Both Linus (most recent) and Steve (a while ago) reported that perf
related callbacks have massive stack bloat.
The problem is that software events need a pt_regs in order to
properly report the event location and unwind stack. And because we
could not assume one was present we allocated one on stack and filled
it with minimal bits required for operation.
Now, pt_regs is quite large, so this is undesirable. Furthermore it
turns out that most sites actually have a pt_regs pointer available,
making this even more onerous, as the stack space is pointless waste.
This patch addresses the problem by observing that software events
have well defined nesting semantics, therefore we can use static
per-cpu storage instead of on-stack.
Linus made the further observation that all but the scheduler callers
of perf_sw_event() have a pt_regs available, so we change the regular
perf_sw_event() to require a valid pt_regs (where it used to be
optional) and add perf_sw_event_sched() for the scheduler.
We have a scheduler specific call instead of a more generic _noregs()
like construct because we can assume non-recursion from the scheduler
and thereby simplify the code further (_noregs would have to put the
recursion context call inline in order to assertain which __perf_regs
element to use).
One last note on the implementation of perf_trace_buf_prepare(); we
allow .regs = NULL for those cases where we already have a pt_regs
pointer available and do not need another.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Reported-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Javi Merino <javi.merino@arm.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Petr Mladek <pmladek@suse.cz>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Tom Zanussi <tom.zanussi@linux.intel.com>
Cc: Vaibhav Nagarnaik <vnagarnaik@google.com>
Link: http://lkml.kernel.org/r/20141216115041.GW3337@twins.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
If the trace_seq of ftrace_raw_output_prep() is full this function
returns TRACE_TYPE_PARTIAL_LINE, otherwise it returns zero.
The problem is that TRACE_TYPE_PARTIAL_LINE happens to be zero!
The thing is, the caller of ftrace_raw_output_prep() expects a
success to be zero. Change that to expect it to be
TRACE_TYPE_HANDLED.
Link: http://lkml.kernel.org/r/20141114112522.GA2988@dhcp128.suse.cz
Reminded-by: Petr Mladek <pmladek@suse.cz>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Adding a trace_seq_has_overflowed() which returns true if the trace_seq
had too much written into it allows us to simplify the code.
Instead of checking the return value of every call to trace_seq_printf()
and friends, they can all be called normally, and at the end we can
return !trace_seq_has_overflowed() instead.
Several functions also return TRACE_TYPE_PARTIAL_LINE when the trace_seq
overflowed and TRACE_TYPE_HANDLED otherwise. Another helper function
was created called trace_handle_return() which takes a trace_seq and
returns these enums. Using this helper function also simplifies the
code.
This change also makes it possible to remove the return values of
trace_seq_printf() and friends. They should instead just be
void functions.
Link: http://lkml.kernel.org/r/20141114011410.365183157@goodmis.org
Reviewed-by: Petr Mladek <pmladek@suse.cz>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently the __field() macro in TRACE_EVENT is only good for primitive
values, such as integers and pointers, but it fails on complex data types
such as structures or unions. This is because the __field() macro
determines if the variable is signed or not with the test of:
(((type)(-1)) < (type)1)
Unfortunately, that fails when type is a structure.
Since trace events should support structures as fields a new macro
is created for such a case called __field_struct() which acts exactly
the same as __field() does but it does not do the signed type check
and just uses a constant false for that answer.
Cc: Tony Luck <tony.luck@gmail.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
If a trace event uses a dynamic array for something other than a string
then there's currently no way the TP_printk() can figure out what size
it is. A __get_dynamic_array_len() is required to know the length.
This also simplifies the __get_bitmask() macro which required it as well,
but instead just hardcoded it.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Being able to show a cpumask of events can be useful as some events
may affect only some CPUs. There is no standard way to record the
cpumask and converting it to a string is rather expensive during
the trace as traces happen in hotpaths. It would be better to record
the raw event mask and be able to parse it at print time.
The following macros were added for use with the TRACE_EVENT() macro:
__bitmask()
__assign_bitmask()
__get_bitmask()
To test this, I added this to the sched_migrate_task event, which
looked like this:
TRACE_EVENT(sched_migrate_task,
TP_PROTO(struct task_struct *p, int dest_cpu, const struct cpumask *cpus),
TP_ARGS(p, dest_cpu, cpus),
TP_STRUCT__entry(
__array( char, comm, TASK_COMM_LEN )
__field( pid_t, pid )
__field( int, prio )
__field( int, orig_cpu )
__field( int, dest_cpu )
__bitmask( cpumask, num_possible_cpus() )
),
TP_fast_assign(
memcpy(__entry->comm, p->comm, TASK_COMM_LEN);
__entry->pid = p->pid;
__entry->prio = p->prio;
__entry->orig_cpu = task_cpu(p);
__entry->dest_cpu = dest_cpu;
__assign_bitmask(cpumask, cpumask_bits(cpus), num_possible_cpus());
),
TP_printk("comm=%s pid=%d prio=%d orig_cpu=%d dest_cpu=%d cpumask=%s",
__entry->comm, __entry->pid, __entry->prio,
__entry->orig_cpu, __entry->dest_cpu,
__get_bitmask(cpumask))
);
With the output of:
ksmtuned-3613 [003] d..2 485.220508: sched_migrate_task: comm=ksmtuned pid=3615 prio=120 orig_cpu=3 dest_cpu=2 cpumask=00000000,0000000f
migration/1-13 [001] d..5 485.221202: sched_migrate_task: comm=ksmtuned pid=3614 prio=120 orig_cpu=1 dest_cpu=0 cpumask=00000000,0000000f
awk-3615 [002] d.H5 485.221747: sched_migrate_task: comm=rcu_preempt pid=7 prio=120 orig_cpu=0 dest_cpu=1 cpumask=00000000,000000ff
migration/2-18 [002] d..5 485.222062: sched_migrate_task: comm=ksmtuned pid=3615 prio=120 orig_cpu=2 dest_cpu=3 cpumask=00000000,0000000f
Link: http://lkml.kernel.org/r/1399377998-14870-6-git-send-email-javi.merino@arm.com
Link: http://lkml.kernel.org/r/20140506132238.22e136d1@gandalf.local.home
Suggested-by: Javi Merino <javi.merino@arm.com>
Tested-by: Javi Merino <javi.merino@arm.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Register/unregister tracepoint probes with struct tracepoint pointer
rather than tracepoint name.
This change, which vastly simplifies tracepoint.c, has been proposed by
Steven Rostedt. It also removes 8.8kB (mostly of text) to the vmlinux
size.
From this point on, the tracers need to pass a struct tracepoint pointer
to probe register/unregister. A probe can now only be connected to a
tracepoint that exists. Moreover, tracers are responsible for
unregistering the probe before the module containing its associated
tracepoint is unloaded.
text data bss dec hex filename
10443444 4282528 10391552 25117524 17f4354 vmlinux.orig
10434930 4282848 10391552 25109330 17f2352 vmlinux
Link: http://lkml.kernel.org/r/1396992381-23785-2-git-send-email-mathieu.desnoyers@efficios.com
CC: Ingo Molnar <mingo@kernel.org>
CC: Frederic Weisbecker <fweisbec@gmail.com>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: Frank Ch. Eigler <fche@redhat.com>
CC: Johannes Berg <johannes.berg@intel.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
[ SDR - fixed return val in void func in tracepoint_module_going() ]
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
But there were a few features that were added.
Uprobes now work with event triggers and multi buffers.
Uprobes have support under ftrace and perf.
The big feature is that the function tracer can now be used within the
multi buffer instances. That is, you can now trace some functions
in one buffer, others in another buffer, all functions in a third buffer
and so on. They are basically agnostic from each other. This only
works for the function tracer and not for the function graph trace,
although you can have the function graph tracer running in the top level
buffer (or any tracer for that matter) and have different function tracing
going on in the sub buffers.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1
iQEcBAABAgAGBQJTOthtAAoJEKQekfcNnQGu5c8H/Ana/U+0tmksp1dbHkRHsKSH
+Fsv4Jeu8gf1NaFKHEhkUTcFtnzE6qAPV2VCrcJwXbhAhhwZm+LjrnWdoy3215S3
cQW4LftLEonh2cM36Cos74TulMEYN6XmL6dQZV+CILKQkDrWU4qJjQ64okXEkqrd
9iG3p/mSXyvJcmnyg61ALnMOhZDLsXY3djBhWBPhiTPGS6BRb9zh4Pmw6Zv0n2rJ
U93Gt/3AQrv1ybu73dUxqP0abp60oXOiWoF/R2jcbKqIM+K9RPJX79unCV3jq3u9
f+6jMlB9PgAMqQj6ihJdwxKDDuzwyrVdEPnsgvl4jarCBCtVVwhKedBaKN/KS8k=
=HdXY
-----END PGP SIGNATURE-----
Merge tag 'trace-3.15' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace
Pull tracing updates from Steven Rostedt:
"Most of the changes were largely clean ups, and some documentation.
But there were a few features that were added:
Uprobes now work with event triggers and multi buffers and have
support under ftrace and perf.
The big feature is that the function tracer can now be used within the
multi buffer instances. That is, you can now trace some functions in
one buffer, others in another buffer, all functions in a third buffer
and so on. They are basically agnostic from each other. This only
works for the function tracer and not for the function graph trace,
although you can have the function graph tracer running in the top
level buffer (or any tracer for that matter) and have different
function tracing going on in the sub buffers"
* tag 'trace-3.15' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace: (45 commits)
tracing: Add BUG_ON when stack end location is over written
tracepoint: Remove unused API functions
Revert "tracing: Move event storage for array from macro to standalone function"
ftrace: Constify ftrace_text_reserved
tracepoints: API doc update to tracepoint_probe_register() return value
tracepoints: API doc update to data argument
ftrace: Fix compilation warning about control_ops_free
ftrace/x86: BUG when ftrace recovery fails
ftrace: Warn on error when modifying ftrace function
ftrace: Remove freelist from struct dyn_ftrace
ftrace: Do not pass data to ftrace_dyn_arch_init
ftrace: Pass retval through return in ftrace_dyn_arch_init()
ftrace: Inline the code from ftrace_dyn_table_alloc()
ftrace: Cleanup of global variables ftrace_new_pgs and ftrace_update_cnt
tracing: Evaluate len expression only once in __dynamic_array macro
tracing: Correctly expand len expressions from __dynamic_array macro
tracing/module: Replace include of tracepoint.h with jump_label.h in module.h
tracing: Fix event header migrate.h to include tracepoint.h
tracing: Fix event header writeback.h to include tracepoint.h
tracing: Warn if a tracepoint is not set via debugfs
...
I originally wrote commit 35bb4399bd to shrink the size of the overhead of
tracepoints by several kilobytes. Later, I received a patch from Vaibhav
Nagarnaik that fixed a bug in the same code that this commit touches. Not
only did it fix a bug, it also removed code and shrunk the size of the
overhead of trace events even more than this commit did.
Since this commit is scheduled for 3.15 and Vaibhav's patch is already in
mainline, I need to revert this patch in order to keep it from conflicting
with Vaibhav's patch. Not to mention, Vaibhav's patch makes this patch
obsolete.
Link: http://lkml.kernel.org/r/20140320225637.0226041b@gandalf.local.home
Cc: Vaibhav Nagarnaik <vnagarnaik@google.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
In event format strings, the array size is reported in two locations.
One in array subscript and then via the "size:" attribute. The values
reported there have a mismatch.
For e.g., in sched:sched_switch the prev_comm and next_comm character
arrays have subscript values as [32] where as the actual field size is
16.
name: sched_switch
ID: 301
format:
field:unsigned short common_type; offset:0; size:2; signed:0;
field:unsigned char common_flags; offset:2; size:1; signed:0;
field:unsigned char common_preempt_count; offset:3; size:1;signed:0;
field:int common_pid; offset:4; size:4; signed:1;
field:char prev_comm[32]; offset:8; size:16; signed:1;
field:pid_t prev_pid; offset:24; size:4; signed:1;
field:int prev_prio; offset:28; size:4; signed:1;
field:long prev_state; offset:32; size:8; signed:1;
field:char next_comm[32]; offset:40; size:16; signed:1;
field:pid_t next_pid; offset:56; size:4; signed:1;
field:int next_prio; offset:60; size:4; signed:1;
After bisection, the following commit was blamed:
92edca0 tracing: Use direct field, type and system names
This commit removes the duplication of strings for field->name and
field->type assuming that all the strings passed in
__trace_define_field() are immutable. This is not true for arrays, where
the type string is created in event_storage variable and field->type for
all array fields points to event_storage.
Use __stringify() to create a string constant for the type string.
Also, get rid of event_storage and event_storage_mutex that are not
needed anymore.
also, an added benefit is that this reduces the overhead of events a bit more:
text data bss dec hex filename
8424787 2036472 1302528 11763787 b3804b vmlinux
8420814 2036408 1302528 11759750 b37086 vmlinux.patched
Link: http://lkml.kernel.org/r/1392349908-29685-1-git-send-email-vnagarnaik@google.com
Cc: Laurent Chavey <chavey@google.com>
Cc: stable@vger.kernel.org # 3.10+
Signed-off-by: Vaibhav Nagarnaik <vnagarnaik@google.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Use a temporary variable to store the expansion of the len expression.
If the evaluation is expensive, this commit will ensure it is evaluated
only once inside ftrace_get_offsets_<call>.
Link: http://lkml.kernel.org/r/1393651938-16418-3-git-send-email-filbranden@google.com
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
This fixes expansion of the len argument in __dynamic_array macros.
The previous code from commit 7d536cb3f would not fully evaluate the
expression before multiplying its result by the size of the type.
This went unnoticed because the length stored in the high 16 bits of the
offset (which is the one that was broken here) is only used by
filter_pred_strloc which only acts on strings for which the size of the
type is 1.
Link: http://lkml.kernel.org/r/1393651938-16418-2-git-send-email-filbranden@google.com
Signed-off-by: Filipe Brandenburger <filbranden@google.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The functions that assign the contents for the ftrace events are
defined by the TRACE_EVENT() macros. Each event has its own unique
way to assign data to its buffer. When you have over 500 events,
that means there's 500 functions assigning data uniquely for each
event (not really that many, as DECLARE_EVENT_CLASS() and multiple
DEFINE_EVENT()s will only need a single function).
By making helper functions in the core kernel to do some of the work
instead, we can shrink the size of the kernel down a bit.
With a kernel configured with 502 events, the change in size was:
text data bss dec hex filename
12987390 1913504 9785344 24686238 178ae9e /tmp/vmlinux
12959102 1913504 9785344 24657950 178401e /tmp/vmlinux.patched
That's a total of 28288 bytes, which comes down to 56 bytes per event.
Link: http://lkml.kernel.org/r/20120810034708.370808175@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The code that shows array fields for events is defined for all events.
This can add up quite a bit when you have over 500 events.
By making helper functions in the core kernel to do the work
instead, we can shrink the size of the kernel down a bit.
With a kernel configured with 502 events, the change in size was:
text data bss dec hex filename
12990946 1913568 9785344 24689858 178bcc2 /tmp/vmlinux
12987390 1913504 9785344 24686238 178ae9e /tmp/vmlinux.patched
That's a total of 3556 bytes, which comes down to 7 bytes per event.
Although it's not much, this code is just called at initialization of
the events.
Link: http://lkml.kernel.org/r/20120810034708.084036335@goodmis.org
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The code for trace events to format the raw recorded event data
into human readable format in the 'trace' file is repeated for every
event in the system. When you have over 500 events, this can add up
quite a bit.
By making helper functions in the core kernel to do the work
instead, we can shrink the size of the kernel down a bit.
With a kernel configured with 502 events, the change in size was:
text data bss dec hex filename
12991007 1913568 9785344 24689919 178bcff /tmp/vmlinux.orig
12990946 1913568 9785344 24689858 178bcc2 /tmp/vmlinux.patched
Note, this version does not save as much as the version of this patch
I had a few years ago. That is because in the mean time, commit
f71130de5c ("tracing: Add a helper function for event print functions")
did a lot of the work my original patch did. But this change helps
slightly, and is part of a larger clean up to reduce the size much further.
Link: http://lkml.kernel.org/r/20120810034707.378538034@goodmis.org
Cc: Li Zefan <lizefan@huawei.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The event trigger code that checks for callback triggers before and
after recording of an event has lots of flags checks. This code is
duplicated throughout the ftrace events, kprobes and system calls.
They all do the exact same checks against the event flags.
Added helper functions ftrace_trigger_soft_disabled(),
event_trigger_unlock_commit() and event_trigger_unlock_commit_regs()
that consolidated the code and these are used instead.
Link: http://lkml.kernel.org/r/20140106222703.5e7dbba2@gandalf.local.home
Acked-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Tested-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Add a generic event_command.set_trigger_filter() op implementation and
have the current set of trigger commands use it - this essentially
gives them all support for filters.
Syntactically, filters are supported by adding 'if <filter>' just
after the command, in which case only events matching the filter will
invoke the trigger. For example, to add a filter to an
enable/disable_event command:
echo 'enable_event:system:event if common_pid == 999' > \
.../othersys/otherevent/trigger
The above command will only enable the system:event event if the
common_pid field in the othersys:otherevent event is 999.
As another example, to add a filter to a stacktrace command:
echo 'stacktrace if common_pid == 999' > \
.../somesys/someevent/trigger
The above command will only trigger a stacktrace if the common_pid
field in the event is 999.
The filter syntax is the same as that described in the 'Event
filtering' section of Documentation/trace/events.txt.
Because triggers can now use filters, the trigger-invoking logic needs
to be moved in those cases - e.g. for ftrace_raw_event_calls, if a
trigger has a filter associated with it, the trigger invocation now
needs to happen after the { assign; } part of the call, in order for
the trigger condition to be tested.
There's still a SOFT_DISABLED-only check at the top of e.g. the
ftrace_raw_events function, so when an event is soft disabled but not
because of the presence of a trigger, the original SOFT_DISABLED
behavior remains unchanged.
There's also a bit of trickiness in that some triggers need to avoid
being invoked while an event is currently in the process of being
logged, since the trigger may itself log data into the trace buffer.
Thus we make sure the current event is committed before invoking those
triggers. To do that, we split the trigger invocation in two - the
first part (event_triggers_call()) checks the filter using the current
trace record; if a command has the post_trigger flag set, it sets a
bit for itself in the return value, otherwise it directly invoks the
trigger. Once all commands have been either invoked or set their
return flag, event_triggers_call() returns. The current record is
then either committed or discarded; if any commands have deferred
their triggers, those commands are finally invoked following the close
of the current event by event_triggers_post_call().
To simplify the above and make it more efficient, the TRIGGER_COND bit
is introduced, which is set only if a soft-disabled trigger needs to
use the log record for filter testing or needs to wait until the
current log record is closed.
The syscall event invocation code is also changed in analogous ways.
Because event triggers need to be able to create and free filters,
this also adds a couple external wrappers for the existing
create_filter and free_filter functions, which are too generic to be
made extern functions themselves.
Link: http://lkml.kernel.org/r/7164930759d8719ef460357f143d995406e4eead.1382622043.git.tom.zanussi@linux.intel.com
Signed-off-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Add a 'trigger' file for each trace event, enabling 'trace event
triggers' to be set for trace events.
'trace event triggers' are patterned after the existing 'ftrace
function triggers' implementation except that triggers are written to
per-event 'trigger' files instead of to a single file such as the
'set_ftrace_filter' used for ftrace function triggers.
The implementation is meant to be entirely separate from ftrace
function triggers, in order to keep the respective implementations
relatively simple and to allow them to diverge.
The event trigger functionality is built on top of SOFT_DISABLE
functionality. It adds a TRIGGER_MODE bit to the ftrace_event_file
flags which is checked when any trace event fires. Triggers set for a
particular event need to be checked regardless of whether that event
is actually enabled or not - getting an event to fire even if it's not
enabled is what's already implemented by SOFT_DISABLE mode, so trigger
mode directly reuses that. Event trigger essentially inherit the soft
disable logic in __ftrace_event_enable_disable() while adding a bit of
logic and trigger reference counting via tm_ref on top of that in a
new trace_event_trigger_enable_disable() function. Because the base
__ftrace_event_enable_disable() code now needs to be invoked from
outside trace_events.c, a wrapper is also added for those usages.
The triggers for an event are actually invoked via a new function,
event_triggers_call(), and code is also added to invoke them for
ftrace_raw_event calls as well as syscall events.
The main part of the patch creates a new trace_events_trigger.c file
to contain the trace event triggers implementation.
The standard open, read, and release file operations are implemented
here.
The open() implementation sets up for the various open modes of the
'trigger' file. It creates and attaches the trigger iterator and sets
up the command parser. If opened for reading set up the trigger
seq_ops.
The read() implementation parses the event trigger written to the
'trigger' file, looks up the trigger command, and passes it along to
that event_command's func() implementation for command-specific
processing.
The release() implementation does whatever cleanup is needed to
release the 'trigger' file, like releasing the parser and trigger
iterator, etc.
A couple of functions for event command registration and
unregistration are added, along with a list to add them to and a mutex
to protect them, as well as an (initially empty) registration function
to add the set of commands that will be added by future commits, and
call to it from the trace event initialization code.
also added are a couple trigger-specific data structures needed for
these implementations such as a trigger iterator and a struct for
trigger-specific data.
A couple structs consisting mostly of function meant to be implemented
in command-specific ways, event_command and event_trigger_ops, are
used by the generic event trigger command implementations. They're
being put into trace.h alongside the other trace_event data structures
and functions, in the expectation that they'll be needed in several
trace_event-related files such as trace_events_trigger.c and
trace_events.c.
The event_command.func() function is meant to be called by the trigger
parsing code in order to add a trigger instance to the corresponding
event. It essentially coordinates adding a live trigger instance to
the event, and arming the triggering the event.
Every event_command func() implementation essentially does the
same thing for any command:
- choose ops - use the value of param to choose either a number or
count version of event_trigger_ops specific to the command
- do the register or unregister of those ops
- associate a filter, if specified, with the triggering event
The reg() and unreg() ops allow command-specific implementations for
event_trigger_op registration and unregistration, and the
get_trigger_ops() op allows command-specific event_trigger_ops
selection to be parameterized. When a trigger instance is added, the
reg() op essentially adds that trigger to the triggering event and
arms it, while unreg() does the opposite. The set_filter() function
is used to associate a filter with the trigger - if the command
doesn't specify a set_filter() implementation, the command will ignore
filters.
Each command has an associated trigger_type, which serves double duty,
both as a unique identifier for the command as well as a value that
can be used for setting a trigger mode bit during trigger invocation.
The signature of func() adds a pointer to the event_command struct,
used to invoke those functions, along with a command_data param that
can be passed to the reg/unreg functions. This allows func()
implementations to use command-specific blobs and supports code
re-use.
The event_trigger_ops.func() command corrsponds to the trigger 'probe'
function that gets called when the triggering event is actually
invoked. The other functions are used to list the trigger when
needed, along with a couple mundane book-keeping functions.
This also moves event_file_data() into trace.h so it can be used
outside of trace_events.c.
Link: http://lkml.kernel.org/r/316d95061accdee070aac8e5750afba0192fa5b9.1382622043.git.tom.zanussi@linux.intel.com
Signed-off-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Idea-by: Steve Rostedt <rostedt@goodmis.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Pull perf fixes from Ingo Molnar:
"Misc kernel and tooling fixes"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
tools lib traceevent: Fix conversion of pointer to integer of different size
perf/trace: Properly use u64 to hold event_id
perf: Remove fragile swevent hlist optimization
ftrace, perf: Avoid infinite event generation loop
tools lib traceevent: Fix use of multiple options in processing field
perf header: Fix possible memory leaks in process_group_desc()
perf header: Fix bogus group name
perf tools: Tag thread comm as overriden
Vince's perf-trinity fuzzer found yet another 'interesting' problem.
When we sample the irq_work_exit tracepoint with period==1 (or
PERF_SAMPLE_PERIOD) and we add an fasync SIGNAL handler we create an
infinite event generation loop:
,-> <IPI>
| irq_work_exit() ->
| trace_irq_work_exit() ->
| ...
| __perf_event_overflow() -> (due to fasync)
| irq_work_queue() -> (irq_work_list must be empty)
'--------- arch_irq_work_raise()
Similar things can happen due to regular poll() wakeups if we exceed
the ring-buffer wakeup watermark, or have an event_limit.
To avoid this, dis-allow sampling this particular tracepoint.
In order to achieve this, create a special perf_perm function pointer
for each event and call this (when set) on trying to create a
tracepoint perf event.
[ roasted: use expr... to allow for ',' in your expression ]
Reported-by: Vince Weaver <vincent.weaver@maine.edu>
Tested-by: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Dave Jones <davej@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Link: http://lkml.kernel.org/r/20131114152304.GC5364@laptop.programming.kicks-ass.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The trace event filters are still tied to event calls rather than
event files, which means you don't get what you'd expect when using
filters in the multibuffer case:
Before:
# echo 'bytes_alloc > 8192' > /sys/kernel/debug/tracing/events/kmem/kmalloc/filter
# cat /sys/kernel/debug/tracing/events/kmem/kmalloc/filter
bytes_alloc > 8192
# mkdir /sys/kernel/debug/tracing/instances/test1
# echo 'bytes_alloc > 2048' > /sys/kernel/debug/tracing/instances/test1/events/kmem/kmalloc/filter
# cat /sys/kernel/debug/tracing/events/kmem/kmalloc/filter
bytes_alloc > 2048
# cat /sys/kernel/debug/tracing/instances/test1/events/kmem/kmalloc/filter
bytes_alloc > 2048
Setting the filter in tracing/instances/test1/events shouldn't affect
the same event in tracing/events as it does above.
After:
# echo 'bytes_alloc > 8192' > /sys/kernel/debug/tracing/events/kmem/kmalloc/filter
# cat /sys/kernel/debug/tracing/events/kmem/kmalloc/filter
bytes_alloc > 8192
# mkdir /sys/kernel/debug/tracing/instances/test1
# echo 'bytes_alloc > 2048' > /sys/kernel/debug/tracing/instances/test1/events/kmem/kmalloc/filter
# cat /sys/kernel/debug/tracing/events/kmem/kmalloc/filter
bytes_alloc > 8192
# cat /sys/kernel/debug/tracing/instances/test1/events/kmem/kmalloc/filter
bytes_alloc > 2048
We'd like to just move the filter directly from ftrace_event_call to
ftrace_event_file, but there are a couple cases that don't yet have
multibuffer support and therefore have to continue using the current
event_call-based filters. For those cases, a new USE_CALL_FILTER bit
is added to the event_call flags, whose main purpose is to keep the
old behavior for those cases until they can be updated with
multibuffer support; at that point, the USE_CALL_FILTER flag (and the
new associated call_filter_check_discard() function) can go away.
The multibuffer support also made filter_current_check_discard()
redundant, so this change removes that function as well and replaces
it with filter_check_discard() (or call_filter_check_discard() as
appropriate).
Link: http://lkml.kernel.org/r/f16e9ce4270c62f46b2e966119225e1c3cca7e60.1382620672.git.tom.zanussi@linux.intel.com
Signed-off-by: Tom Zanussi <tom.zanussi@linux.intel.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
perf_trace_buf_prepare() + perf_trace_buf_submit(task => NULL)
make no sense if hlist_empty(head). Change perf_trace_##call()
to check ->perf_events beforehand and do nothing if it is empty.
This removes the overhead for tasks without events associated
with them. For example, "perf record -e sched:sched_switch -p1"
attaches the counter(s) to the single task, but every task in
system will do perf_trace_buf_prepare/submit() just to realize
that it was not attached to this event.
However, we can only do this if __task == NULL, so we also add
the __builtin_constant_p(__task) check.
With this patch "perf bench sched pipe" shows approximately 4%
improvement when "perf record -p1" runs in parallel, many thanks
to Steven for the testing.
Link: http://lkml.kernel.org/r/20130806160847.GA2746@redhat.com
Tested-by: David Ahern <dsahern@gmail.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The next patch tries to avoid the costly perf_trace_buf_* calls
when possible but there is a problem. We can only do this if
__task == NULL, perf_tp_event(task != NULL) has the additional
code for this case.
Unfortunately, TP_perf_assign/__perf_xxx which changes the default
values of __count/__task variables for perf_trace_buf_submit() is
called "too late", after we already did perf_trace_buf_prepare(),
and the optimization above can't work.
So this patch simply embeds __perf_xxx() into TP_ARGS(), this way
DECLARE_EVENT_CLASS() can use the result of assignments hidden in
"args" right after ftrace_get_offsets_##call() which is mostly
trivial. This allows us to have the fast-path "__task != NULL"
check at the start, see the next patch.
Link: http://lkml.kernel.org/r/20130806160844.GA2739@redhat.com
Tested-by: David Ahern <dsahern@gmail.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
To simplify the review of the next patches:
1. We are going to reimplent __perf_task/counter and embedd them
into TP_ARGS(). expand TRACE_EVENT(sched_stat_runtime) into
DECLARE_EVENT_CLASS() + DEFINE_EVENT(), this way they can use
different TP_ARGS's.
2. Change perf_trace_##call() macro to do perf_fetch_caller_regs()
right before perf_trace_buf_prepare().
This way it evaluates TP_ARGS() asap, the next patch explores
this fact.
Note: after 87f44bbc perf_trace_buf_prepare() doesn't need
"struct pt_regs *regs", perhaps it makes sense to remove this
argument. And perhaps we can teach perf_trace_buf_submit()
to accept regs == NULL and do fetch_caller_regs(CALLER_ADDR1)
in this case.
3. Cosmetic, but the typecast from "void*" buys nothing. It just
adds the noise, remove it.
Link: http://lkml.kernel.org/r/20130806160841.GA2736@redhat.com
Acked-by: Peter Zijlstra <peterz@infradead.org>
Tested-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Every perf_trace_buf_prepare() caller does
WARN_ONCE(size > PERF_MAX_TRACE_SIZE, message) and "message" is
almost the same.
Shift this WARN_ONCE() into perf_trace_buf_prepare(). This changes
the meaning of _ONCE, but I think this is fine.
- 4947014 2932448 10104832 17984294 1126b26 vmlinux
+ 4948422 2932448 10104832 17985702 11270a6 vmlinux
on my build.
Link: http://lkml.kernel.org/r/20130617170211.GA19813@redhat.com
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Each TRACE_EVENT() adds several helper functions. If two or more trace events
share the same structure and print format, they can also share most of these
helper functions and save a lot of space from duplicate code. This is why the
DECLARE_EVENT_CLASS() and DEFINE_EVENT() were created.
Some events require a trigger to be called at registering and unregistering of
the event and to do so they use TRACE_EVENT_FN().
If multiple events require a trigger, they currently have no choice but to use
TRACE_EVENT_FN() as there's no DEFINE_EVENT_FN() available. This unfortunately
causes a lot of wasted duplicate code created.
By adding a DEFINE_EVENT_FN(), these events can still use a
DECLARE_EVENT_CLASS() and then define their own triggers.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Link: http://lkml.kernel.org/r/51C3236C.8030508@hds.com
Signed-off-by: Seiji Aguchi <seiji.aguchi@hds.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
The macro _TRACE_PROFILE_INIT was removed a long time ago,
but an "#undef" guard was left behind. Remove it.
Link: http://lkml.kernel.org/r/514684EE.6000805@huawei.com
Signed-off-by: zhangwei(Jovi) <jovi.zhangwei@huawei.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
In order to let triggers enable or disable events, we need a 'soft'
method for doing so. For example, if a function probe is added that
lets a user enable or disable events when a function is called, that
change must be done without taking locks or a mutex, and definitely
it can't sleep. But the full enabling of a tracepoint is expensive.
By adding a 'SOFT_DISABLE' flag, and converting the flags to be updated
without the protection of a mutex (using set/clear_bit()), this soft
disable flag can be used to allow critical sections to enable or disable
events from being traced (after the event has been placed into "SOFT_MODE").
Some caveats though: The comm recorder (to map pids with a comm) can not
be soft disabled (yet). If you disable an event with with a "soft"
disable and wait a while before reading the trace, the comm cache may be
replaced and you'll get a bunch of <...> for comms in the trace.
Reading the "enable" file for an event that is disabled will now give
you "0*" where the '*' denotes that the tracepoint is still active but
the event itself is "disabled".
[ fixed _BIT used in & operation : thanks to Dan Carpenter and smatch ]
Cc: Dan Carpenter <dan.carpenter@oracle.com>
Cc: Tom Zanussi <tom.zanussi@linux.intel.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
As we've added __init annotation to field-defining functions, we should
add __refdata annotation to event_call variables, which reference those
functions.
Link: http://lkml.kernel.org/r/51343C1F.2050502@huawei.com
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Move duplicate code in event print functions to a helper function.
This shrinks the size of the kernel by ~13K.
text data bss dec hex filename
6596137 1743966 10138672 18478775 119f6b7 vmlinux.o.old
6583002 1743849 10138672 18465523 119c2f3 vmlinux.o.new
Link: http://lkml.kernel.org/r/51258746.2060304@huawei.com
Signed-off-by: Li Zefan <lizefan@huawei.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Pass the struct ftrace_event_file *ftrace_file to the
trace_event_buffer_lock_reserve() (new function that replaces the
trace_current_buffer_lock_reserver()).
The ftrace_file holds a pointer to the trace_array that is in use.
In the case of multiple buffers with different trace_arrays, this
allows different events to be recorded into different buffers.
Also fixed some of the stale comments in include/trace/ftrace.h
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
The trace events for ftrace are all defined via global variables.
The arrays of events and event systems are linked to a global list.
This prevents multiple users of the event system (what to enable and
what not to).
By adding descriptors to represent the event/file relation, as well
as to which trace_array descriptor they are associated with, allows
for more than one set of events to be defined. Once the trace events
files have a link between the trace event and the trace_array they
are associated with, we can create multiple trace_arrays that can
record separate events in separate buffers.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
When doing per-cpu helper optimizing work, find that this code is so puzzled.
1. It's mark as comment text, maybe a sample function for guidelines
or a todo work.
2. But, this sample code is odd where struct perf_trace_buf is nonexistent.
commit ce71b9 delete struct perf_trace_buf definition.
Author: Frederic Weisbecker <fweisbec@gmail.com>
Date: Sun Nov 22 05:26:55 2009 +0100
tracing: Use the perf recursion protection from trace event
Is it necessary to keep there?
just compile test.
Link: http://lkml.kernel.org/r/50949FC9.6050202@gmail.com
Signed-off-by: Shan Wei <davidshan@tencent.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Have the ring buffer commit function use the irq_work infrastructure to
wake up any waiters waiting on the ring buffer for new data. The irq_work
was created for such a purpose, where doing the actual wake up at the
time of adding data is too dangerous, as an event or function trace may
be in the midst of the work queue locks and cause deadlocks. The irq_work
will either delay the action to the next timer interrupt, or trigger an IPI
to itself forcing an interrupt to do the work (in a safe location).
With irq_work, all ring buffer commits can safely do wakeups, removing
the need for the ring buffer commit "nowake" variants, which were used
by events and function tracing. All commits can now safely use the
normal commit, and the "nowake" variants can be removed.
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
A few events are interesting not only for a current task.
For example, sched_stat_* events are interesting for a task
which wakes up. For this reason, it will be good if such
events will be delivered to a target task too.
Now a target task can be set by using __perf_task().
The original idea and a draft patch belongs to Peter Zijlstra.
I need these events for profiling sleep times. sched_switch is used for
getting callchains and sched_stat_* is used for getting time periods.
These events are combined in user space, then it can be analyzed by
perf tools.
Inspired-by: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Arun Sharma <asharma@fb.com>
Signed-off-by: Andrew Vagin <avagin@openvz.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1342016098-213063-1-git-send-email-avagin@openvz.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Each event adds some points to its counters. By default it adds 1,
and a number of points may be transmited in event's parameters.
E.g. sched:sched_stat_runtime adds how long process has been running.
But this functionality was broken by v2.6.31-rc5-392-gf413cdb
and now the event's parameters doesn't affect on a number of points.
TP_perf_assign isn't defined, so __perf_count(c) isn't executed and
__count is always equal to 1.
Signed-off-by: Andrew Vagin <avagin@openvz.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1317052535-1765247-2-git-send-email-avagin@openvz.org
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Filesystem, like Btrfs, has some "ULL" macros, and when these macros are passed
to tracepoints'__print_symbolic(), there will be 64->32 truncate WARNINGS during
compiling on 32bit box.
Signed-off-by: Liu Bo <liubo2009@cn.fujitsu.com>
Link: http://lkml.kernel.org/r/4DACE6E0.7000507@cn.fujitsu.com
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently the trace_event structures are placed in the _ftrace_events
section, and at link time, the linker makes one large array of all
the trace_event structures. On boot up, this array is read (much like
the initcall sections) and the events are processed.
The problem is that there is no guarantee that gcc will place complex
structures nicely together in an array format. Two structures in the
same file may be placed awkwardly, because gcc has no clue that they
are suppose to be in an array.
A hack was used previous to force the alignment to 4, to pack the
structures together. But this caused alignment issues with other
architectures (sparc).
Instead of packing the structures into an array, the structures' addresses
are now put into the _ftrace_event section. As pointers are always the
natural alignment, gcc should always pack them tightly together
(otherwise initcall, extable, etc would also fail).
By having the pointers to the structures in the section, we can still
iterate the trace_events without causing unnecessary alignment problems
with other architectures, or depending on the current behaviour of
gcc that will likely change in the future just to tick us kernel developers
off a little more.
The _ftrace_event section is also moved into the .init.data section
as it is now only needed at boot up.
Suggested-by: David Miller <davem@davemloft.net>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Currently we have in something like the sched_switch event:
field:char prev_comm[TASK_COMM_LEN]; offset:12; size:16; signed:1;
When a userspace tool such as perf tries to parse this, the
TASK_COMM_LEN is meaningless. This is done because the TRACE_EVENT() macro
simply uses a #len to show the string of the length. When the length is
an enum, we get a string that means nothing for tools.
By adding a static buffer and a mutex to protect it, we can store the
string into that buffer with snprintf and show the actual number.
Now we get:
field:char prev_comm[16]; offset:12; size:16; signed:1;
Something much more useful.
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
As for the raw syscalls events, individual syscall events won't
leak system wide information on task bound tracing. Allow non
privileged users to use them in such workflow.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Jason Baron <jbaron@redhat.com>
This introduces the new TRACE_EVENT_FLAGS() macro in order
to set up initial event flags value.
This macro must simply follow the definition of a trace event
and take the event name and the flag value as parameters:
TRACE_EVENT(my_event, .....
....
);
TRACE_EVENT_FLAGS(my_event, 1)
This will set up 1 as the initial my_event->flags value.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Jason Baron <jbaron@redhat.com>
Drop the cpparg() macro that wraps CPP parameters. We already have
the PARAM() macro for that, no need to have several versions.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
__print_flags() and __print_symbolic() use percpu trace_seq:
1) Its memory is allocated at compile time, it wastes memory if we don't use tracing.
2) It is percpu data and it wastes more memory for multi-cpus system.
3) It disables preemption when it executes its core routine
"trace_seq_printf(s, "%s: ", #call);" and introduces latency.
So we move this trace_seq to struct trace_iterator.
Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
LKML-Reference: <4C078350.7090106@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Because kprobes and syscalls need special processing to register
events, the class->reg() method was created to handle the differences.
But instead of creating a default ->reg for perf and ftrace events,
the code was scattered with:
if (class->reg)
class->reg();
else
default_reg();
This is messy and can also lead to bugs.
This patch cleans up this code and creates a default reg() entry for
the events allowing for the code to directly call the class->reg()
without the condition.
Reported-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Drop this argument now that we always want to rewind only to the
state of the first caller.
It means frame pointers are not necessary anymore to reliably get
the source of an event. But this also means we need this helper
to be a macro now, as an inline function is not an option since
we need to know when to provide a default implentation.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
Cc: David Miller <davem@davemloft.net>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
