Impact: Rework the perfcounter output ABI

use sys_read() only for instant data and provide mmap() output for all
async overflow data.

The first mmap() determines the size of the output buffer. The mmap()
size must be a PAGE_SIZE multiple of 1+pages, where pages must be a
power of 2 or 0. Further mmap()s of the same fd must have the same
size. Once all maps are gone, you can again mmap() with a new size.

In case of 0 extra pages there is no data output and the first page
only contains meta data.

When there are data pages, a poll() event will be generated for each
full page of data. Furthermore, the output is circular. This means
that although 1 page is a valid configuration, its useless, since
we'll start overwriting it the instant we report a full page.

Future work will focus on the output format (currently maintained)
where we'll likey want each entry denoted by a header which includes a
type and length.

Further future work will allow to splice() the fd, also containing the
async overflow data -- splice() would be mutually exclusive with
mmap() of the data.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Orig-LKML-Reference: <20090323172417.470536358@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: new feature giving performance improvement

This adds the ability for userspace to do an mmap on a hardware counter
fd and get access to a read-only page that contains the information
needed to translate a hardware counter value to the full 64-bit
counter value that would be returned by a read on the fd.  This is
useful on architectures that allow user programs to read the hardware
counters, such as PowerPC.

The mmap will only succeed if the counter is a hardware counter
monitoring the current process.

On my quad 2.5GHz PowerPC 970MP machine, userspace can read a counter
and translate it to the full 64-bit value in about 30ns using the
mmapped page, compared to about 830ns for the read syscall on the
counter, so this does give a significant performance improvement.
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Orig-LKML-Reference: <20090323172417.297057964@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Tracepoint events like lock_acquire and software counters like
pagefaults can recurse into the perf counter code again, avoid that.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Orig-LKML-Reference: <20090323172417.152096433@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Since the bitfields turned into a bit of a mess, remove them and rely on
good old masks.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Orig-LKML-Reference: <20090323172417.059499915@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: build fix for powerpc

Commit db3a944aca35ae61 ("perf_counter: revamp syscall input ABI")
expanded the hw_event.type field into a union of structs containing
bitfields.  In particular it introduced a type field and a raw_type
field, with the intention that the 1-bit raw_type field should
overlay the most-significant bit of the 8-bit type field, and in fact
perf_counter_alloc() now assumes that (or at least, assumes that
raw_type doesn't overlay any of the bits that are 1 in the values of
PERF_TYPE_{HARDWARE,SOFTWARE,TRACEPOINT}).

Unfortunately this is not true on big-endian systems such as PowerPC,
where bitfields are laid out from left to right, i.e. from most
significant bit to least significant.  This means that setting
hw_event.type = PERF_TYPE_SOFTWARE will set hw_event.raw_type to 1.

This fixes it by making the layout depend on whether or not
__BIG_ENDIAN_BITFIELD is defined.  It's a bit ugly, but that's what
we get for using bitfields in a user/kernel ABI.

Also, that commit didn't fix up some places in arch/powerpc/kernel/
perf_counter.c where hw_event.raw and hw_event.event_id were used.
This fixes them too.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Impact: cleanup

Having 3 slightly different copies of the same code around does nobody
any good. First step in revamping the output format.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Orig-LKML-Reference: <20090319194233.929962222@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: modify ABI

The hardware/software classification in hw_event->type became a little
strained due to the addition of tracepoint tracing.

Instead split up the field and provide a type field to explicitly specify
the counter type, while using the event_id field to specify which event to
use.

Raw counters still work as before, only the raw config now goes into
raw_event.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Orig-LKML-Reference: <20090319194233.836807573@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: new perfcounters feature

Enable usage of tracepoints as perf counter events.

tracepoint event ids can be found in /debug/tracing/event/*/*/id
and (for now) are represented as -65536+id in the type field.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Orig-LKML-Reference: <20090319194233.744044174@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

I noticed that the counter_list only includes top-level counters, thus
perf_swcounter_event() will miss sw-counters in groups.

Since perf_swcounter_event() also wants an RCU safe list, create a new
event_list that includes all counters and uses RCU list ops and use call_rcu
to free the counter structure.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Use hrtimers to profile timer based sampling for the software time
counters.

This allows platforms without hardware counter support to still
perform sample based profiling.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Provide separate sw counters for major and minor page faults.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Provide generic software counter infrastructure that supports
software events.

This will be used to allow sample based profiling based on software
events such as pagefaults. The current infrastructure can only
provide a count of such events, no place information.
Signed-off-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

The sum of bits is 65 currently not 64 - so reduce the # of reserved bits
from 55 to 54.

Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: ABI change

This expands several fields in the perf_counter_hw_event struct and adds
a "flags" argument to the perf_counter_open system call, in order that
features can be added in future without ABI changes.

In particular the record_type field is expanded to 64 bits, and the
space for flag bits has been expanded from 32 to 64 bits.

This also adds some new fields:

* read_format (64 bits) is intended to provide a way to specify what
  userspace wants to get back when it does a read() on a simple
  (non-interrupting) counter;

* exclude_idle (1 bit) provides a way for userspace to ask that events
  that occur when the cpu is idle be excluded;

* extra_config_len will provide a way for userspace to supply an
  arbitrary amount of extra machine-specific PMU configuration data
  immediately following the perf_counter_hw_event struct, to allow
  sophisticated users to program things such as instruction matching
  CAMs and address range registers;

* __reserved_3 and __reserved_4 provide space for future expansion.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

This fixes three issues noticed by Arnd Bergmann:

- Add #ifdef __KERNEL__ and move some things around in perf_counter.h
  to make sure only the bits that userspace needs are exported to
  userspace.

- Use __u64, __s64, __u32 types in the structs exported to userspace
  rather than u64, s64, u32.

- Make the sys_perf_counter_open syscall available to the SPUs on
  Cell platforms.

And one issue that I noticed in looking at the code again:

- Wrap the perf_counter_open syscall with SYSCALL_DEFINE4 so we get
  the proper handling of int arguments on ppc64 (and some other 64-bit
  architectures).
Reported-by: NArnd Bergmann <arnd@arndb.de>
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Jaswinder Singh Rajput reported that commit 23a185ca caused the
context switch and migration software counters to report zero always.
With that commit, the software counters only count events that occur
between sched-in and sched-out for a task.  This is necessary for the
counter enable/disable prctls and ioctls to work.  However, the
context switch and migration counts are incremented after sched-out
for one task and before sched-in for the next.  Since the increment
doesn't occur while a task is scheduled in (as far as the software
counters are concerned) it doesn't count towards any counter.

Thus the context switch and migration counters need to count events
that occur at any time, provided the counter is enabled, not just
those that occur while the task is scheduled in (from the perf_counter
subsystem's point of view).  The problem though is that the software
counter code can't tell the difference between being enabled and being
scheduled in, and between being disabled and being scheduled out,
since we use the one pair of enable/disable entry points for both.
That is, the high-level disable operation simply arranges for the
counter to not be scheduled in any more, and the high-level enable
operation arranges for it to be scheduled in again.

One way to solve this would be to have sched_in/out operations in the
hw_perf_counter_ops struct as well as enable/disable.  However, this
takes a simpler approach: it adds a 'prev_state' field to the
perf_counter struct that allows a counter's enable method to know
whether the counter was previously disabled or just inactive
(scheduled out), and therefore whether the enable method is being
called as a result of a high-level enable or a schedule-in operation.

This then allows the context switch, migration and page fault counters
to reset their hw.prev_count value in their enable functions only if
they are called as a result of a high-level enable operation.
Although page faults would normally only occur while the counter is
scheduled in, this changes the page fault counter code too in case
there are ever circumstances where page faults get counted against a
task while its counters are not scheduled in.
Reported-by: NJaswinder Singh Rajput <jaswinder@kernel.org>
Signed-off-by: NPaul Mackerras <paulus@samba.org>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: new perf_counter feature

This extends the perf_counter_hw_event struct with bits that specify
that events in user, kernel and/or hypervisor mode should not be
counted (i.e. should be excluded), and adds code to program the PMU
mode selection bits accordingly on x86 and powerpc.

For software counters, we don't currently have the infrastructure to
distinguish which mode an event occurs in, so we currently fail the
counter initialization if the setting of the hw_event.exclude_* bits
would require us to distinguish.  Context switches and CPU migrations
are currently considered to occur in kernel mode.

On x86, this changes the previous policy that only root can count
kernel events.  Now non-root users can count kernel events or exclude
them.  Non-root users still can't use NMI events, though.  On x86 we
don't appear to have any way to control whether hypervisor events are
counted or not, so hw_event.exclude_hv is ignored.

On powerpc, the selection of whether to count events in user, kernel
and/or hypervisor mode is PMU-wide, not per-counter, so this adds a
check that the hw_event.exclude_* settings are the same as other events
on the PMU.  Counters being added to a group have to have the same
settings as the other hardware counters in the group.  Counters and
groups can only be enabled in hw_perf_group_sched_in or power_perf_enable
if they have the same settings as any other counters already on the
PMU.  If we are not running on a hypervisor, the exclude_hv setting
is ignored (by forcing it to 0) since we can't ever get any
hypervisor events.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Ratelimit performance counter interrupts to 100KHz per CPU.

This replaces the irq-delta-time based method.
Signed-off-by: NMike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Starting kerneltop with only -c 100 seems to be a bad idea, it can
easily lock the system due to perfcounter IRQ overload.

So add throttling: if a new IRQ arrives in a shorter than
PERFMON_MIN_PERIOD_NS time, turn off perfcounters and untrottle them
from the next timer tick.
Signed-off-by: NMike Galbraith <efault@gmx.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: New perf_counter features

This primarily adds a way for perf_counter users to enable and disable
counters and groups.  Enabling or disabling a counter or group also
enables or disables all of the child counters that have been cloned
from it to monitor children of the task monitored by the top-level
counter.  The userspace interface to enable/disable counters is via
ioctl on the counter file descriptor.

Along the way this extends the code that handles child counters to
handle child counter groups properly.  A group with multiple counters
will be cloned to child tasks if and only if the group leader has the
hw_event.inherit bit set - if it is set the whole group is cloned as a
group in the child task.

In order to be able to enable or disable all child counters of a given
top-level counter, we need a way to find them all.  Hence I have added
a child_list field to struct perf_counter, which is the head of the
list of children for a top-level counter, or the link in that list for
a child counter.  That list is protected by the perf_counter.mutex
field.

This also adds a mutex to the perf_counter_context struct.  Previously
the list of counters was protected just by the lock field in the
context, which meant that perf_counter_init_task had to take that lock
and then take whatever lock/mutex protects the top-level counter's
child_list.  But the counter enable/disable functions need to take
that lock in order to traverse the list, then for each counter take
the lock in that counter's context in order to change the counter's
state safely, which will lead to a deadlock.

To solve this, we now have both a mutex and a spinlock in the context,
and taking either is sufficient to ensure the list of counters can't
change - you have to take both before changing the list.  Now
perf_counter_init_task takes the mutex instead of the lock (which
incidentally means that inherit_counter can use GFP_KERNEL instead of
GFP_ATOMIC) and thus avoids the possible deadlock.  Similarly the new
enable/disable functions can take the mutex while traversing the list
of child counters without incurring a possible deadlock when the
counter manipulation code locks the context for a child counter.

We also had an misfeature that the first counter added to a context
would possibly not go on until the next sched-in, because we were
using ctx->nr_active to detect if the context was running on a CPU.
But nr_active is the number of active counters, and if that was zero
(because the context didn't have any counters yet) it would look like
the context wasn't running on a cpu and so the retry code in
__perf_install_in_context wouldn't retry.  So this adds an 'is_active'
field that is set when the context is on a CPU, even if it has no
counters.  The is_active field is only used for task contexts, not for
per-cpu contexts.

If we enable a subsidiary counter in a group that is active on a CPU,
and the arch code can't enable the counter, then we have to pull the
whole group off the CPU.  We do this with group_sched_out, which gets
moved up in the file so it comes before all its callers.  This also
adds similar logic to __perf_install_in_context so that the "all on,
or none" invariant of groups is preserved when adding a new counter to
a group.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Impact: New perf_counter features

A pinned counter group is one that the user wants to have on the CPU
whenever possible, i.e. whenever the associated task is running, for
a per-task group, or always for a per-cpu group.  If the system
cannot satisfy that, it puts the group into an error state where
it is not scheduled any more and reads from it return EOF (i.e. 0
bytes read).  The group can be released from error state and made
readable again using prctl(PR_TASK_PERF_COUNTERS_ENABLE).  When we
have finer-grained enable/disable controls on counters we'll be able
to reset the error state on individual groups.

An exclusive group is one that the user wants to be the only group
using the CPU performance monitor hardware whenever it is on.  The
counter group scheduler will not schedule an exclusive group if there
are already other groups on the CPU and will not schedule other groups
onto the CPU if there is an exclusive group scheduled (that statement
does not apply to groups containing only software counters, which can
always go on and which do not prevent an exclusive group from going on).
With an exclusive group, we will be able to let users program PMU
registers at a low level without the concern that those settings will
perturb other measurements.

Along the way this reorganizes things a little:
- is_software_counter() is moved to perf_counter.h.
- cpuctx->active_oncpu now records the number of hardware counters on
  the CPU, i.e. it now excludes software counters.  Nothing was reading
  cpuctx->active_oncpu before, so this change is harmless.
- A new cpuctx->exclusive field records whether we currently have an
  exclusive group on the CPU.
- counter_sched_out moves higher up in perf_counter.c and gets called
  from __perf_counter_remove_from_context and __perf_counter_exit_task,
  where we used to have essentially the same code.
- __perf_counter_sched_in now goes through the counter list twice, doing
  the pinned counters in the first loop and the non-pinned counters in
  the second loop, in order to give the pinned counters the best chance
  to be scheduled in.

Note that only a group leader can be exclusive or pinned, and that
attribute applies to the whole group.  This avoids some awkwardness in
some corner cases (e.g. where a group leader is closed and the other
group members get added to the context list).  If we want to relax that
restriction later, we can, and it is easier to relax a restriction than
to apply a new one.

This doesn't yet handle the case where a pinned counter is inherited
and goes into error state in the child - the error state is not
propagated up to the parent when the child exits, and arguably it
should.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Impact: extend perf_counter infrastructure

This adds an optional hw_perf_group_sched_in() arch function that enables
a whole group of counters in one go.  It returns 1 if it added the group
successfully, 0 if it did nothing (and therefore the core needs to add
the counters individually), or a negative number if an error occurred.
It should add all the counters and enable any software counters in the
group, or else add none of them and return an error.

There are a couple of related changes/improvements in the group handling
here:

* As an optimization, group_sched_out() and group_sched_in() now check the
  state of the group leader, and do nothing if the leader is not active
  or disabled.

* We now call hw_perf_save_disable/hw_perf_restore around the complete
  set of counter enable/disable calls in __perf_counter_sched_in/out,
  to give the arch code the opportunity to defer updating the hardware
  state until the hw_perf_restore call if it wants.

* We no longer stop adding groups after we get to a group that has more
  than one counter.  We will ultimately add an option for a group to be
  exclusive.  The current code doesn't really implement exclusive groups
  anyway, since a group could end up going on with other counters that
  get added before it.
Signed-off-by: NPaul Mackerras <paulus@samba.org>

Impact: build fix on ia64

KOSAKI Motohiro reported that -tip doesnt build on ia64 because
asm/perf_counter.h only exists on x86 for now. Fix it.
Reported-by: NKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Tested-by: NKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Acked-by: NKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Generalize "bus cycles" hw events - and map them to CPU_CLK_Unhalted.Ref
on x86. (which is a good enough approximation)
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: remove dead code

nr_inherited was not maintained correctly (not decremented) - and also
not used - remove it.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Allow lowlevel ->enable() op to return an error if a counter can not be
added. This can be used to handle counter constraints.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: rename field names

Shorten them.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: refactor the x86 code for fixed-mode PMCs

Extend the data structures and rename the existing facilities
to allow for a 'generic' versus 'fixed' counter distinction.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: remove debug checks
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: add new feature, new sw counter

Add a counter that counts the number of cross-CPU migrations a
task is suffering.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: add new feature, new sw counter

Add a counter that counts the number of context-switches a task
is doing.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: implement new performance feature

Counter inheritance can be used to run performance counters in a workload,
transparently - and pipe back the counter results to the parent counter.

Inheritance for performance counters works the following way: when creating
a counter it can be marked with the .inherit=1 flag. Such counters are then
'inherited' by all child tasks (be they fork()-ed or clone()-ed). These
counters get inherited through exec() boundaries as well (except through
setuid boundaries).

The counter values get added back to the parent counter(s) when the child
task(s) exit - much like stime/utime statistics are gathered. So inherited
counters are ideal to gather summary statistics about an application's
behavior via shell commands, without having to modify that application.

The timec.c command utilizes counter inheritance:

  http://redhat.com/~mingo/perfcounters/timec.c

Sample output:

   $ ./timec -e 1 -e 3 -e 5 ls -lR /usr/include/ >/dev/null

   Performance counter stats for 'ls':

           163516953 instructions
                2295 cache-misses
             2855182 branch-misses
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: restructure code

Change counter math from absolute values to clear delta logic.

We try to extract elapsed deltas from the raw hw counter - and put
that into the generic counter.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: cleanup

Introduce a proper enum for the 3 states of a counter:

	PERF_COUNTER_STATE_OFF		= -1
	PERF_COUNTER_STATE_INACTIVE	=  0
	PERF_COUNTER_STATE_ACTIVE	=  1

and rename counter->active to counter->state and propagate the
changes everywhere.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Add a way for self-monitoring tasks to disable/enable counters summarily,
via a prctl:

	PR_TASK_PERF_COUNTERS_DISABLE		31
	PR_TASK_PERF_COUNTERS_ENABLE		32
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: add new perf-counter type

The 'task clock' counter counts the amount of time a task is executing,
in nanoseconds. It stops ticking when a task is scheduled out either due
to it blocking, sleeping or it being preempted.

This counter type is a Linux kernel based abstraction, it is available
even if the hardware does not support native hardware performance counters.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: cleanup

Rename them to better match up the usual IRQ disable/enable APIs:

 hw_perf_disable_all()  => hw_perf_save_disable()
 hw_perf_restore_ctrl() => hw_perf_restore()
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: add new perf-counter type

The 'CPU clock' counter counts the amount of CPU clock time that is
elapsing, in nanoseconds. (regardless of how much of it the task is
spending on a CPU executing)

This counter type is a Linux kernel based abstraction, it is available
even if the hardware does not support native hardware performance counters.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: restructure code, introduce hw_ops driver abstraction

Introduce this abstraction to handle counter details:

 struct hw_perf_counter_ops {
	void (*hw_perf_counter_enable)	(struct perf_counter *counter);
	void (*hw_perf_counter_disable)	(struct perf_counter *counter);
	void (*hw_perf_counter_read)	(struct perf_counter *counter);
 };

This will be useful to support assymetric hw details, and it will also
be useful to implement "software counters". (Counters that count kernel
managed sw events such as pagefaults, context-switches, wall-clock time
or task-local time.)
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: add group counters

This patch adds the "counter groups" abstraction.

Groups of counters behave much like normal 'single' counters, with a
few semantic and behavioral extensions on top of that.

A counter group is created by creating a new counter with the open()
syscall's group-leader group_fd file descriptor parameter pointing
to another, already existing counter.

Groups of counters are scheduled in and out in one atomic group, and
they are also roundrobin-scheduled atomically.

Counters that are member of a group can also record events with an
(atomic) extended timestamp that extends to all members of the group,
if the record type is set to PERF_RECORD_GROUP.
Signed-off-by: NIngo Molnar <mingo@elte.hu>