When pai records data it may fail in select() system call.
This error is not reported and the pai program exits with success.
Change this and exit with proper exit code.
Fixes: d7b1cbad8b ("cpumf/pai: Add Processor Activity Instrumentation tool")
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Running command pai -r <file> might encounter entries of type
PERF_RECORD_SWITCH. Entries of that type are currently not handled
and cause an error message:
# pai -r ~/paicrypto.000
unknown header-type 14 unknown header-type 14 \
unknown header-type 14 ...
The error message is not terminated by a newline.
Handle entries of this type PERF_RECORD_SWITCH. These records
do not carry any payload at all, just a bit is set in the
header::misc member. This bit set determines context switch out.
Output after:
# ./pai -r ~/paicrypto.000
0x4b814018f4f3 6 cs-out
0x4b817bc3c936 6 cs-in
0x4b817bc5246c 6 cs-out
0x4b817bd90e9a 6 cs-in
....
#
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Pai collects data from per CPU ring buffers and stores them in the
memory mapped output file. When data is collected from many CPUs at
the same time, writing data to output file can be slow.
Improve this and allow the pai recording to run with higher
real time priority. This is the same approach as done by the perf tool.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Acked-by: Sumanth Korikkar <sumanthk@linux.ibm.com>
Signed-off-by: Steffen Eiden <seiden@linux.ibm.com>
pai dumps core when invoked with summary flag -S and
certain painnpa.XXXXX data files:
# ~/s390-tools/cpumf/pai -r -S
Segmentation fault (core dumped)
#
Fix this and check the existence of event raw data before
reading it. This avoid accessing NULL pointer trying to read raw
data when no counter values have been appended as raw data by the
kernel.
Fixes: dc73c77d73 ("cpumf/pai: Omit detailed event information when summary flag selected")
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Use PMU named pai_ext to extract the PMU type number required
for the installation of PAI NNPA counters using perf_event_open()
system call.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
When the summary flag is selected do not print detailed information
on every single event. Just print the summary information on each
non-zero counter value.
Suggested-by: Christian Borntraeger <borntraeger@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
During recording, progress information is displayed on the screen
on how many event have been processed and at which offset these
events have been extracted from the ring buffer.
This information is usually of no interest to the user. Omit this
data and only print them when requested via verbose flag.
Suggested-by: Christian Borntraeger <borntraeger@linux.ibm.com>
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Use util_file_read_va() function to read out the event number
given the event file. Also use util_strdup() function to save
the event name. This function terminates on Out Of Memory condition.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Jan Höppner <hoeppner@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
The Processor Activity Instrumentation facility (PAI) Extension 1
adds support for the counter set for Neural Network Processing Assist
(NNPA) counters.
NNPA counter values are appended as raw data to the data report.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Jan Höppner <hoeppner@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
The Processor Activity Instrumentation facility (PAI) uses the
perf_event_open system call to report data and the crypto counter
values. The data can be inspected with the perf tool in a very
generic way using raw data report and a hex dump utility.
The pai tool also uses the perf_event_open system call to record the
data, but uses specialized event attributes for recording detailed
information such as the recording of fork, exit, exec system call
events and context switch events.
Cryptography counter values are appended as raw data attachments.
The pai command can be used for recording and reporting.
Also a list of CPUs can be specified.
The recording ring buffers are created per CPU
and are named paicrypto.XXX where XXX stands for the CPU number.
The numbering scheme has leading zeros, such as 007 for CPU number 7.
For invocation and possible command line options see the man page.
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Jan Höppner <hoeppner@linux.ibm.com>
Signed-off-by: Jan Höppner <hoeppner@linux.ibm.com>
Thomas Richter