[PATCH 0/2 v2] Syscalls tracing

[Frederic Weisbecker <fweisbec@gmail.com>]

tracing/syscalls: core infrastructure to trace syscalls

This new iteration addresses a good part of the previous reviews.

As suggested by Ingo Molnar and Peter Zijlstra, the syscalls
prototypes probing is done by abusing the SYSCALL_DEFINE family
macros.
We now store automatically the arguments names, their types, their number
and the name of the syscall.

Also some fixes on output newlines and dangerous exporting of global_trace
are provided.

An example of the trace:

echo syscall > /debugfs/tracing/current_tracer


# tracer: syscall
#
#           TASK-PID    CPU#    TIMESTAMP  FUNCTION
#              | |       |          |         |
           <...>-5080  [001]   132.192228: sys_dup2(oldfd: a, newfd: 1) 
           <...>-5080  [001]   132.192239: sys_dup2 -> 0x1
           <...>-5080  [001]   132.192242: sys_fcntl(fd: a, cmd: 1, arg: 0) 
           <...>-5080  [001]   132.192245: sys_fcntl -> 0x1
           <...>-5080  [001]   132.192248: sys_close(fd: a) 
           <...>-5080  [001]   132.192250: sys_close -> 0x0
           <...>-5080  [001]   132.192265: sys_rt_sigprocmask(how: 0, set: 0, oset: 6cf808, sigsetsize: 8) 
           <...>-5080  [001]   132.192267: sys_rt_sigprocmask -> 0x0
           <...>-5080  [001]   132.192271: sys_rt_sigaction(sig: 2, act: 7fffff4338f0, oact: 7fffff433850, sigsetsize: 8) 
           <...>-5080  [001]   132.192273: sys_rt_sigaction -> 0x0
           <...>-5080  [001]   132.192285: sys_rt_sigprocmask(how: 0, set: 0, oset: 6cf808, sigsetsize: 8) 
           <...>-5080  [001]   132.192287: sys_rt_sigprocmask -> 0x0
           <...>-5080  [001]   132.192415: sys_write(fd: 1, buf: 15dfc08, count: 21) 
           <...>-5080  [001]   132.192436: sys_write -> 0x21
           <...>-4754  [000]   132.192478: sys_read(fd: 8, buf: 2a9340e, count: 1fee) 
           <...>-5080  [001]   132.192487: sys_rt_sigprocmask(how: 0, set: 7fffff432a70, oset: 7fffff4329f0, sigsetsize: 8) 

And if you ask for the parameters types:

echo syscall_arg_type > trace_options

# tracer: syscall
#
#           TASK-PID    CPU#    TIMESTAMP  FUNCTION
#              | |       |          |         |
           <...>-5080  [001]   132.192228: sys_dup2(unsigned int oldfd: a, unsigned int newfd: 1) 
           <...>-5080  [001]   132.192239: sys_dup2 -> 0x1
           <...>-5080  [001]   132.192242: sys_fcntl(unsigned int fd: a, unsigned int cmd: 1, unsigned long arg: 0) 
           <...>-5080  [001]   132.192245: sys_fcntl -> 0x1
           <...>-5080  [001]   132.192248: sys_close(unsigned int fd: a) 
           <...>-5080  [001]   132.192250: sys_close -> 0x0
           <...>-5080  [001]   132.192265: sys_rt_sigprocmask(int how: 0, sigset_t * set: 0, sigset_t * oset: 6cf808, size_t sigsetsize: 8) 
           <...>-5080  [001]   132.192267: sys_rt_sigprocmask -> 0x0
           <...>-5080  [001]   132.192271: sys_rt_sigaction(int sig: 2, const struct sigaction * act: 7fffff4338f0, struct sigaction * oact: 7fffff433850, size_t sigsetsize: 8) 
           <...>-5080  [001]   132.192273: sys_rt_sigaction -> 0x0
           <...>-5080  [001]   132.192285: sys_rt_sigprocmask(int how: 0, sigset_t * set: 0, sigset_t * oset: 6cf808, size_t sigsetsize: 8) 
           <...>-5080  [001]   132.192287: sys_rt_sigprocmask -> 0x0
           <...>-5080  [001]   132.192415: sys_write(unsigned int fd: 1, const char * buf: 15dfc08, size_t count: 21) 
           <...>-5080  [001]   132.192436: sys_write -> 0x21

TODO:

- add a single mask on the struct syscall_metadata to provide quickly which arguments is
  a pointer (usually type __user *p) so that the user can decide if he wants to save them of tracing time

- now that we have each parameter type as strings, add a new field on struct syscall_metadata to have the parameter types
  encoded as single enum values (for quick checks) so that we can use specific callbacks for each parameter type
  to be displayed.

NOTE: this is still not overlapping with a potential future merge of utrace, since the low-level hooks on
the syscalls remain somewhat basic.

NOTE2: I've only tested it on x86-64 for now, so only x86-64 support is provided.
--

Frederic Weisbecker (2):
  tracing/syscalls: core infrastructure for syscalls tracing
  tracing/syscalls: support for syscalls tracing on x86-64

 arch/x86/Kconfig                   |    1 +
 arch/x86/include/asm/ftrace.h      |    7 +
 arch/x86/include/asm/thread_info.h |    9 +-
 arch/x86/kernel/ftrace.c           |   63 +++++++++
 arch/x86/kernel/ptrace.c           |    7 +
 include/asm-generic/vmlinux.lds.h  |   11 ++-
 include/linux/ftrace.h             |   29 +++++
 include/linux/syscalls.h           |   60 +++++++++-
 kernel/trace/Kconfig               |   10 ++
 kernel/trace/Makefile              |    1 +
 kernel/trace/trace.h               |   19 +++
 kernel/trace/trace_syscalls.c      |  243 ++++++++++++++++++++++++++++++++++++
 12 files changed, 454 insertions(+), 6 deletions(-)
 create mode 100644 kernel/trace/trace_syscalls.c