[ktap] Re: [PATCH] add new symbol module to search symbols in DSO (for uprobes)
- From: Jovi Zhangwei <jovi.zhangwei@xxxxxxxxx>
- To: ktap@xxxxxxxxxxxxx
- Date: Tue, 5 Nov 2013 12:34:48 +0800
Hi Azat,
On Tue, Nov 5, 2013 at 1:57 AM, Azat Khuzhin <a3at.mail@xxxxxxxxx> wrote:
> This will allow to don't hardcoding addresses of functions in ktap
> scripts, that use uprobes, but instead, use normal function names.
>
> Before this patch:
> $ ktap -e 'trace probe:/path/to/bin:0xDEADBEAF' # instead some real address
>
> After this path:
> $ ktap -e 'trace probe:/path/to/bin:foo'
>
> Works also for ret probes (%return)
> Requires libelf.
>
> Signed-off-by: Azat Khuzhin <a3at.mail@xxxxxxxxx>
> ---
Really thanks your patch, I always want to support debugging info
handling into ktap,
but I didn't have enough time, I appreciate you can take this task.
Some comments embedded below.
> Makefile | 5 ++-
> include/ktap_types.h | 2 +
> userspace/eventdef.c | 34 +++++++++++++++
> userspace/symbol.c | 114
> +++++++++++++++++++++++++++++++++++++++++++++++++++
> userspace/symbol.h | 32 +++++++++++++++
> 5 files changed, 186 insertions(+), 1 deletion(-)
> create mode 100644 userspace/symbol.c
> create mode 100644 userspace/symbol.h
>
> diff --git a/Makefile b/Makefile
> index 372b41a..344c82a 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -59,6 +59,8 @@ $(UDIR)/tstring.o: $(INTP)/tstring.c $(INC)/*
> $(QUIET_CC)$(CC) $(DEBUGINFO_FLAG) $(KTAPC_CFLAGS) -o $@ -c $<
> $(UDIR)/object.o: $(INTP)/object.c $(INC)/*
> $(QUIET_CC)$(CC) $(DEBUGINFO_FLAG) $(KTAPC_CFLAGS) -o $@ -c $<
> +$(UDIR)/symbol.o: $(UDIR)/symbol.c $(INC)/*
> + $(QUIET_CC)$(CC) $(DEBUGINFO_FLAG) $(KTAPC_CFLAGS) -o $@ -c $<
>
> KTAPOBJS =
> KTAPOBJS += $(UDIR)/lex.o
> @@ -73,9 +75,10 @@ KTAPOBJS += $(UDIR)/opcode.o
> KTAPOBJS += $(UDIR)/table.o
> KTAPOBJS += $(UDIR)/tstring.o
> KTAPOBJS += $(UDIR)/object.o
> +KTAPOBJS += $(UDIR)/symbol.o
>
> ktap: $(KTAPOBJS)
> - $(QUIET_LINK)$(CC) $(KTAPC_CFLAGS) -o $@ $(KTAPOBJS) -lpthread
> + $(QUIET_LINK)$(CC) $(KTAPC_CFLAGS) -o $@ $(KTAPOBJS) -lpthread -lelf
>
This will be a problem for embedded system which don't have libelf,
then they cannot
use ktap in those small system just because lack libelf.
So I suggest let libelf linking to be an option, like perf Makefile,
there have NO_LIBELF
option in perf Makefile. If user compiler ktap without libelf support,
then they cannot use
debugging info, but still can use direct memory address or other
useful functionality in ktap.
You can take more reference on perf Makefile in Linux kernel.
> KMISC := /lib/modules/$(KVERSION)/ktapvm/
>
> diff --git a/include/ktap_types.h b/include/ktap_types.h
> index 1dae9d8..4e19f0b 100644
> --- a/include/ktap_types.h
> +++ b/include/ktap_types.h
> @@ -600,5 +600,7 @@ extern ktap_global_state dummy_global_state;
> #define KTAP_QL(x) "'" x "'"
> #define KTAP_QS KTAP_QL("%s")
>
> +#define STRINGIFY(type) #type
> +
> #endif /* __KTAP_TYPES_H__ */
>
> diff --git a/userspace/eventdef.c b/userspace/eventdef.c
> index 76a68ac..dab964a 100644
> --- a/userspace/eventdef.c
> +++ b/userspace/eventdef.c
> @@ -29,6 +29,7 @@
> #include "../include/ktap_types.h"
> #include "../include/ktap_opcodes.h"
> #include "ktapc.h"
> +#include "symbol.h"
>
> static char tracing_events_path[] = "/sys/kernel/debug/tracing/events";
>
> @@ -305,6 +306,35 @@ static int parse_events_add_kprobe(char *old_event)
>
> #define UPROBE_EVENTS_PATH "/sys/kernel/debug/tracing/uprobe_events"
>
> +static char* parse_events_resolve_symbol(char *event, char *r)
> +{
> + char *colon;
> + colon = strchr(event, ':');
> +
> + char *binary = strndup(event, colon - event);
> + char *symbol = strndup(colon + 1 /* skip ":" */,
> + r ? (r - 1 /* skip "%" */ - colon) : strlen((const char
> *)colon));
> +
I guess this would be a problem if use argument fetching in uprobe, for example,
trace probe:/lib/libc.so:foo $arg1 $arg2 {
print(argevent)
}
Basically kprobe and uprobe tracing syntax was borrowed from perf-probe,
and should be compatible with perf-probe syntax as much as possible.
BTW, perf-probe was contributed by Masami(Maintainer of kprobe in Linux kernel).
So in current stage, it's fine to not care about the argument symbol parsing,
just parse the function symbol is enough.
Would you like to change the code a little to handle parse symbol if argument
appended? Thanks.
More info could take reference by perf-probe command option.
Jovi.
Other related posts:
