tools: kvm_stat: Add comments

A lot of the code works with the perf events about which only sparse
documentation was available until 2012. Having that information now,
we can clarify what is done in the code.

Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This commit is contained in:
Janosch Frank 2016-05-18 13:26:25 +02:00 committed by Paolo Bonzini
parent f0cf040f84
commit fabc712866

View File

@ -10,6 +10,15 @@
#
# This work is licensed under the terms of the GNU GPL, version 2. See
# the COPYING file in the top-level directory.
"""The kvm_stat module outputs statistics about running KVM VMs
Three different ways of output formatting are available:
- as a top-like text ui
- in a key -> value format
- in an all keys, all values format
The data is sampled from the KVM's debugfs entries and its perf events.
"""
import curses
import sys
@ -217,8 +226,10 @@ IOCTL_NUMBERS = {
}
class Arch(object):
"""Class that encapsulates global architecture specific data like
syscall and ioctl numbers.
"""Encapsulates global architecture specific data.
Contains the performance event open syscall and ioctl numbers, as
well as the VM exit reasons for the architecture it runs on.
"""
@staticmethod
@ -306,12 +317,22 @@ def parse_int_list(list_string):
def get_online_cpus():
"""Returns a list of cpu id integers."""
with open('/sys/devices/system/cpu/online') as cpu_list:
cpu_string = cpu_list.readline()
return parse_int_list(cpu_string)
def get_filters():
"""Returns a dict of trace events, their filter ids and
the values that can be filtered.
Trace events can be filtered for special values by setting a
filter string via an ioctl. The string normally has the format
identifier==value. For each filter a new event will be created, to
be able to distinguish the events.
"""
filters = {}
filters['kvm_userspace_exit'] = ('reason', USERSPACE_EXIT_REASONS)
if ARCH.exit_reasons:
@ -322,6 +343,14 @@ libc = ctypes.CDLL('libc.so.6', use_errno=True)
syscall = libc.syscall
class perf_event_attr(ctypes.Structure):
"""Struct that holds the necessary data to set up a trace event.
For an extensive explanation see perf_event_open(2) and
include/uapi/linux/perf_event.h, struct perf_event_attr
All fields that are not initialized in the constructor are 0.
"""
_fields_ = [('type', ctypes.c_uint32),
('size', ctypes.c_uint32),
('config', ctypes.c_uint64),
@ -342,6 +371,20 @@ class perf_event_attr(ctypes.Structure):
self.read_format = PERF_FORMAT_GROUP
def perf_event_open(attr, pid, cpu, group_fd, flags):
"""Wrapper for the sys_perf_evt_open() syscall.
Used to set up performance events, returns a file descriptor or -1
on error.
Attributes are:
- syscall number
- struct perf_event_attr *
- pid or -1 to monitor all pids
- cpu number or -1 to monitor all cpus
- The file descriptor of the group leader or -1 to create a group.
- flags
"""
return syscall(ARCH.sc_perf_evt_open, ctypes.pointer(attr),
ctypes.c_int(pid), ctypes.c_int(cpu),
ctypes.c_int(group_fd), ctypes.c_long(flags))
@ -353,6 +396,8 @@ PATH_DEBUGFS_TRACING = '/sys/kernel/debug/tracing'
PATH_DEBUGFS_KVM = '/sys/kernel/debug/kvm'
class Group(object):
"""Represents a perf event group."""
def __init__(self):
self.events = []
@ -360,6 +405,22 @@ class Group(object):
self.events.append(event)
def read(self):
"""Returns a dict with 'event name: value' for all events in the
group.
Values are read by reading from the file descriptor of the
event that is the group leader. See perf_event_open(2) for
details.
Read format for the used event configuration is:
struct read_format {
u64 nr; /* The number of events */
struct {
u64 value; /* The value of the event */
} values[nr];
};
"""
length = 8 * (1 + len(self.events))
read_format = 'xxxxxxxx' + 'Q' * len(self.events)
return dict(zip([event.name for event in self.events],
@ -367,6 +428,7 @@ class Group(object):
os.read(self.events[0].fd, length))))
class Event(object):
"""Represents a performance event and manages its life cycle."""
def __init__(self, name, group, trace_cpu, trace_pid, trace_point,
trace_filter, trace_set='kvm'):
self.name = name
@ -375,10 +437,19 @@ class Event(object):
trace_filter, trace_set)
def __del__(self):
"""Closes the event's file descriptor.
As no python file object was created for the file descriptor,
python will not reference count the descriptor and will not
close it itself automatically, so we do it.
"""
if self.fd:
os.close(self.fd)
def setup_event_attribute(self, trace_set, trace_point):
"""Returns an initialized ctype perf_event_attr struct."""
id_path = os.path.join(PATH_DEBUGFS_TRACING, 'events', trace_set,
trace_point, 'id')
@ -388,9 +459,19 @@ class Event(object):
def setup_event(self, group, trace_cpu, trace_pid, trace_point,
trace_filter, trace_set):
"""Sets up the perf event in Linux.
Issues the syscall to register the event in the kernel and
then sets the optional filter.
"""
event_attr = self.setup_event_attribute(trace_set, trace_point)
# First event will be group leader.
group_leader = -1
# All others have to pass the leader's descriptor instead.
if group.events:
group_leader = group.events[0].fd
@ -408,15 +489,33 @@ class Event(object):
self.fd = fd
def enable(self):
"""Enables the trace event in the kernel.
Enabling the group leader makes reading counters from it and the
events under it possible.
"""
fcntl.ioctl(self.fd, ARCH.ioctl_numbers['ENABLE'], 0)
def disable(self):
"""Disables the trace event in the kernel.
Disabling the group leader makes reading all counters under it
impossible.
"""
fcntl.ioctl(self.fd, ARCH.ioctl_numbers['DISABLE'], 0)
def reset(self):
"""Resets the count of the trace event in the kernel."""
fcntl.ioctl(self.fd, ARCH.ioctl_numbers['RESET'], 0)
class TracepointProvider(object):
"""Data provider for the stats class.
Manages the events/groups from which it acquires its data.
"""
def __init__(self):
self.group_leaders = []
self.filters = get_filters()
@ -424,6 +523,20 @@ class TracepointProvider(object):
self._pid = 0
def get_available_fields(self):
"""Returns a list of available event's of format 'event name(filter
name)'.
All available events have directories under
/sys/kernel/debug/tracing/events/ which export information
about the specific event. Therefore, listing the dirs gives us
a list of all available events.
Some events like the vm exit reasons can be filtered for
specific values. To take account for that, the routine below
creates special fields with the following format:
event name(filter name)
"""
path = os.path.join(PATH_DEBUGFS_TRACING, 'events', 'kvm')
fields = walkdir(path)[1]
extra = []
@ -436,6 +549,8 @@ class TracepointProvider(object):
return fields
def setup_traces(self):
"""Creates all event and group objects needed to be able to retrieve
data."""
if self._pid > 0:
# Fetch list of all threads of the monitored pid, as qemu
# starts a thread for each vcpu.
@ -499,6 +614,7 @@ class TracepointProvider(object):
@fields.setter
def fields(self, fields):
"""Enables/disables the (un)wanted events"""
self._fields = fields
for group in self.group_leaders:
for index, event in enumerate(group.events):
@ -517,12 +633,16 @@ class TracepointProvider(object):
@pid.setter
def pid(self, pid):
"""Changes the monitored pid by setting new traces."""
self._pid = pid
# The garbage collector will get rid of all Event/Group
# objects and open files after removing the references.
self.group_leaders = []
self.setup_traces()
self.fields = self._fields
def read(self):
"""Returns 'event name: current value' for all enabled events."""
ret = defaultdict(int)
for group in self.group_leaders:
for name, val in group.read().iteritems():
@ -531,12 +651,19 @@ class TracepointProvider(object):
return ret
class DebugfsProvider(object):
"""Provides data from the files that KVM creates in the kvm debugfs
folder."""
def __init__(self):
self._fields = self.get_available_fields()
self._pid = 0
self.do_read = True
def get_available_fields(self):
""""Returns a list of available fields.
The fields are all available KVM debugfs files
"""
return walkdir(PATH_DEBUGFS_KVM)[2]
@property
@ -592,6 +719,12 @@ class DebugfsProvider(object):
return 0
class Stats(object):
"""Manages the data providers and the data they provide.
It is used to set filters on the provider's data and collect all
provider data.
"""
def __init__(self, providers, pid, fields=None):
self.providers = providers
self._pid_filter = pid
@ -601,6 +734,7 @@ class Stats(object):
self.update_provider_filters()
def update_provider_filters(self):
"""Propagates fields filters to providers."""
def wanted(key):
if not self._fields_filter:
return True
@ -615,6 +749,7 @@ class Stats(object):
provider.fields = provider_fields
def update_provider_pid(self):
"""Propagates pid filters to providers."""
for provider in self.providers:
provider.pid = self._pid_filter
@ -638,6 +773,8 @@ class Stats(object):
self.update_provider_pid()
def get(self):
"""Returns a dict with field -> (value, delta to last value) of all
provider data."""
for provider in self.providers:
new = provider.read()
for key in provider.fields:
@ -653,6 +790,7 @@ LABEL_WIDTH = 40
NUMBER_WIDTH = 10
class Tui(object):
"""Instruments curses to draw a nice text ui."""
def __init__(self, stats):
self.stats = stats
self.screen = None
@ -687,6 +825,7 @@ class Tui(object):
curses.endwin()
def update_drilldown(self):
"""Sets or removes a filter that only allows fields without braces."""
if not self.stats.fields_filter:
self.stats.fields_filter = r'^[^\(]*$'
@ -694,9 +833,11 @@ class Tui(object):
self.stats.fields_filter = None
def update_pid(self, pid):
"""Propagates pid selection to stats object."""
self.stats.pid_filter = pid
def refresh(self, sleeptime):
"""Refreshes on-screen data."""
self.screen.erase()
if self.stats.pid_filter > 0:
self.screen.addstr(0, 0, 'kvm statistics - pid {0}'
@ -734,6 +875,11 @@ class Tui(object):
self.screen.refresh()
def show_filter_selection(self):
"""Draws filter selection mask.
Asks for a valid regex and sets the fields filter accordingly.
"""
while True:
self.screen.erase()
self.screen.addstr(0, 0,
@ -756,6 +902,11 @@ class Tui(object):
continue
def show_vm_selection(self):
"""Draws PID selection mask.
Asks for a pid until a valid pid or 0 has been entered.
"""
while True:
self.screen.erase()
self.screen.addstr(0, 0,
@ -787,6 +938,7 @@ class Tui(object):
continue
def show_stats(self):
"""Refreshes the screen and processes user input."""
sleeptime = 0.25
while True:
self.refresh(sleeptime)
@ -809,6 +961,7 @@ class Tui(object):
continue
def batch(stats):
"""Prints statistics in a key, value format."""
s = stats.get()
time.sleep(1)
s = stats.get()
@ -817,6 +970,7 @@ def batch(stats):
print '%-42s%10d%10d' % (key, values[0], values[1])
def log(stats):
"""Prints statistics as reiterating key block, multiple value blocks."""
keys = sorted(stats.get().iterkeys())
def banner():
for k in keys:
@ -837,6 +991,7 @@ def log(stats):
line += 1
def get_options():
"""Returns processed program arguments."""
description_text = """
This script displays various statistics about VMs running under KVM.
The statistics are gathered from the KVM debugfs entries and / or the
@ -906,6 +1061,7 @@ Requirements:
return options
def get_providers(options):
"""Returns a list of data providers depending on the passed options."""
providers = []
if options.tracepoints:
@ -918,6 +1074,7 @@ def get_providers(options):
return providers
def check_access(options):
"""Exits if the current user can't access all needed directories."""
if not os.path.exists('/sys/kernel/debug'):
sys.stderr.write('Please enable CONFIG_DEBUG_FS in your kernel.')
sys.exit(1)