lkml.org 
[lkml]   [2009]   [Mar]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/2 v2] Syscalls tracing
On Mon, Mar 23, 2009 at 12:32:35PM -0400, Mathieu Desnoyers wrote:
> * Frederic Weisbecker (fweisbec@gmail.com) wrote:
> > On Fri, Mar 13, 2009 at 12:47:43PM -0400, Mathieu Desnoyers wrote:
> > > * Frederic Weisbecker (fweisbec@gmail.com) wrote:
> > > > On Fri, Mar 13, 2009 at 03:42:10PM +0100, Frederic Weisbecker wrote:
> > > > > tracing/syscalls: core infrastructure to trace syscalls
> > > > >
> > > > > This new iteration addresses a good part of the previous reviews.
> > > >
> > > >
> > > > Ah I just discovered that you applied the previous version today.
> > > > But the v2 is not a delta :-s
> > > >
> > > > I can rebase them but not until Sunday.
> > > >
> > >
> > > Hi Frederic,
> > >
> > > I like the approach you are taking for syscall tracing. As I read your
> > > code, I see that your infrastructure basically exports the information
> > > under ascii format. How hard do you think that would be to get this to
> > > work for binary output with LTTng ? Ascii output is not exactly a
> > > high-performance output for a hot path such as system call entry/exit.
> > > We could :
> >
> >
> > Do you mean the syscalls metadata? (syscall name, argument types, number
> > of arguments) or the syscall events?
> >
> > I'm not sure it would be valuable to export the metadata as binary because
> > they are not very large as a whole. But a kind of enumeration of types
> > will be added as I explain below.
> >
> > But I plan to export the events as binary values, though the usual tracing
> > files.
> >
>
> Syscall metadata could be exported in ascii. Syscall events could be
> exported as binary, because they generate a high trace data throughput.


Yeah, agreed, that what I meant.


> > > 1 - Iterate on the table at trace start to dump the syscall names, and
> > > their arguments, into a syscall-table data channel.
> >
> >
> > Yeah.
> > For now the arguments parameters types are saved as strings as you know.
> > I would like to add a new field on the syscall_metadata struct which will
> > contain the arg types as enum values so that the tracer will be able to:
> >
> > - Define custom callback for custom types to display better the arguments
> > I think I could provide a default output for each parameter type and
> > let the ability to override this default output with custom callback.
> > For example, it would be great to have the open flag as O_RDONLY instead
> > of the numerical value.
>
> Yes, and we should make sure we save the information in the trace
> buffers in the same way it is presented in the kernel (e.g. a flag
> within a unsigned int), and use a type table to map the trace data to
> some human-readable information. This type table could be in text
> format or some fixed, easy to parse, format.
>
> Basically, the O_RDONLY flag would simply be written as a "int" into
> the trace buffers (no special callback required), but a mappping to
> O_RDONLY, OR_WRONLY, etc.. would be registered in the type table for
> this specific event types.


Indeed a general type table which could be overriden or completed by
a per event table, for those which require it.


> >
> > - probably create a raw format string value for each syscall so that we can
> > export the binary values directly to userspace and the user can format all
> > that with the pattern we give to him.
> >
>
> LTTng now provides the ability to access the event type table table from
> within the kernel, so we could save the information in binary form into
> the trace buffers and then, only when the text output is required,
> format it into text format. This would unify the binary and text output
> mechanisms called at the tracing site, which is the most critical
> tracing execution site.


That's what is done currently. Nothing is stored in plain formatted
ascii in the ring buffer for the syscall tracing.
It's all raw binary. Ie, we dump the registers contents and the number
of the syscall.

Only on output time we perform the formatting, which later could be
a choice between a binary export and a formatted text.


> I also wonder about the impact of the syscall tracing approach you are
> taking compared to adding instrumentation directly into the kernel code
> (e.g. directly in sys_open), which would have the following advantages :
>
> - The data has already been copied to kernel-space, so there is no need
> to duplicate copy_from_user/to_user() calls.


Indeed, I haven't thought about it.


> - This second approach would let us trace syscall code called directly
> from within the kernel (e.g. from kernel threads).
>
> What do you think ?


Yes, but it means adding hundreds of trace{points,events} over the kernel.
Moreover, the current CPP probe let us very easily factorize it, making it very
scalable.

And actually I don't think two copy_from_user will really change a lot the
tracing throughput.
The only difference with your approach is that we'll have:

sys_call do the copy_from_user()
memcpy(ring_buffer, arg_copied, n)

instead of:

copy_from_user(ring_buffer, __user arg, n)
sys_call do the copy_from_user

With the second solution we risk a tlb miss on ring_buffer stage,
or worst: a page fault.
But once it is done, when the syscall performs its copy_from_user, there
are few chances that it encounters one more time a tlb miss or a page fault
for the same page since it will be called very soon after.

Of course this is all theory, and only practice could prove it.

Concerning the syscalls performed by the kernel threads, this is not
the same focus of tracing. When we trace a syscall, we mostly want to see
what happens when the kernel is servicing a user.
But well, it can be debated.

I don't know much about it. But when a kernel thread performs a syscall,
I guess it is often wrapped by an in-kernel helper. For example to create
a thread, a call to fork is done, but this is wrapped in a call to
kthread_create(). So a tracepoint can be placed here.
I think a lot of examples match this pattern concerning the kernel threads
and a tracepoint in such higher level helper for the kthreads seems to
me better to capture a kernel thread creation event.

Furthermore, AFAICS, only the kthreadd thread forks (and init, which has its
own helper for that too).


>
> >
> > > 2 - We might have to create one event type per syscall argument types.
> > > Or one event per syscall, this would be great because we would benefit
> > > from the event identification infrastructure. And actually, the syscall
> > > declaration macros could declare tiny callbacks that would declare a
> > > marker name (syscall, <syscall_name>) along with the code that saves the
> > > system call pr_regs parameters into the event payload.
> > >
> > > Any thoughts ?
> >
> >
> > I will surely have to provide a way to define custom callbacks for several
> > syscalls (as shown with the open() flag example).
> > But the problem is that I will not be able to do that statically like the
> > metadata is currently built.
> > Such a thing should be built when we enable the syscall tracing once I guess
> > (iterate over the syscalls metadata, assign the callback that matches a syscall,
> > etc...).
>
> In LTTng, the metadata is already generated dynamically. For instance,
> when we want to list the connected interrupt handlers, we have to
> support the fact that drivers can be loaded/unloaded. Therefore, we list
> the interrupt handlers at trace start, but we also instrument the
> insertion/removal of handlers while the trace runs, so we know at each
> point in time which handlers are connected. This mechanism would fit
> well with your system call type dump.


Cool, and how could it be used by the syscalls to define specific
callbacks for their types (for example with thee open() flag, or any generic
types like common int)?


> >
> > But I think I will need sometimes complex callbacks to handle the output of
> > some complex arguments, so I don't know if it is possible to handled all that
> > through a marker or TRACE_EVENT style.
> >
>
> In my lttng tree, ltt/probes/block-trace.c already has something like
> what you need for the open() flags :
>
>
> void probe_block_bio_complete(struct request_queue *q, struct bio *bio)
> {
> trace_mark_tp(block, bio_complete, block_bio_complete,
> probe_block_bio_complete,
> "sector %llu size %u rw(FAILFAST_DRIVER,FAILFAST_TRANSPORT,"
> "FAILFAST_DEV,DISCARD,META,SYNC,BARRIER,AHEAD,RW) %lX "
> "not_uptodate #1u%d",
> (unsigned long long)bio->bi_sector, bio->bi_size,
> bio->bi_rw, !bio_flagged(bio, BIO_UPTODATE));
> }
>
> an event type parser could then map the data saved in the "rw" unsigned
> long argument to the bitfield composed of :
>
> FAILFAST_DRIVER
> FAILFAST_TRANSPORT
> FAILFAST_DEV
> DISCARD
> META
> SYNC
> BARRIER
> AHEAD
> RW
>


Oh I see!


> It's very basic, and I'm sure we could improve this, but it gives you
> the idea of how I plan to deal with this.
>
> Mathieu


I understand better now.
The idea would be now to join the syscalls metadata with such quick
handlers. We will have to think about how to join these in a proper
way.

Thanks!

Frederic.


>
> > Thanks!
> >
> > Frederic.
> >
> >
> > >
> > > Mathieu
> > >
> > > >
> > > > > 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
> > > > >
> > > >
> > >
> > > --
> > > Mathieu Desnoyers
> > > OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
> >
>
> --
> Mathieu Desnoyers
> OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68



\
 
 \ /
  Last update: 2009-03-23 20:29    [from the cache]
©2003-2011 Jasper Spaans. Advertise on this site