On kernels build without CONFIG_TRACING kvm_stat will bail out even
when traces are not used. This is not very helpful, especially if the
user can't install a new kernel. Instead, we should warn the user and
fall back to debugfs statistics.
These changes check if trace statistics were selected without kernel
support, warn with a small timeout, set the debugfs statistics option
to True and the tracefs one to False.
Fixes: 7aa4ee5 ('scripts/kvm/kvm_stat: Improve debugfs access checking')
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1454485291-43849-2-git-send-email-frankja@linux.vnet.ibm.com>
[Exit if -t is passed explicitly. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Added a description text that explains what the script does and which
requirements have to be met to let it run.
The help formatter class is needed as the default optparse formatter
makes the text unreadable.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-35-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Interactively changing the filter is much more useful than the
drilldown, because it is more versatile.
With this patch, the filter can be changed by pressing 'f' in the text
ui and entering a new filter regex.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-34-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When filtering, the group leader event should not be disabled, as all
other events under it will also be disabled. Also we should make sure
that values from disabled fields will not be displayed.
This also filters the fields from the log and batch output for better
readability.
Also the drilldown update now directly checks for the stats' field
filter and (un)sets drilldown accordingly.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-33-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Setting the hard limit as a unprivileged user either returns an error
when it is higher than the current one or irreversibly sets it lower.
Therefore we leave the hardlimit untouched as long as we don't need to
raise it as this needs CAP_SYS_RESOURCE.
This gives admins the possibility to run the script as an unprivileged
user to increase security.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-32-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The struct read_format, which denotes the returned values on a read
states that the values are u64 and not long long which is used for
struct unpacking.
Therefore the 'q' long long formatter was exchanged with 'Q' which is
the format for u64 data.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-31-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
All initializations of the ctypes struct that don't need additional
information were moved to its init method. The unneeded
initializations for sample_type and sample_period were removed as they
do not affect the counters that are read.
This improves readability of the setup_event_attribute by halfing its
LOC.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-30-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The key names in log mode were capped to 10 characters which is not
enough for distinguishing between keys. Capping was therefore removed.
In batch mode the spacing between keys and values was too narrow and
therefore had to be extended to 42.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-29-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The tui function itself had a few sub-functions and therefore
basically already was class-like. Making it an actual one with proper
methods improved readability.
The curses wrapper was dropped in favour of __entry/exit__ methods
that implement the same behaviour.
Also renamed single character variable name, so the name reflects the
content.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-28-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The architecture detection method directly accesses vmx and smv exit
reason constants. Therefore we don't need it anymore.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-27-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Using global variables and multiple initialization functions for arch
specific data makes the code hard to read. By grouping them in the
Arch classes we encapsulate and initialize them in one place.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-26-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Added additional newlines for readability.
Factored out attribute and event setup code into own methods.
Exchanged file() with preferred open().
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-25-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Introduced separating newlines for readability and removed special
treatment/variable of the group leader. Renamed fmt to read_format.
The group leader's file descriptor will not be turned into a file
object anymore, instead os.read is used to read from the descriptor.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-24-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Converted class definition to new style and renamed improper named
variables.
Introduced property for fields_filter.
Moved member variable declaration to init, so one can see all class
variables when reading the init method.
Completely clear the values dict, as we don't need to keep single values.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-23-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The variable was only used in one class but still was defined
globally. Additionaly the detect_platform routine which prepares the
data that goes into the variable was called on each start of the
script, no matter if the class was needed.
To make the variable local to the TracepointProvider class, a new
function that calls detect_platform and returns the filters was
introduced.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-22-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reading /sys/devices/system/cpu/online makes opening the cpu
directories unnecessary and works on more/older systems.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-21-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Variables with bad names like f and m were renamed to their full name,
so it is clearer which data they contain.
Unneeded variables were removed and the field generating code was
moved in an own function.
dict.iteritems() was removed as directly iterating over a dictionary
also yields the needed keys.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-20-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
As previous commit authors used a mixture of setters/getters and
direct access to class variables consolidating them the python way
improved readability.
Properties allow us to assign a value to a class variable through a
setter without the need to call the setter ourselves.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-19-git-send-email-frankja@linux.vnet.ibm.com>
[prop.setter is new in Python 2.6, which is the earliest supported
version. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The underscore in front of the function name does not comply with the
python coding guidelines.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-18-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The online cpus detection method is in the Stats class but does not
use any class variables.
Moving it out of the class to the platform detection function makes
the Stats class more readable.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-17-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
s390 machines can also be detected via uname -m, i.e. python's
os.uname, no need for more complicated checks.
Calling uname once and saving its value for multiple checks is
perfectly sufficient. We don't expect the machine's architecture to
change when the script is running anyway.
On multi-cpu systems x86_init currently will get called multiple
times, returning makes sure we don't waste cicles on that.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-16-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
As num cpus * 1000 is NOT a sensible rlimit, we need to calculate a
more accurate rlimit.
The number of open files is directly dependent on the cpu count and on
the number of trace points per cpu. A additional constant works as a
buffer for files that are needed by python or do get opened when the
script runs.
Hence we have:
cpus * traces + constant
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-15-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In 2008 a patch was written that introduced ctypes.get_errno() and
set_errno() as official interfaces to the libc errno variable. Using
them we can avoid accessing private libc variables.
The patch was included in python 2.6.
Also we need to raise the right exception, with the right parameters
and a helpful message.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-14-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When it is next to the TracepointProvider less scrolling is needed to
change related, surrounding code.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-13-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Filter, id and byte are builtin python modules which should not be
redefined by local variables.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-12-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Keyword assignments should not not have spaces around the equal
character according to PEP8.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-11-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The main function should be the main location for initialization and
helps encapsulating variables into a scope. This way they don't have
to be global and might be mistaken for local ones.
As the providers variable is scoped now it can't be accessed from
within the Stats class. Hence, the global access to the variable was
changed to a local one.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-10-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Access checking with F_OK was replaced with the better readable
os.path.exists().
On Linux exists() returns False when the user doesn't have sufficient
permissions for statting the directory. Therefore the error message
now states that sufficient rights are needed when the check fails.
Also added check for /sys/kernel/debug/tracing/.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-9-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Paths to debugfs and trace dirs are now specified globally to remove
redundancies in the code.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-8-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The exit reasons dictionaries were defined number -> value but later
on were accessed the other way around. Therefore a invert function
inverted them.
Defining them the right way removes the need to invert them and
therefore also speeds up the script's setup process.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-7-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Updating globals over the globals().update() method is not the
standard way of changing globals. Marking variables as global and
modifying them the standard way is better readable.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-6-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Only two of the constants are actually needed to set up the events, so
the others were removed. All variables that used them were also removed.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-5-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Constants should be uppercase with separating underscores, as
requested in PEP8. This helps identifying them when reading the code.
Reviewed-by: Jason J. Herne <jjherne@linux.vnet.ibm.com>
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-4-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Os.walk gives back lists of directories and files, no need to filter
directories from the list that listdir gives back.
To make it better understandable a wrapper with docstring was
introduced.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-3-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Removed multiple imports of the same module and moved all imports to
the top.
It is not necessary to import a module each time one of its
functions/classes is used.
For readability each import should get its own line.
Signed-off-by: Janosch Frank <frankja@linux.vnet.ibm.com>
Message-Id: <1452525484-32309-2-git-send-email-frankja@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The curses user interface shows both the accumulated total and the
current event counts. Add column headers so it's clear what the numbers
mean.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Ademar Reis <areis@redhat.com>
Reviewed-by: Wei Huang <wei@redhat.com>
Message-Id: <1425338947-10296-2-git-send-email-stefanha@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This patch defines the list of kvm_exit reasons for aarch64. This list is
based on the Exception Class (EC) field of HSR register. With this patch
users can trace the execution of guest VMs better. A sample output from
command "kvm_stat -1 -t" is shown as the following:
<...>
kvm_exit(WATCHPT_HYP) 0 0
kvm_exit(WFI) 9422 9361
NOTE: This patch requires TRACE_EVENT(kvm_exit) to include exit_reason
field in TP_ARGS. A patch to upstream kernel has been submitted.
Signed-off-by: Wei Huang <wei@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
While running kvm_stat using tracepoint on ARM64 hardware (e.g. "kvm_stat
-1 -t"), the initial values of some kvm_userspace_exit counters were found
to be very suspecious. For instance the tracing tool showed that S390_TSCH
was called many times on ARM64 machine, which apparently was wrong.
This patch adds RESET ioctl support for perf monitoring. Before calling
ioctl to enable a perf event, this patch resets the counter first. With
this patch, the init counter values become correct on ARM64 hardware.
Example:
==== before patch ====
kvm_userspace_exit(S390_SIEIC) 1426 0
kvm_userspace_exit(S390_TSCH) 339 0
==== after patch ====
kvm_userspace_exit(S390_SIEIC) 0 0
kvm_userspace_exit(S390_TSCH) 0 0
Signed-off-by: Wei Huang <wei@redhat.com>
kvm_stat uses syscall() to call perf_event_open(). If this function
call fails, the returned value is -1, which doesn't tell the details
of such failure (i.e. ENOSYS or EINVAL). This patch retrieves errno
and prints it when syscall() fails. The error message will look like
"Exception: perf_event_open failed, errno = 38".
Signed-off-by: Wei Huang <wei@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This patch updates the exit reasons for x86_vmx, x86_svm, and userspace
to the latest definition.
Signed-off-by: Wei Huang <wei@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This patch enables aarch64 support for kvm_stat. The platform detection
is based on OS uname.
Signed-off-by: Wei Huang <wei@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add support for powerpc platforms. We use uname -m, which allows us to
detect ppc, ppc64 and ppc64le/el.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Unfortunately ioctl numbers are platform specific, so abstract them out
of the code so they can be overridden. As it happens x86 and s390 share
the same values, so nothing needs to change yet.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The current platform detection is a little bit messy. We look for lines
in /proc/cpuinfo starting with 'flags' OR 'vendor-id', and scan both
for values we know will only occur in one or the other. We also keep
scanning once we've found a value, which could be a feature, but isn't
in this case.
We'd also like to add another platform, powerpc, which will just make it
worse. So clean it up in preparation.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In kvm_stat we have a dictionary of exit reasons for s390. Firstly these
are not s390 specific, they are the generic exit reasons. So rename the
dictionary to reflect that, and add it separately to filters[].
Secondly, the values are defined using hex, but in the kernel header
they are decimal. That means values above 9 in kvm_stat are incorrect.
While we're there, fix the whitespace to match the rest of the file.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In kvm_stat we grovel through /sys to find out how many cpus are in the
system. However if a cpu is offline it will still be present in /sys,
and the perf_event_open() will fail.
Modify the logic to only return online cpus. We need to be careful on
systems which don't support cpu hotplug, the online file will not be
present at all.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The old stats contain information not available in the tracepoints.
By default, keep the old behavior, but allow choosing which set of stats
to present, or even both.
Inspired by a patch from Marcelo Tosatti.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Correct sys_perf_event_open syscall number for s390 architecture
- the hardcoded syscall number 298 is for x86 but should
be different for other architectures.
In case we figure out via /proc/cpuinfo that we are running
on s390 the appropriate syscall number is used from map
syscall_numbers; other architectures can extend this.
Signed-off-by: Heinz Graalfs <graalfs@linux.vnet.ibm.com>
Signed-off-by: Jens Freimann <jfrei@linux.vnet.ibm.com>
Signed-off-by: Alexander Graf <agraf@suse.de>
Add s390_exit_reasons so kvm_stat doesn't crash when called on s390.
Look for 'vendor_id' in /proc/cpuinfo as well, instead of just for
'flags', so we can determine if we run on S390.
Signed-off-by: Jens Freimann <jfrei@linux.vnet.ibm.com>
Signed-off-by: Alexander Graf <agraf@suse.de>