| Age | Commit message (Collapse) | Author |
|---|
|
Currently the a -ve failure value is being handled as an error
for the number of CPUs. Also check for zero cpus as this should
be considered as invalid. This also stops a zero sized calloc
based on zero cpus being reported as an error by static analysis.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
|
|
Signed-off-by: Lisa Nguyen <lisa.nguyen@linaro.org>
|
|
Coverity Scan detected a resource leak in build_cstate_info() and
output_cstate_info() functions. Refer to defect #84131.
Valgrind confirms the memory leak as well:
==41915== 4,768 bytes in 1 blocks are definitely lost in loss record 6 of 8
==41915== at 0x4C2F948: calloc (vg_replace_malloc.c:711)
==41915== by 0x40232B: build_cstate_info (idlestat.c:318)
==41915== by 0x403DD1: output_cstate_info (idlestat.c:937)
==41915== by 0x404D53: idlestat_store (idlestat.c:1286)
==41915== by 0x4055A8: main (idlestat.c:1532)
Fix this issue by rewritting the syntax for the calloc function inside
build_cstate_info(), so the pointer can be easily deferenced later. Also,
deallocate any memory that was created locally inside output_cstate_info().
Signed-off-by: Lisa Nguyen <lisa.nguyen@linaro.org>
|
|
Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
|
|
cppcheck detected a potential buffer overflow with a %s scanf width format
specifier being larger than the destination buffer. Reduce the scan size to
match the expected buffer size.
Fixes error:
[idlestat.c:974]: (error) Width 32 given in format string (no. 2) is larger
than destination buffer 'irqname[17]', use %16s to prevent overflowing it.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
|
|
fpath is not being free'd causing a small memory leak, free it.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
|
|
clang's scan-build detected the following error "Undefined or
garbage value returned to caller" because ret is not intialised.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Amit Kucheria <amit.kucheria@linaro.org>
|
|
The function merge_pstates did not work as advertised. This patch fixes
the implementation.
Signed-off-by: Tuukka Tikkanen <idlestat@tic0.net>
Signed-off-by: Amit Kucheria <amit.kucheria@linaro.org>
|
|
Several parts of idlestat code did not cope with some cpu(s)
being offline. This patch adds information to cpu topology
data structure to record online status. Online-sensitive
parts of the code may query the status using is_cpu_online()
and skip performing activities related to offline cpus.
The following activities are skipped of cpu is offline:
- Trace: Obtaining initial frequency
- Trace: Storing initial frequency
- Trace: Adding pseudo pstate transition 0 Hz to end of trace
- Trace: Storing cpu C-state information
- Tracefile_idlestat: Reading cpu C-state information
- Topology: Mapping "datas" to "topology"
Signed-off-by: Tuukka Tikkanen <idlestat@tic0.net>
Signed-off-by: Amit Kucheria <amit.kucheria@linaro.org>
|
|
Make it possible to compile with either AOSP and NDK without
warnings. For AOSP, putting the idlestat to AOSP_ROOT/external
works fine. For NDK, the following command works fine.
ndk-build NDK_PROJECT_PATH=`pwd` \
APP_BUILD_SCRIPT=`pwd`/Android.mk \
APP_PLATFORM=19 APP_ABI=armeabi-v7a
1. Most code modifications to eliminate warnings
2. Both AOSP and NDK use -Wl,--gc-sections, which will remove code
compiled from tracefile_*.c and *_report.c, so we need
-Wl,--no-gc-sections
3. Android basename() is declared in <libgen.h>
4. Android doesn't have old <values.h>, use <floats.h> instead
Signed-off-by: Koan-Sin Tan <freedom.tan@linaro.org>
Reviewed-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Android complation checks that all parameters are used. Add
check for this condition to default compilation as well. For
cases where the parameter is legitimately not used, add a
macro UNUSED to flag such cases for the compiler. This macro
is defined in a new header file. Currently this macro is
defined functional if the preprocessor macro __GNUC__ is
defined, otherwise the macro expands to an empty comment.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
load_text_data_line() has been split out of the original trace
file loading loop. The original code incremented event count
at the points of code that had successfully completed importing
a single event. As the new function additionally now returns
a return code indicating success, this incrementation can be
done at the call site. This removes the need to pass a pointer
to the count variable and allows further cleanup of the code.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The parameter cpu is not used in write_cstate_info.
Note: The function is intended as an abstraction for writing
cstate information to trace file. As the trace file format
may change in future or multiple output formats may be
forthcoming, this function is an abstraction for storing
the state. Some future format may require the cpu number,
so this parameter may be reintroduced later. However, as
we currently do not know whether it is needed or not, it
can be removed.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Function calculate_energy_model (no longer) uses the program
options structure. It may have been used earlier for the
verbosity control in some messages, but that system has been
reworked. Remove the unused paramter.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
alloc_pstate() searched the array of existing frequencies as if
the frequencies were in descending order (by frequency) but
then inserted the new frequency as if the frequencies were
to be sorted into ascending order. The search order was changed
to ascending order.
cluster_get_highest_freq() actually returned the lowest active
frequency in the cluster. core_get_highest_freq() acted in the
similar way. These have been fixed.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Idlestat adds phony frequency entries to the end of trace file
to ensure all the last running state is accounted for. This
phony frequency was unsigned int "-1" (aka 2^32-1 or 2^64-1),
which was then listed in any energy model template generated
from the trace file.
This patch changes the phony frequency to 0 and makes the
energy model template generation ignore 0 frequency.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The P-state (frequency) for cores and clusters was not calculated.
This patch adds recording of core/cluster frequency as the
highest frequency of all non-idle cpus contained within the core/
cluster.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
As the kernel trace buffer is a finite resource, the default
trace buffer size idlestat sets may either be impossibly
large or insufficient to hold the data. This patch adds two
new command line options to control buffer size and whether
or not it will be polled during the trace period.
The polling functionality while tracing is not yet implemented.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Calls to core_get_least_cstate and cluster_get_least_cstate
were not included in the patch that intended to add them. This
caused all core/cluster c-states to be recorded as if the
core/cluster entered exactly the same state as what was
entered by any cpu in the core/cluster that changed its state.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The return value from call to fscanf was previously ignored.
Now the function cpuidle_get_target_residency will return -1
if fscanf fails to parse the value.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The report_ops and trace_ops arrays are created by the linker.
Static analysis tool has no idea of this and make a false positive
detection for the loops that access the arrays. This patch
introduces comments that suppress the detection for the
lines in question.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Signed-off-by: Amit Kucheria <amit.kucheria@linaro.org>
Reviewed-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Makes it easier to add/remove options and keep them sync'ed.
Signed-off-by: Amit Kucheria <amit.kucheria@linaro.org>
Reviewed-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
It should've been part of dd89f9215bf7c2e345716eb09dfa975f1f1d15dc "idlestat:
Add support for baseline datas".
While we're at it, remove the '-d' option which is not used anymore.
Signed-off-by: Amit Kucheria <amit.kucheria@linaro.org>
Reviewed-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Many of the source files were missing copyright headers. Headers have been
added to all .c and .h files and existing headers were updates with
contributor information.
The version number was bumped up to 0.5 due to having many new features
since last version number update.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The functions used to calculate core/cluster c-state statistics from
individual cpu statistics are no longer used and can be removed.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Current implementation for core/cluster data calculation is to look
for intersections of same state residency in cpu data after loading
all of the data. This fails to correctly record core/cluster data
for cases where all the contained cpus are in a low power state, but
are not using the exact same state.
To resolve this, core/cluster state needs to be evaluated at the
same time as the cpu events are being processes. The store_data()
function is modified to record not only the affected cpu changes,
but also the affected core and cluster. Supporting changes are added
throughout the source tree.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The function store_data() does not use the parameter count for any
purpose. This patch removes the parameter.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The function cluster_data() has no callers, remove it.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Idlestat modifies the kernel trace options via sysfs interface every
time a trace is created. The user may have prefer to have the original
trace options restored once idlestat is finished with the task. This
patch adds code to save and restore the trace options that might be
modified by idlestat.
idlestore_store_trace_options(): Saves the options by 1) record the
original trace buffer size and 2) scan the whole ftrace event
hierarchy, storing the paths of enabled event types into a list.
idlestat_restore_trace_options(): Restores the options including
buffer size and enabled events. Disables all other events.
Signed-off-by: Koan-Sin Tan <freedom.tan@linaro.org>
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
As pstates are allocated in the pstate arrays only when seen in
traces and each trace has its own private arrays, main trace
and baseline may end up having different sets of pstates in
their arrays. This patch iterates through the pstate arrays
and inserts pstate entries into the arrays so that both traces
have the same pstates in same order. This allows easy comparison
between the two traces.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
While passing around the per report data pointer has been previously
implemented, there was no mechanism for the report ops to allocate
or release the data as needed. This patch moves allocation of
report_data pointer from struct options to main() and adds calls
necessary for allocating (and initializing) and releasing the data.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The functions display_cstates and display_pstates are in control of
dumping c/p state information for each cpu. This patch adds handling
of baseline states to both functions. If the selected report mode
does not provide baseline handling function(s), the operation is
silently skipped.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
pstate_single_state and wakeup_single_state have had most of
their names copied from the cstate variant. As the items
displayed by these functions are frequencies and interrupt
sources, rename the functions to pstate_single_freq and
wakeup_single_irq.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
This patch adds relay of baseline data for per-cpu entries when
composing the report.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The existing version of alloc_pstate() can create duplicate
entries for the same pstate if called with a frequency
value that already exists. Change the semantics so that
in case the frequency already exists in the array, the
index returned is for the existing frequency.
This change allows removal of (now) obsolete function
get_current_pstate and merging of pstate entries between
main trace and baseline trace for comparison report purposes.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
This patch adds command line option for specifying a baseline
trace file. If specified, the trace is loaded using idlestat_load
and the baseline member in the main datas structure points to
the loaded baseline datas structure. As such, this adds no new
features itself, but allows implementing a comparison feature.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The initial implementation of default, boxless and csv report formats
required making the report format ops structures global. Getoptions()
referred directly to these structures when assigning the active ops
pointer.
This patch adds a name field to the structures and places pointers
to the structures in a special segment so that the list of ops
structures can be iterated like an array without having to create
such an array explicitly. Getoptions() is modified to record only
the name and main() will then obtain the actual structure pointer
by name.
The ops array head and sentinel is combined into the same source
files that have been used for trace file ops arrays. The files
are renamed to ops_head.c and ops_tail.c.
Command line options parsing now checks that there is only one
specification of report format. Usage function lists the available
formats. -C and -B are retained as shortcuts for -r csv and
-r boxless, respectively.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Some reports may require some program options to be set, but other
reports may not require the same options. To keep details of these
checks out of main() and getoptions(), this patch adds a new
operation to report_ops for doing these checks when necessary.
Pointer to struct program_options is passed to the op. If the
options are not suitable for operation, the op must print an error
message for the user and return -1. Otherwise the op should
return 0.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
print_vrb (print verbose) is a preprocessor macro defined in utils.h.
It behaves like fprintf(stderr, ...) if the level argument is greater
than or equal to current options->verbose value (incremented by one
for each -v parameter on command line). Unfortunately the macro does
depend on variable "options" being visible in the scope it is used in,
which prevents using it in any function that does not have that
variable declared in it or passed to it.
This patch makes the macro a real function and disconnects the
dependency of having options visible at call site. Instead while the
options are passed, a call to set_verbose_level() is make whenever
the value of options->verbose changes.
The name and call semantics are adjusted to increase versatility of
the function. Two variants are created:
verbose_printf(int min_level, const char *fmt, ...)
verbose_fprintf(FILE *f, int min_level, const char *fmt, ...)
These two behave like printf and fprintf if level >= current verbose
level, otherwise they do nothing and return 0.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
This patch separates loading of different trace file formats to
independent functions. In order to call the correct function for
each trace file processed, each load function comes with a match
function. Pointers to these functions are stored in a new
struct trace_ops.
In order to avoid static lists of trace file formats (e.g.
implemented as static arrays or a number of registration
function calls in main), pointers to all trace_ops structures
are collected by the linker into a private segment. This creates
an implicit array of the pointers, which may then be enumerated
starting right after a globally linked element trace_ops_head.
A sentinel NULL pointer trails the array. A helper macro
EXPORT_TRACE_OPS is added for creating the special pointers.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
Reviewed-by: Koan-Sin Tan <freedom.tan@linaro.org>
|
|
Pre-patch the topology is rooted to a global variable g_cpu_topo_list.
This prevents loading multiple traces for comparison. This patch
makes the allocation dynamic. The root pointer is stored within
struct cpuidle_datas, allowing the idlestat_load() to retain its
prototype, but most of the functions defined in topology.c have
a new parameter for passing the topology root pointer.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
Reviewed-by: Koan-Sin Tan <freedom.tan@linaro.org>
|
|
Return values from programs are unsigned. Replace return -1 with
return 1 in main().
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Function inter() may return newly allocated structures or one of the
parameters (if the other parameter is NULL). This function is called
in a loop from cluster_data(), core_cluster_data() and
physical_cluster_data() with one of the parameters often being an
earlier return value from inter(). This causes a great number of
structure allocations without matching memory releases.
Code in topology.c takes the result from *_cluster_data(), but it
only releases the top-level structure, without releasing the
data pointed to by top-level structure members.
This patch plugs both types of leaks. Since the memory pointed to
by inter() may or may not be allocated by that call, a new member
has been added to struct cpuidle_cstate to indicate the origin
of the structure.
This patch also adds detection for data aliasing related to inter(),
but does not remove the bug.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
This patch changes the return value of inter() and intersection()
on allocation fail to ptrerror(). Previously the return value was
NULL, which under normal conditions indicates no overlap between
the two sets. Further cascading changes are made to callers of
these functions and the callers of callers etc.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
This patch modifies malloc/calloc/realloc calls in idlestat.c to
conform to following rules:
1. Newly allocated memory is zeroed by using calloc instead of malloc
or by calling memset after realloc. The only exception is when the newly
allocated memory is storage for a string or always completely overwritten
with a call to memcpy or a single assignment statement.
2. Calls to memset(ptr, 0, size) that followed calloc have been removed.
3. perror() message body after allocation failure is uniformly __func__,
resulting in "func_name: Out of memory".
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The element count has already been incremented by the time of
calling realloc in inter(), so adding +1 to the element
count is unnecessary.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
Some functions such as inter() and intersection() use NULL return
value to indicate both errors (e.g. out of memory) and a result
(empty intersection). Make ptrerror() return a unique non-NULL
pointer value so that the error and NULL result can be told apart.
Previously the parameter to ptrerror() was unconditionally
interpreted as a pointer to string to be passed to perror(). Now
the function checks the value and calls perror only if the
parameter value is not NULL.
A new function, is_err(), has been added to check if any pointer
value is equal to ptrerror() return value.
error(), ptrerror() and is_err() have been moved to utils.c.
The functions that used ptrerror() before this patch have been
modified to consistently return ptrerror() instead of NULL in
cases of errors (affects the return statements where no call to
perror() is desired). Likewise the callers now use is_err()
instead of checking for NULL pointer.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
|
|
The function intersection() uses pointer parameters without checking
the validity. Add assert() statements making sure the pointers are
not NULL.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
|
|
The initial and final pstates were stored only if the display
options at the time of creating the trace file included display
of P-states. Since the trace may later be imported with the
P-state display on, add the pseudo P-state changes to the trace
unconditionally.
In addition, the return value of get_trace_ts was never checked
for error. The error checking has been added. Since floating
point comparison for equality is unreliable, the function now
returns only status and the actual data is relayed by modifying
memory pointed to by a function parameter. Error messages have
been added to the get_trace_ts function.
Signed-off-by: Tuukka Tikkanen <tuukka.tikkanen@linaro.org>
Reviewed-by: Amit Kucheria <amit.kucheria@linaro.org>
|
