diff options
Diffstat (limited to 'debug.txt')
| -rw-r--r-- | debug.txt | 179 |
1 files changed, 132 insertions, 47 deletions
@@ -1,75 +1,160 @@ .:: debugging ::. -------------------------------------------------------------------------------- -# print syscalls of process with all threads (-f) of running process -strace -f -p <pid> -# only trace certain syscalls +# +# strace(1) +# + +strace [OPTS] [ELF] + -f .......... follow child processes on fork(2) + -p <pid> .... attach to running process + -s <size> ... max string size (default: 32) + -e <expr> ... expression for trace filtering + -o <file> ... log output into <file> + +useful <expr>: + trace=syscall[,syscall] .... trace only syscall listed + trace=file ................. trace all syscall that take a filename as arg + trace=process .............. trace process management related syscalls + signal ..................... trace signals delivered to the process + +use cases: +- trace 'open & socket' syscalls for a running process + childs strace -f -p <pid> -e trace=open,socket -# trace signals delivered to process + +- trace signals delivered to a running process strace -f -p <pid> -e signal -# dump stack of process and all threads -pstack <pid> -# print file flags -# +fg print file flag abbreviations +# +# lsof(8) +# + +lsof +fg -p <pid> + -p <pid> ... list open files for process + +fg ........ show file flags for file descripros + +file flags: + R/W/RW ..... read/write/read-write + CR ......... create + AP ......... append + TR ......... truncate + +uase cases: +- show open files with file flags lsof +fg -p <pid> -# print process virt mem map -# compared to /proc/<>/maps it shows the size of the mappings + +# +# pmap(1) +# + pmap <pid> + ............. dump virtual memory map of process. + compared to /proc/<pid>/maps it shows the size of the mappings + + +# +# pstack(1) +# + +pstack <pid> + ............. dump current stack of process + threads + + +# +# pidstat(1) +# # trace minor/major page faults pidstat -r -p <pid> [interval] + minor_pagefault: happens when the page needed is already in memory but not + allocated to the faulting process, in that case the kernel + only has to create a new page-table entry pointing to the + shared physical page + major_pagefault: happends when the page needed is NOT in memory, the kernel + has to create a new page-table entry and populate the + physical page + + +# +# /usr/bin/time(1) +# # statistics of process run /usr/bin/time -v <cmd> +# +# perf(1) +# + # get supported events perf list - -## show process stats -perf stat -p <pid> -perf stat -p <pid> -I <ms> -# select events of interests -perf stat -p <pid> -e cycles,faults,cache-misses,context-switches - -# -K hide kernel threads -perf top -F 99 -p <pid> -K - -## record measurement into db -# -g uses 'fb' = frame-pointer as default method to collect backtraces -# tell gcc to not optimize fp out -fno-omit-frame-pointer -perf record -F 99 -p <pid> -g -# use post processing and dwarf to collect backtraces -perf record -F 99 -p <pid> --call-graph dwarf -perf record -F 99 -p <pid> -e instructions,cpu-cycles,faults,cache-misses,context-switches - -## report perf db -# print to stdio -perf report -n --stdio -# use tui -perf report -g graph,0.5,caller - -# flamegraph (https://github.com/brendangregg/FlameGraph) -# record events of interest, eg page-faults -perf record -e page-faults --call-graph dwarf -perf script | FlameGraph/stackcollapse-perf.pl > out.perf-folded -FlameGraph/flamegraph.pl out.perf-folded > perf-trace.svg - -# record multiple events and generate flamegraphs -perf record --call-graph dwarf -p $(pgrep -u jstolp -n sim) -F 200 -e cpu-cycles,page-faults + ......... show supported hw/sw events + +perf stat + -p <pid> .. show stats for running process + -I <ms> ... show stats periodically over interval <ms> + -e <ev> ... filter for events + +perf top + -p <pid> .. show stats for running process + -F <hz> ... sampling frequency + -K ........ hide kernel threads + +perf record + -p <pid> ............... record stats for running process + -F <hz> ................ sampling frequency + --call-graph <method> .. [fp, dwarf, lbr] method how to caputre backtrace + fp : use frame-pointer, need -fno-omit-frame-pointer + dwarf: use .cfi debug information + lbr : use hardware last branch record facility + -g ..................... short-hand for --call-graph fp + -e <ev> ................ filter for events + +perf report + -n .................... annotate symbols with nr of samples + --stdio ............... report to stdio, if not presen tui mode + -g graph,0.5,caller ... show caller based call chains with value >0.5 + +useful <ev>: + page-faults + minor-faults + major-faults + cpu-cycles` + task-clock + + +# +# flamegraph(https://github.com/brendangregg/FlameGraph) +# + +# flamegraph for single event trace +perf record -g -p <pid> -e cpu-cycles +perf script | FlameGraph/stackcollapse-perf.pl | FlameGraph/flamegraph.pl > cycles-flamegraph.svg + +# flamegraphs for multiple events trace +perf record -g -p <pid> -e cpu-cycles,page-faults perf script --per-event-dump # fold & generate as above +# -# print supported events -ophelp +# +# OProfile +# -operf -p <pid> +operf -g -p <pid> + -g ...... caputre call-graph information -opreport -c -opreport -l +opreport [opt] FILE + NOOPT ... show time spent per binary image + -l ...... show time spent per symbol + -c ...... show callgraph information (see below) + -a ...... add column with time spent accumulated over child nodes + +ophelp + NOOPT ... show supported hw/sw events -------------------------------------------------------------------------------- vim:sts=2:et:tw=80:cc=80:fo+=t:ft=help |