On Fri, 1 Dec 2017 13:40:39 -0300
Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
> Adding Masami to the CC list, to see if he can come up with a fix for
> the 'perf probe' failure below, improving the message or ignoring
> versioned symbols (with a warning perhaps).
Thanks. Should we ignore the versioned symbols? or can we put a probe
on that? If we have symbol address, we can put a probe on it, even
if it is versioned.
>
> Em Thu, Nov 30, 2017 at 12:04:23AM +0530, bhargavb escreveu:
> > Hello,
> >
> > I am trying to understand user space profiling using linux inbuilt
> > mechanism. (Trying with perf also but could not get complete clear
> > documentation and could not proceed, so retained linux-perf-users in
> > to list as the underlying interface used is same for both).
>
> So yo uwant to trace function entry/exit in some userspace library? You
> can try using the tracefs as below or leave it to perf to do that, first
> you set up the probes, say for some glibc functions:
>
> [root@jouet ~]# perf probe -F -x /lib64/libc-2.25.so | grep ^malloc
> malloc
> malloc_check
> malloc_consolidate.part.1
> malloc_get_state@GLIBC_2.2.5
> malloc_hook_ini
> malloc_info
> malloc_printerr
> malloc_set_state@GLIBC_2.2.5
> malloc_stats
> malloc_trim
> malloc_usable_size
> mallochook
> [root@jouet ~]#
>
> [root@jouet ~]# perf probe -x /lib64/libc-2.25.so malloc*
> Failed to write event: Invalid argument
> Error: Failed to add events.
> [root@jouet ~]#
OK, this is not good, at least it should analyse the reason(versioned)
and warn that (Or, I think we can just remove the suffix after "@").
>
> For now ignore that error, its a bug I just found, what matters is that
> it _has_ put in place probes for most of those functions, as you can see
> by looking at:
>
> [root@jouet ~]# perf probe -l
> probe_libc:malloc (on __malloc in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_check (on malloc_check in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_consolidate (on malloc_consolidate.part.1 in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_hook_ini (on malloc_hook_ini in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_info (on __malloc_info in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_printerr (on malloc_printerr in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_stats (on __malloc_stats in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_trim (on __malloc_trim in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_usable_size (on __malloc_usable_size in /usr/lib64/libc-2.25.so)
> probe_libc:mallochook (on mallochook in /usr/lib64/libc-2.25.so)
> [root@jouet ~]#
>
> Masami, this one also doesn't work:
>
> [root@jouet ~]# perf probe -x /lib64/libc-2.25.so 'malloc*%return'
> Error: event "malloc_printerr" already exists.
> <SNIP>
Hm, this looks like a collision happens on that event name. "-f"(--force)
option allows you to add events with number suffix (e.g. _1, _2, ... )
>
> doing it a bit differently:
>
> [root@jouet ~]# perf probe -F -x /lib64/libc-2.25.so | egrep ^malloc[^@]+$ | while read function ; do perf probe -x /lib64/libc-2.25.so "${function}_return=$function%return" ; done
> Added new event:
> probe_libc:malloc_check_return (on malloc_check%return in /usr/lib64/libc-2.25.so)
>
> You can now use it in all perf tools, such as:
>
> perf record -e probe_libc:malloc_check_return -aR sleep 1
Hmm, would you think we should add __return suffix by default?
Thanks,
>
> <SNIP>
> -----------
>
> This now makes us have:
>
> [root@jouet ~]# perf probe -l
> probe:check_stack_object (on check_stack_object%return@acme/git/linux/mm/usercopy.c with ret)
> probe_libc:malloc (on __malloc in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_check (on malloc_check in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_consolidate (on malloc_consolidate.part.1 in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_hook_ini (on malloc_hook_ini in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_info (on __malloc_info in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_printerr (on malloc_printerr in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_stats (on __malloc_stats in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_trim (on __malloc_trim in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_usable_size (on __malloc_usable_size in /usr/lib64/libc-2.25.so)
> probe_libc:mallochook (on mallochook in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_check_return (on malloc_check%return in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_hook_ini_return (on malloc_hook_ini%return in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_info_return (on __malloc_info%return in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_printerr_return (on malloc_printerr%return in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_stats_return (on __malloc_stats%return in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_trim_return (on __malloc_trim%return in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_usable_size_return (on __malloc_usable_size%return in /usr/lib64/libc-2.25.so)
> probe_libc:mallochook_return (on mallochook%return in /usr/lib64/libc-2.25.so)
> probe_libc:malloc_return (on __malloc%return in /usr/lib64/libc-2.25.so)
> [root@jouet ~]#
>
> There were some usability snags that prevented this from being put in
> place, so do it manually:
>
> [root@jouet ~]# perf probe -x /lib64/libc-2.25.so 'malloc_consolidate_return=malloc_consolidate.part.1%return'
> Added new event:
> probe_libc:malloc_consolidate_return (on malloc_consolidate.part.1%return in /usr/lib64/libc-2.25.so)
>
> You can now use it in all perf tools, such as:
>
> perf record -e probe_libc:malloc_consolidate_return -aR sleep 1
>
> [root@jouet ~]#
>
> Ok, now we can trace this syswide using:
>
> [root@jouet ~]# perf trace --no-syscalls -e probe_libc:*
> <BIG SNIP>
> 339.557 probe_libc:malloc:(7fda55c22970))
> 339.562 probe_libc:malloc_return:(7fda55c22970 <- 7fda55c23918))
> 339.567 probe_libc:malloc_consolidate:(7fda55c1b7b0))
> 339.570 probe_libc:malloc_consolidate_return:(7fda55c1b7b0 <- 7fda55c1f886))
> 339.596 probe_libc:malloc:(7fda55c22970))
> 339.600 probe_libc:malloc_return:(7fda55c22970 <- 7fda55c23918))
> 339.622 probe_libc:malloc:(7fda55c22970))
> 339.626 probe_libc:malloc_return:(7fda55c22970 <- 7fda56ba9a39))
> 339.631 probe_libc:malloc:(7fda55c22970))
> 339.634 probe_libc:malloc_return:(7fda55c22970 <- 7fda578f1713))
> 339.637 probe_libc:malloc:(7fda55c22970))
> 339.641 probe_libc:malloc_return:(7fda55c22970 <- 7fda578f14ae))
> 339.657 probe_libc:malloc_consolidate:(7fda55c1b7b0))
> 339.661 probe_libc:malloc_consolidate_return:(7fda55c1b7b0 <- 7fda55c1dbd6))
> 339.664 probe_libc:malloc:(7fda55c22970))
> 339.667 probe_libc:malloc_return:(7fda55c22970 <- 7fda55c23918))
> 339.670 probe_libc:malloc:(7fda55c22970))
> 339.673 probe_libc:malloc_return:(7fda55c22970 <- 7fda55c23918))
> 339.678 probe_libc:malloc_consolidate:(7fda55c1b7b0))
> 339.681 probe_libc:malloc_consolidate_return:(7fda55c1b7b0 <- 7fda55c1f886))
> 339.685 probe_libc:malloc_consolidate:(7fda55c1b7b0))
> 339.688 probe_libc:malloc_consolidate_return:(7fda55c1b7b0 <- 7fda55c1f886))
> 339.693 probe_libc:malloc_consolidate:(7fda55c1b7b0))
> 339.696 probe_libc:malloc_consolidate_return:(7fda55c1b7b0 <- 7fda55c1f886))
> 339.778 probe_libc:malloc:(7fda55c22970))
> 339.783 probe_libc:malloc_return:(7fda55c22970 <- 7fda55c23918))
> 339.788 probe_libc:malloc_consolidate:(7fda55c1b7b0))
> ^C(7fda55c22970 <- 7fda578f14ae))
> [root@jouet ~]#
>
> It works as well with 'perf record' + 'perf script'.
>
> Other features from perf can be added to the mix, like callchains, etc.
>
> But see below for doing it via tracefs
>
> > I observed below behaviour when I tried as per directions using URL:
> > https://opensource.com/article/17/7/dynamic-tracing-linux-user-and-kernel-space#comment-136366
> > :
> >
> >
> > echo 'p:<func_entry> ./test:0x420' >
> > /sys/kernel/debug/tracing/uprobe_events (for creating
> > uprobe)
> > echo 1 > /sys/kernel/debug/tracing/events/enable
> > (enabling uproble)
> > There is another enable parameter:
> > /sys/kernel/debug/tracing/events/uprobes/enable
> >
> > I understand unless enabled both events/enable and
> > events/uprobes/enable, the /sys/kernel/debug/tracing/trace is not
> > complete. I do not understand the relevance of
> > .....events/uprobes/enable as I could not get any trace data in trace
> > log for 'func_entry' the user space function being added for tracing.
> > Irrespective of the probe function added, the trace log shows complete
> > system trace log during the period of execution of the user
> > application being traced.
>
> Is this what you want:
>
> [root@jouet ~]# cd /sys/kernel/debug/tracing/
> [root@jouet tracing]# cat available_tracers
> hwlat blk mmiotrace function_graph wakeup_dl wakeup_rt wakeup function nop
> [root@jouet tracing]# echo nop > current_tracer
> [root@jouet tracing]# echo 1 > events/probe_libc/malloc/enable
> [root@jouet tracing]# echo 1 > tracing_on
> [root@jouet tracing]# tail trace
> tail-1581 [002] d... 118484.507167: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507172: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507189: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507193: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507198: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507203: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507208: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507214: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507224: malloc: (0x7fdb32d67970)
> tail-1581 [002] d... 118484.507237: malloc: (0x7fdb32d67970)
> [root@jouet tracing]#
>
> ?
>
> > Can this be clearly mentioned as to when to use one of them, role of
> > both and and also, I could not see uprobe function in particular at
> > all but generic trace in /sys/kernel/debug/tracing/trace. Kindly guide
> > me in understanding uprobe enable logic in user space in detail.
> >
> > Regards,
> > Bhargav
> > --
--
Masami Hiramatsu <mhiramat@xxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
