Re: Task watchers v2

[Matt Helsley <matthltc@us.ibm.com>]

On Mon, 2006-12-18 at 13:44 +0800, Zhang, Yanmin wrote:
> On Thu, 2006-12-14 at 16:07 -0800, Matt Helsley wrote:
> > plain text document attachment (task-watchers-v2)
> > Associate function calls with significant events in a task's lifetime much like
> > we handle kernel and module init/exit functions. This creates a table for each
> > of the following events in the task_watchers_table ELF section:
> >
> > WATCH_TASK_INIT at the beginning of a fork/clone system call when the
> > new task struct first becomes available.
> >
> > WATCH_TASK_CLONE just before returning successfully from a fork/clone.
> >
> > WATCH_TASK_EXEC just before successfully returning from the exec
> > system call.
> >
> > WATCH_TASK_UID every time a task's real or effective user id changes.
> >
> > WATCH_TASK_GID every time a task's real or effective group id changes.
> >
> > WATCH_TASK_EXIT at the beginning of do_exit when a task is exiting
> > for any reason.
> >
> > WATCH_TASK_FREE is called before critical task structures like
> > the mm_struct become inaccessible and the task is subsequently freed.
> >
> > The next patch will add a debugfs interface for measuring fork and exit rates
> > which can be used to calculate the overhead of the task watcher infrastructure.
> >
> > Subsequent patches will make use of task watchers to simplify fork, exit,
> > and many of the system calls that set [er][ug]ids.
> It's easier to get such watch capabilities by kprobe/systemtap. Why to
> add new codes to kernel?

Good question! Disclaimer: Everything I know about kprobes I learned
from Documentation/kprobes.txt

The task watchers patches have a few distinguishing capabilities yet
lack capabilities important for kprobes -- so neither is a replacement
for the other. Specifically:

- Task watchers are for use by the kernel for more than profiling and
debugging. They need to work even when kernel debugging and
instrumentation are disabled.

- Task watchers do not need to be dynamically enabled, disabled, or
removed (though dynamic insertion would be nice -- I'm working on that).
In fact I've been told that dynamically enabling, disabling, or removing
them would incur unacceptable complexity and/or cost for an
uninstrumented kernel.

- Task watchers don't require arch support. They use completely generic
code.
	- Since they are written into the code task watchers don't need
	  to modify instructions.

	- Task watchers doesn't need to single-step an instruction

	- Task watchers don't need to know about arch registers, calling
	  conventions, etc. to work

- Task watchers don't need to have the same (possibly extensive)
argument list as the function being "probed". This makes maintenance
easier -- no need to keep the signature of the watchers in synch with
the signature of the "probed" function.

- Task watchers don't require MODULES (2.6.20-rc1-mm1's
arch/i386/Kconfig suggests this is true of kprobes).

- Task watchers don't need kernel symbols.

- Task watchers can affect flow control (see the patch hunks that change
copy_process()) with their return value.

- Task watchers do not need to know the instruction address to be
"probed".

- Task watchers can actually improve kernel performance slightly (up to
2% in extremely fork-heavy workloads for instance).

- Task watchers require local variables -- not necessarily arguments to
the "probed" function.

- Task watchers don't care if preemption is enabled or disabled.

- Task watchers could sleep if they want to.

	So to the best of my knowledge kprobes isn't a replacement for task
watchers nor is task watchers capable of replacing kprobes.

Cheers,
	-Matt Helsley