Function graph tracing is a bit different than the function tracers, as
it is processed after either the ftrace_caller or ftrace_regs_caller
and we only have one place to modify the jump to ftrace_graph_caller,
the jump needs to happen after the restore of registeres.
The function graph tracer is dependent on the function tracer, where
even if the function graph tracing is going on by itself, the save and
restore of registers is still done for function tracing regardless of
if function tracing is happening, before it calls the function graph
code.
If there's no function tracing happening, it is possible to just call
the function graph tracer directly, and avoid the wasted effort to save
and restore regs for function tracing.
This requires adding new flags to the dyn_ftrace records:
FTRACE_FL_TRAMP
FTRACE_FL_TRAMP_EN
The first is set if the count for the record is one, and the ftrace_ops
associated to that record has its own trampoline. That way the mcount code
can call that trampoline directly.
In the future, trampolines can be added to arbitrary ftrace_ops, where you
can have two or more ftrace_ops registered to ftrace (like kprobes and perf)
and if they are not tracing the same functions, then instead of doing a
loop to check all registered ftrace_ops against their hashes, just call the
ftrace_ops trampoline directly, which would call the registered ftrace_ops
function directly.
Without this patch perf showed:
0.05% hackbench [kernel.kallsyms] [k] ftrace_caller
0.05% hackbench [kernel.kallsyms] [k] arch_local_irq_save
0.05% hackbench [kernel.kallsyms] [k] native_sched_clock
0.04% hackbench [kernel.kallsyms] [k] __buffer_unlock_commit
0.04% hackbench [kernel.kallsyms] [k] preempt_trace
0.04% hackbench [kernel.kallsyms] [k] prepare_ftrace_return
0.04% hackbench [kernel.kallsyms] [k] __this_cpu_preempt_check
0.04% hackbench [kernel.kallsyms] [k] ftrace_graph_caller
See that the ftrace_caller took up more time than the ftrace_graph_caller
did.
With this patch:
0.05% hackbench [kernel.kallsyms] [k] __buffer_unlock_commit
0.04% hackbench [kernel.kallsyms] [k] call_filter_check_discard
0.04% hackbench [kernel.kallsyms] [k] ftrace_graph_caller
0.04% hackbench [kernel.kallsyms] [k] sched_clock
The ftrace_caller is no where to be found and ftrace_graph_caller still
takes up the same percentage.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The ftrace dynamic record has a flags element that also has a counter.
Instead of hard coding "rec->flags & ~FTRACE_FL_MASK" all over the
place. Use a macro instead.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
As new flags will be added to the ftrace dynamic record, and since
the flags field is also a counter, converting the numbers used to
do the shifting and masking into a set of macros where we only need
to deal with the max bit count of the counter and the number of bits
for the flags will prevent mistakes in the future.
Dealing with only two numbers is much easier than updating all the
macros that deal with shifting and masking.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Pull tracing updates from Steven Rostedt:
"Lots of tweaks, small fixes, optimizations, and some helper functions
to help out the rest of the kernel to ease their use of trace events.
The big change for this release is the allowing of other tracers, such
as the latency tracers, to be used in the trace instances and allow
for function or function graph tracing to be in the top level
simultaneously"
* tag 'trace-3.16' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace: (44 commits)
tracing: Fix memory leak on instance deletion
tracing: Fix leak of ring buffer data when new instances creation fails
tracing/kprobes: Avoid self tests if tracing is disabled on boot up
tracing: Return error if ftrace_trace_arrays list is empty
tracing: Only calculate stats of tracepoint benchmarks for 2^32 times
tracing: Convert stddev into u64 in tracepoint benchmark
tracing: Introduce saved_cmdlines_size file
tracing: Add __get_dynamic_array_len() macro for trace events
tracing: Remove unused variable in trace_benchmark
tracing: Eliminate double free on failure of allocation on boot up
ftrace/x86: Call text_ip_addr() instead of the duplicated code
tracing: Print max callstack on stacktrace bug
tracing: Move locking of trace_cmdline_lock into start/stop seq calls
tracing: Try again for saved cmdline if failed due to locking
tracing: Have saved_cmdlines use the seq_read infrastructure
tracing: Add tracepoint benchmark tracepoint
tracing: Print nasty banner when trace_printk() is in use
tracing: Add funcgraph_tail option to print function name after closing braces
tracing: Eliminate duplicate TRACE_GRAPH_PRINT_xx defines
tracing: Add __bitmask() macro to trace events to cpumasks and other bitmasks
...
Most archs with HAVE_ARCH_CALLER_ADDR have pretty much the same
definitions of CALLER_ADDRx(n). Instead of duplicating the code for all
the archs, define a ftrace_return_address0() and
ftrace_return_address(n) that can be overwritten by the archs if they
need to do something different. Instead of 7 macros in every arch, we
now only have at most 2 (and actually only 1 as
ftrace_return_address0() should be the same for all archs).
The CALLER_ADDRx(n) will now be defined in linux/ftrace.h and use the
ftrace_return_address*(n?) macros. This removes a lot of the duplicate
code.
Link: http://lkml.kernel.org/p/1400585464-30333-1-git-send-email-takahiro.akashi@linaro.org
Signed-off-by: AKASHI Takahiro <takahiro.akashi@linaro.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
As the decision to what needs to be done (converting a call to the
ftrace_caller to ftrace_caller_regs or to convert from ftrace_caller_regs
to ftrace_caller) can easily be determined from the rec->flags of
FTRACE_FL_REGS and FTRACE_FL_REGS_EN, there's no need to have the
ftrace_check_record() return either a UPDATE_MODIFY_CALL_REGS or a
UPDATE_MODIFY_CALL. Just he latter is enough. This added flag causes
more complexity than is required. Remove it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Move and rename get_ftrace_addr() and get_ftrace_addr_old() to
ftrace_get_addr_new() and ftrace_get_addr_curr() respectively.
This moves these two helper functions in the generic code out from
the arch specific code, and renames them to have a better generic
name. This will allow other archs to use them as well as makes it
a bit easier to work on getting separate trampolines for different
functions.
ftrace_get_addr_new() returns the trampoline address that the mcount
call address will be converted to.
ftrace_get_addr_curr() returns the trampoline address of what the
mcount call address currently jumps to.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
A race exists between module loading and enabling of function tracer.
CPU 1 CPU 2
----- -----
load_module()
module->state = MODULE_STATE_COMING
register_ftrace_function()
mutex_lock(&ftrace_lock);
ftrace_startup()
update_ftrace_function();
ftrace_arch_code_modify_prepare()
set_all_module_text_rw();
<enables-ftrace>
ftrace_arch_code_modify_post_process()
set_all_module_text_ro();
[ here all module text is set to RO,
including the module that is
loading!! ]
blocking_notifier_call_chain(MODULE_STATE_COMING);
ftrace_init_module()
[ tries to modify code, but it's RO, and fails!
ftrace_bug() is called]
When this race happens, ftrace_bug() will produces a nasty warning and
all of the function tracing features will be disabled until reboot.
The simple solution is to treate module load the same way the core
kernel is treated at boot. To hardcode the ftrace function modification
of converting calls to mcount into nops. This is done in init/main.c
there's no reason it could not be done in load_module(). This gives
a better control of the changes and doesn't tie the state of the
module to its notifiers as much. Ftrace is special, it needs to be
treated as such.
The reason this would work, is that the ftrace_module_init() would be
called while the module is in MODULE_STATE_UNFORMED, which is ignored
by the set_all_module_text_ro() call.
Link: http://lkml.kernel.org/r/1395637826-3312-1-git-send-email-indou.takao@jp.fujitsu.com
Reported-by: Takao Indoh <indou.takao@jp.fujitsu.com>
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: stable@vger.kernel.org # 2.6.38+
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Instead of having a list of global functions that are called,
as only one global function is allow to be enabled at a time, there's
no reason to have a list.
Instead, simply have all the users of the global ops, use the global ops
directly, instead of registering their own ftrace_ops. Just switch what
function is used before enabling the function tracer.
This removes a lot of code as well as the complexity involved with it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Create a "set_ftrace_filter" and "set_ftrace_notrace" files in the instance
directories to let users filter of functions to trace for the given instance.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Passing data to the function callback was originally done by adding the
ftrace_ops in another structure, and using the container_of() to get
the field. But this adds a bit more complexity than it is worth, and
adding a simple .private field to ftrace_ops makes things a lot easier.
But be warned, the .private data should not be freed once it is used
unless the ftrace_ops itself has gone through the necessary freeing
routines. A simple synchronize_sched() is not enough as functions
can be traced that are called outside the view of RCU and all its
concoctions.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Trace event triggers added a lseek that uses the ftrace_filter_lseek()
function. Unfortunately, when function tracing is not configured in
that function is not defined and the kernel fails to build.
This is the second time that function was added to a file ops and
it broke the build due to requiring special config dependencies.
Make a generic tracing_lseek() that all the tracing utilities may
use.
Also, modify the old ftrace_filter_lseek() to return 0 instead of
1 on WRONLY. Not sure why it was a 1 as that does not make sense.
This also changes the old tracing_seek() to modify the file pos
pointer on WRONLY as well.
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Tested-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Acked-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Add a traceoff_on_warning option in both the kernel command line as well
as a sysctl option. When set, any WARN*() function that is hit will cause
the tracing_on variable to be cleared, which disables writing to the
ring buffer.
This is useful especially when tracing a bug with function tracing. When
a warning is hit, the print caused by the warning can flood the trace with
the functions that producing the output for the warning. This can make the
resulting trace useless by either hiding where the bug happened, or worse,
by overflowing the buffer and losing the trace of the bug totally.
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Fix a deadlock on ftrace_regex_lock which happens when setting
an enable_event trigger on dynamic kprobe event as below.
----
sh-2.05b# echo p vfs_symlink > kprobe_events
sh-2.05b# echo vfs_symlink:enable_event:kprobes:p_vfs_symlink_0 > set_ftrace_filter
=============================================
[ INFO: possible recursive locking detected ]
3.9.0+ #35 Not tainted
---------------------------------------------
sh/72 is trying to acquire lock:
(ftrace_regex_lock){+.+.+.}, at: [<ffffffff810ba6c1>] ftrace_set_hash+0x81/0x1f0
but task is already holding lock:
(ftrace_regex_lock){+.+.+.}, at: [<ffffffff810b7cbd>] ftrace_regex_write.isra.29.part.30+0x3d/0x220
other info that might help us debug this:
Possible unsafe locking scenario:
CPU0
----
lock(ftrace_regex_lock);
lock(ftrace_regex_lock);
*** DEADLOCK ***
----
To fix that, this introduces a finer regex_lock for each ftrace_ops.
ftrace_regex_lock is too big of a lock which protects all
filter/notrace_hash operations, but it doesn't need to be a global
lock after supporting multiple ftrace_ops because each ftrace_ops
has its own filter/notrace_hash.
Link: http://lkml.kernel.org/r/20130509054417.30398.84254.stgit@mhiramat-M0-7522
Cc: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Tom Zanussi <tom.zanussi@intel.com>
Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
[ Added initialization flag and automate mutex initialization for
non ftrace.c ftrace_probes. ]
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Pull tracing updates from Steven Rostedt:
"Along with the usual minor fixes and clean ups there are a few major
changes with this pull request.
1) Multiple buffers for the ftrace facility
This feature has been requested by many people over the last few
years. I even heard that Google was about to implement it themselves.
I finally had time and cleaned up the code such that you can now
create multiple instances of the ftrace buffer and have different
events go to different buffers. This way, a low frequency event will
not be lost in the noise of a high frequency event.
Note, currently only events can go to different buffers, the tracers
(ie function, function_graph and the latency tracers) still can only
be written to the main buffer.
2) The function tracer triggers have now been extended.
The function tracer had two triggers. One to enable tracing when a
function is hit, and one to disable tracing. Now you can record a
stack trace on a single (or many) function(s), take a snapshot of the
buffer (copy it to the snapshot buffer), and you can enable or disable
an event to be traced when a function is hit.
3) A perf clock has been added.
A "perf" clock can be chosen to be used when tracing. This will cause
ftrace to use the same clock as perf uses, and hopefully this will
make it easier to interleave the perf and ftrace data for analysis."
* tag 'trace-3.10' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace: (82 commits)
tracepoints: Prevent null probe from being added
tracing: Compare to 1 instead of zero for is_signed_type()
tracing: Remove obsolete macro guard _TRACE_PROFILE_INIT
ftrace: Get rid of ftrace_profile_bits
tracing: Check return value of tracing_init_dentry()
tracing: Get rid of unneeded key calculation in ftrace_hash_move()
tracing: Reset ftrace_graph_filter_enabled if count is zero
tracing: Fix off-by-one on allocating stat->pages
kernel: tracing: Use strlcpy instead of strncpy
tracing: Update debugfs README file
tracing: Fix ftrace_dump()
tracing: Rename trace_event_mutex to trace_event_sem
tracing: Fix comment about prefix in arch_syscall_match_sym_name()
tracing: Convert trace_destroy_fields() to static
tracing: Move find_event_field() into trace_events.c
tracing: Use TRACE_MAX_PRINT instead of constant
tracing: Use pr_warn_once instead of open coded implementation
ring-buffer: Add ring buffer startup selftest
tracing: Bring Documentation/trace/ftrace.txt up to date
tracing: Add "perf" trace_clock
...
Conflicts:
kernel/trace/ftrace.c
kernel/trace/trace.c
As ftrace_filter_lseek is now used with ftrace_pid_fops, it needs to
be moved out of the #ifdef CONFIG_DYNAMIC_FTRACE section as the
ftrace_pid_fops is defined when DYNAMIC_FTRACE is not.
Cc: stable@vger.kernel.org
Cc: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently set_ftrace_pid and set_graph_function files use seq_lseek
for their fops. However seq_open() is called only for FMODE_READ in
the fops->open() so that if an user tries to seek one of those file
when she open it for writing, it sees NULL seq_file and then panic.
It can be easily reproduced with following command:
$ cd /sys/kernel/debug/tracing
$ echo 1234 | sudo tee -a set_ftrace_pid
In this example, GNU coreutils' tee opens the file with fopen(, "a")
and then the fopen() internally calls lseek().
Link: http://lkml.kernel.org/r/1365663302-2170-1-git-send-email-namhyung@kernel.org
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: stable@vger.kernel.org
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The function tracing control loop used by perf spits out a warning
if the called function is not a control function. This is because
the control function references a per cpu allocated data structure
on struct ftrace_ops that is not allocated for other types of
functions.
commit 0a016409e4 "ftrace: Optimize the function tracer list loop"
Had an optimization done to all function tracing loops to optimize
for a single registered ops. Unfortunately, this allows for a slight
race when tracing starts or ends, where the stub function might be
called after the current registered ops is removed. In this case we
get the following dump:
root# perf stat -e ftrace:function sleep 1
[ 74.339105] WARNING: at include/linux/ftrace.h:209 ftrace_ops_control_func+0xde/0xf0()
[ 74.349522] Hardware name: PRIMERGY RX200 S6
[ 74.357149] Modules linked in: sg igb iTCO_wdt ptp pps_core iTCO_vendor_support i7core_edac dca lpc_ich i2c_i801 coretemp edac_core crc32c_intel mfd_core ghash_clmulni_intel dm_multipath acpi_power_meter pcspk
r microcode vhost_net tun macvtap macvlan nfsd kvm_intel kvm auth_rpcgss nfs_acl lockd sunrpc uinput xfs libcrc32c sd_mod crc_t10dif sr_mod cdrom mgag200 i2c_algo_bit drm_kms_helper ttm qla2xxx mptsas ahci drm li
bahci scsi_transport_sas mptscsih libata scsi_transport_fc i2c_core mptbase scsi_tgt dm_mirror dm_region_hash dm_log dm_mod
[ 74.446233] Pid: 1377, comm: perf Tainted: G W 3.9.0-rc1 #1
[ 74.453458] Call Trace:
[ 74.456233] [<ffffffff81062e3f>] warn_slowpath_common+0x7f/0xc0
[ 74.462997] [<ffffffff810fbc60>] ? rcu_note_context_switch+0xa0/0xa0
[ 74.470272] [<ffffffff811041a2>] ? __unregister_ftrace_function+0xa2/0x1a0
[ 74.478117] [<ffffffff81062e9a>] warn_slowpath_null+0x1a/0x20
[ 74.484681] [<ffffffff81102ede>] ftrace_ops_control_func+0xde/0xf0
[ 74.491760] [<ffffffff8162f400>] ftrace_call+0x5/0x2f
[ 74.497511] [<ffffffff8162f400>] ? ftrace_call+0x5/0x2f
[ 74.503486] [<ffffffff8162f400>] ? ftrace_call+0x5/0x2f
[ 74.509500] [<ffffffff810fbc65>] ? synchronize_sched+0x5/0x50
[ 74.516088] [<ffffffff816254d5>] ? _cond_resched+0x5/0x40
[ 74.522268] [<ffffffff810fbc65>] ? synchronize_sched+0x5/0x50
[ 74.528837] [<ffffffff811041a2>] ? __unregister_ftrace_function+0xa2/0x1a0
[ 74.536696] [<ffffffff816254d5>] ? _cond_resched+0x5/0x40
[ 74.542878] [<ffffffff8162402d>] ? mutex_lock+0x1d/0x50
[ 74.548869] [<ffffffff81105c67>] unregister_ftrace_function+0x27/0x50
[ 74.556243] [<ffffffff8111eadf>] perf_ftrace_event_register+0x9f/0x140
[ 74.563709] [<ffffffff816254d5>] ? _cond_resched+0x5/0x40
[ 74.569887] [<ffffffff8162402d>] ? mutex_lock+0x1d/0x50
[ 74.575898] [<ffffffff8111e94e>] perf_trace_destroy+0x2e/0x50
[ 74.582505] [<ffffffff81127ba9>] tp_perf_event_destroy+0x9/0x10
[ 74.589298] [<ffffffff811295d0>] free_event+0x70/0x1a0
[ 74.595208] [<ffffffff8112a579>] perf_event_release_kernel+0x69/0xa0
[ 74.602460] [<ffffffff816254d5>] ? _cond_resched+0x5/0x40
[ 74.608667] [<ffffffff8112a640>] put_event+0x90/0xc0
[ 74.614373] [<ffffffff8112a740>] perf_release+0x10/0x20
[ 74.620367] [<ffffffff811a3044>] __fput+0xf4/0x280
[ 74.625894] [<ffffffff811a31de>] ____fput+0xe/0x10
[ 74.631387] [<ffffffff81083697>] task_work_run+0xa7/0xe0
[ 74.637452] [<ffffffff81014981>] do_notify_resume+0x71/0xb0
[ 74.643843] [<ffffffff8162fa92>] int_signal+0x12/0x17
To fix this a new ftrace_ops flag is added that denotes the ftrace_list_end
ftrace_ops stub as just that, a stub. This flag is now checked in the
control loop and the function is not called if the flag is set.
Thanks to Jovi for not just reporting the bug, but also pointing out
where the bug was in the code.
Link: http://lkml.kernel.org/r/514A8855.7090402@redhat.com
Link: http://lkml.kernel.org/r/1364377499-1900-15-git-send-email-jovi.zhangwei@huawei.com
Tested-by: WANG Chao <chaowang@redhat.com>
Reported-by: WANG Chao <chaowang@redhat.com>
Reported-by: zhangwei(Jovi) <jovi.zhangwei@huawei.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When a function probe is created, each function that the probe is
attached to, a "callback" method is called. On release of the probe,
each function entry calls the "free" method.
First, "callback" is a confusing name and does not really match what
it does. Callback sounds like it will be called when the probe
triggers. But that's not the case. This is really an "init" function,
so lets rename it as such.
Secondly, both "init" and "free" do not pass enough information back
to the handlers. Pass back the ops, ip and data for each time the
method is called. We have the information, might as well use it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Steven Rostedt (Red Hat)
Linus Torvalds
AKASHI Takahiro
Sasha Levin
Jiri Slaby
Tom Zanussi
Namhyung Kim
Li Zefan
Masami Hiramatsu