How to Trace Just System Call Events with Ftrace Without Showing Any Other Functions in the Linux Kernel

How can I use ftrace to get the in-kernel call graph called by a system call?

First you need to get the function name right - e.g. the function name to use for tracing open syscalls is sys_open.

To do this the "proper" way, it's necessary to have function_graph support in the kernel. On the x86 architecture this depends on CC_OPTIMIZE_FOR_SIZE being disabled, but on x86_64 it doesn't.

In my case I didn't bother to compile a custom kernel to disable CC_OPTIMIZE_FOR_SIZE, I just did

trace-cmd record -p function --func-stack

and included various functions that looked like they might be called by adding several -l options. This was enough to figure out what I wanted to know.

How to capture all generated events with ftrace without any loss

I solved this problem by also capturing the event syscalls:sys_enter_execve. Between the two of them I was able to get every instance of exec called.

Does ftrace allow capture of system call arguments to the Linux kernel, or only function names?

I have limited experience with ftrace, although I have used it for for function stack traces and latency issues. (People with more experience can possibly suggest) Its pretty much the same experience using trace-cmd and kernelshark.

However, if you want to trace syscalls, function params, kernel APIs and return values etc. within the kernel space a better choice would be to go with systemtap. It has an extensive list of Samples & Doc which is good for function call tracing, argument values passed etc. You may want to look at some samples and taylor them to your requirement. See general/para-callgraph-verbose.stp and process/sleeptime.stp

"

general/para-callgraph-verbose.stp - Callgraph Tracing with Verbose Arguments
keywords: TRACE CALLGRAPH

Print a timed per-thread microsecond-timed callgraph, complete with pretty-printed function parameters and return values. The first parameter names the function probe points to trace. The optional second parameter names the probe points for trigger functions, which acts to enable tracing for only those functions that occur while the current thread is nested within the trigger.

stap para-callgraph-verbose.stp 'kernel.function("*@fs/proc*.c")' \
'kernel.function("vfs_read")' -c "cat /proc/sys/vm/* || true"

process/strace.stp - Trace system calls
keywords: _BEST PROCESS SYSCALL

The script loosely emulates strace, when applied to individual processes or hierarchies (via -c/-x), or the entire system (without -c/-x). A few output configuration parameters may be set with -G.

stap strace.stp -c "sleep 1"

"

Note you will need to install the correct version of the debug kernel and kernel-devel rpms/deb for stap to work correctly. For this just use stap-prep and install the dependencies shown depending on the flavour you are on.

---

Related Topics