|
From: Frederic W. <fwe...@gm...> - 2009-10-03 01:25:47
|
On Fri, Oct 02, 2009 at 05:48:34PM -0400, Masami Hiramatsu wrote: > Hi, > > These patches introduce 'perf probe' command and update kprobe-tracer. > perf probe command allows you to add new probe points by C line number > and local variable names. > > This version fixes some bugs, changes subcommand name from kprobe to > probe and use spaces for separator instead of ',' for visibility (this > also make it easy to support probe list from stdin). > > Usage > ----- > perf probe [<options>] -P 'PROBEDEF' [-P 'PROBEDEF' ...] > > -k, --vmlinux <file> vmlinux/module pathname > -r, --release <rel> kernel release > -P, --probe <p|r:[GRP/]NAME FUNC[+OFFS][@SRC]|@SRC:LINE [ARG ...]> > probe point definition, where > p: kprobe probe > r: kretprobe probe > GRP: Group name (optional) > NAME: Event name > FUNC: Function name > OFFS: Offset from function entry (in byte) > SRC: Source code path > LINE: Line number > ARG: Probe argument (local variable name or > kprobe-tracer argument format is supported.) > > Examples > -------- > 1) Add a new kprobe probe on a line of C source code. > ./perf probe -P 'p:myprobe @fs/read_write.c:285 file buf count' > Adding new event: p:myprobe vfs_read+57 file=%bx buf=%si count=%ax Nice! Great thing. One neat, at a first glance, file=%bx buf=%si look like a format definition. How about using file=bx ? Or does that introduce any ambiguities with kprobes definitions syntax? > > 2) Add a new kretprobe probe on a function return. > ./perf probe -P 'r:myretprobe vfs_read $rv' > Adding new event: r:myretprobe vfs_read+0 $rv The '$' character may perhaps also confuse bash scripts that create perf probe. > 3) Check it in the perf list. > ./perf list > ... > rNNN [raw hardware event descriptor] > > kprobes:myprobe [Tracepoint event] > kprobes:myretprobe [Tracepoint event] > skb:kfree_skb [Tracepoint event] > ... > > 4) Record the event by perf > ./perf record -f -e kprobes:myprobe:record -F 1 -a ls > ... > [ perf record: Woken up 1 times to write data ] > [ perf record: Captured and wrote 0.081 MB perf.data (~3540 samples) ] > > 5) Trace the event > ./perf trace > perf-11445 [000] 95862.048894383: myprobe: (c04bbed5) file=dae15e80 buf=b78b2000 count=400 > perf-11445 [000] 95862.049066533: myprobe: (c04bbed5) file=dae15d80 buf=b78b2000 count=400 > perf-11445 [000] 95862.049134394: myprobe: (c04bbed5) file=dae15d80 buf=b78b2000 count=400 > perf-11445 [000] 95862.049171495: myprobe: (c04bbed5) file=dae15a80 buf=b78b2000 count=400 > > NOTE > ---- > perf still fails to parse format if arguments have special charactors > (e.g. $rv, +10($sp) etc.) So, tracing myretprobe will fail with this > version. This will be solved by naming arguments automatically if it > doesn't have C-language name. > > TODO > ---- > - Support sys_perf_counter_open (non-root) Hmm, we really want it to be only usable by the root (except if the policy is tuned to allow that, but that's managed by perf already). > - Input from stdin/output to stdout > - Non-auto static variable > - Fields of data structures (var->field) > - Type support > - Bit fields > - Array (var[N]) > - Dynamic array indexing (var[var2]) > - String/dynamic arrays (var:string, var[N..M]) > - Force Type casting ((type)var) > - Non-inline search > - libdw, libdwfl > - etc. > > Thank you, > Cool, I'm reviewing/testing it and if no rough problem arise I'll apply it. Thanks a lot! |