:
+       mutex_unlock(&tracepoints_mutex);
+       return ret;
+}
+EXPORT_SYMBOL_GPL(tracepoint_probe_register);
+
+/**
+ * tracepoint_probe_unregister -  Disconnect a probe from a tracepoint
+ * @name: tracepoint name
+ * @probe: probe function pointer
+ *
+ * We do not need to call a synchronize_sched to make sure the probes have
+ * finished running before doing a module unload, because the module unload
+ * itself uses stop_machine(), which insures that every preempt disabled section
+ * have finished.
+ */
+int tracepoint_probe_unregister(const char *name, void *probe)
+{
+       struct tracepoint_entry *entry;
+       void *old;
+       int ret = -ENOENT;
+
+       mutex_lock(&tracepoints_mutex);
+       entry = get_tracepoint(name);
+       if (!entry)
+               goto end;
+       if (entry->rcu_pending)
+               rcu_barrier();
+       old = tracepoint_entry_remove_probe(entry, probe);
+       mutex_unlock(&tracepoints_mutex);
+       tracepoint_update_probes();             /* may update entry */
+       mutex_lock(&tracepoints_mutex);
+       entry = get_tracepoint(name);
+       if (!entry)
+               goto end;
+       tracepoint_entry_free_old(entry, old);
+       remove_tracepoint(name);        /* Ignore busy error message */
+       ret = 0;
+end:
+       mutex_unlock(&tracepoints_mutex);
+       return ret;
+}
+EXPORT_SYMBOL_GPL(tracepoint_probe_unregister);
+
+/**
+ * tracepoint_get_iter_range - Get a next tracepoint iterator given a range.
+ * @tracepoint: current tracepoints (in), next tracepoint (out)
+ * @begin: beginning of the range
+ * @end: end of the range
+ *
+ * Returns whether a next tracepoint has been found (1) or not (0).
+ * Will return the first tracepoint in the range if the input tracepoint is
+ * NULL.
+ */
+int tracepoint_get_iter_range(struct tracepoint **tracepoint,
+       struct tracepoint *begin, struct tracepoint *end)
+{
+       if (!*tracepoint && begin != end) {
+               *tracepoint = begin;
+               return 1;
+       }
+       if (*tracepoint >= begin && *tracepoint < end)
+               return 1;
+       return 0;
+}
+EXPORT_SYMBOL_GPL(tracepoint_get_iter_range);
+
+static void tracepoint_get_iter(struct tracepoint_iter *iter)
+{
+       int found = 0;
+
+       /* Core kernel tracepoints */
+       if (!iter->module) {
+               found = tracepoint_get_iter_range(&iter->tracepoint,
+                               __start___tracepoints, __stop___tracepoints);
+               if (found)
+                       goto end;
+       }
+       /* tracepoints in modules. */
+       found = module_get_iter_tracepoints(iter);
+end:
+       if (!found)
+               tracepoint_iter_reset(iter);
+}
+
+void tracepoint_iter_start(struct tracepoint_iter *iter)
+{
+       tracepoint_get_iter(iter);
+}
+EXPORT_SYMBOL_GPL(tracepoint_iter_start);
+
+void tracepoint_iter_next(struct tracepoint_iter *iter)
+{
+       iter->tracepoint++;
+       /*
+        * iter->tracepoint may be invalid because we blindly incremented it.
+        * Make sure it is valid by marshalling on the tracepoints, getting the
+        * tracepoints from following modules if necessary.
+        */
+       tracepoint_get_iter(iter);
+}
+EXPORT_SYMBOL_GPL(tracepoint_iter_next);
+
+void tracepoint_iter_stop(struct tracepoint_iter *iter)
+{
+}
+EXPORT_SYMBOL_GPL(tracepoint_iter_stop);
+
+void tracepoint_iter_reset(struct tracepoint_iter *iter)
+{
+       iter->module = NULL;
+       iter->tracepoint = NULL;
+}
+EXPORT_SYMBOL_GPL(tracepoint_iter_reset);