Re: [RFC PATCH -tip 0/5] kprobes: Abolish jprobe APIs

[Steven Rostedt]

On Sat, 7 Oct 2017 18:35:40 +0900
Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:


> > The ftrace context
> > ==================
> > 
> > WARNING: The ability to add a callback to almost any function within the
> > kernel comes with risks. A callback can be called from any context
> > (normal, softirq, irq, and NMI). Callbacks can also be called just before
> > going to idle, during CPU bring up and takedown, or going to user space.
> > This requires extra care to what can be done inside a callback. A callback
> > can be called outside the protective scope of RCU.
> > 
> > The ftrace infrastructure has some protections agains recursions and RCU
> > but one must still be very careful how they use the callbacks.  
> 
> Q: As far as I know the ftrace handler is called under preempt-disabled,
>  don't we need to mention that here?

Actually it is not. Well not always. I should document this. If the
ftrace handler has the right flags and the arch supports it, the ftrace
handler may be called directly by the function being traced via a
simple trampoline. The trampoline is called directly by the function,
and that trampoline will do minimal register set up to call the
callback function. There will be no preempt disabling when it gets
called.

Now, if more than one function is registered to the same function, or
if the function is not recursive safe, or requires to be per-cpu, then
ftrace itself may disable preemption before calling the callback. The
callback should be prepare for either. And yes, this should be added to
the document.

> 
> > 
> > The ftrace_ops structure
> > ========================
> > 
> > To register a function callback, a ftrace_ops is required. This structure
> > is used to tell ftrace what function should be called as the callback
> > as well as what protections the callback will perform and not require
> > ftrace to handle.
> > 
> > There are only two fields that are needed to be set when registering
> > an ftrace_ops with ftrace. The rest should be NULL.
> > 
> > struct ftrace_ops ops = {
> >        .func			= my_callback_func,
> >        .flags			= MY_FTRACE_FLAGS
> >        .private			= any_private_data_structure,
> > };
> > 
> > Both .flags and .private are optional. Only .func is required.
> > 
> > To enable tracing call:
> > 
> >   register_ftrace_function(&ops);
> > 
> > To disable tracing call:
> > 
> >   unregister_ftrace_function(@ops);  
>                                ^^^^Is this &ops?

Oops, yes :-)

> 
> > 
> > 
> > The callback function
> > =====================
> > 
> > The prototype of the callback function is as follows (as of v4.14):  
>                                                                 ^^^^^^^^^^^^^
> If we put this document under Documentation/, do we still need it?
> (since it should be updated along with the code)

It should be, but I'm also a realist. Stuff in Documentation is slow to
update. I added this there so that it can catch my eye and if I see
that, and know that I updated the functions, I know that the document
needs to be updated too.

> 
> >  void callback_func(unsigned long ip, unsigned long parent_ip,
> > 		    struct ftrace_ops *op, struct pt_regs *regs);
> > 
> > @ip - This is the instruction pointer of the function that is being traced.
> >       (where the fentry or mcount is within the function)
> > 
> > @parent_ip - This is the instruction pointer of the function that called the
> >       the function being traced (where the call of the function occurred).
> > 
> > @op - This is a pointer to ftrace_ops that was used to register the callback.
> >       This can be used to pass data to the callback via the private pointer.
> > 
> > @regs - If the FTRACE_OPS_FL_SAVE_REGS or FTRACE_OPS_FL_SAVE_REGS_IF_SUPPORTED
> >       flags are set in the ftrace_ops structure, then this will be pointing
> >       to the pt_regs structure like it would be if an breakpoint was placed
> >       at the start of the function where ftrace was tracing. Otherwise it
> >       either contains garbage, or NULL.
> > 
> > 
> > The ftrace FLAGS
> > ================
> > 
> > The ftrace_ops flags are all defined and documented in include/linux/ftrace.h.
> > Some of the flags are used for internal infrastructure of ftrace, but the
> > ones that users should be aware of are the following:
> > 
> > (All of these are prefixed with FTRACE_OPS_FL_)
> > 
> > PER_CPU - When set, the callback can be enabled or disabled per cpu with the
> >       following functions:
> > 
> >       void ftrace_function_local_enable(struct ftrace_ops *ops);
> >       void ftrace_function_local_disable(struct ftrace_ops *ops);
> > 
> >       These two functions must be called with preemption disabled.
> > 
> > SAVE_REGS - If the callback requires reading or modifying the pt_regs
> >       passed to the callback, then it must set this flag. Registering
> >       a ftrace_ops with this flag set on an architecture that does not
> >       support passing of pt_regs to the callback, will fail.
> > 
> > SAVE_REGS_IF_SUPPORTED - Similar to SAVE_REGS but the registering of a
> >       ftrace_ops on an architecture that does not support passing of regs
> >       will not fail with this flag set. But the callback must check if
> >       regs is NULL or not to determine if the architecture supports it.
> > 
> > RECURSION_SAFE - By default, a wrapper is added around the callback to
> >       make sure that recursion of the function does not occur. That is
> >       if a function within the callback itself is also traced, ftrace
> >       will prevent the callback from being called again. But this wrapper
> >       adds some overhead, and if the callback is safe from recursion,
> >       it can set this flag to disable the ftrace protection.  
> 
> ( Ah, I didn't noticed this flag )

Yes, with this flag set, preemption may not be disabled. I should
document this a bit better.

> 
> > 
> > IPMODIFY - Requires SAVE_REGS set. If the callback is to "hijack" the
> >       traced function (have another function called instead of the traced
> >       function), it requires setting this flag. This is what live kernel
> >       patches uses. Without this flag the pt_regs->ip can not be modified.
> >       Note, only one ftrace_ops with IPMODIFY set may be registered to
> >       any given function at a time.
> > 
> > RCU - If this is set, then the callback will only be called by functions
> >       where RCU is "watching". This is required if the callback function
> >       performs any rcu_read_lock() operation.
> > 
> > 
> > Filtering what functions to trace
> > =================================
> > 
> > If a callback is only to be called from specific functions, a filter must be
> > set up. The filters are added by name, or ip if it is known.
> > 
> >  int ftrace_set_filter(struct ftrace_ops *ops, unsigned char *buf,
> > 		       int len, int reset);
> > 
> >  @ops - the ops to set the filter with
> >  @buf - the string that holds the function filter text.
> >  @len - the length of the string.
> >  @reset - non zero to reset all filters before applying this filter.
> > 
> >  Filters denote which functions should be enabled when tracing is enabled.
> >  If @buf is NULL and reset is set, all functions will be enabled for tracing.
> > 
> > 
> > The @buf can also be a glob expression to enable all functions that
> > match a specific pattern.
> > 
> > To just trace the schedule function:
> > 
> >  ret = ftrace_set_filter(&ops, "schedule", strlen("schedule"), 0);
> > 
> > To add more functions, call the ftrace_set_filter() more than once with the
> > @reset parameter set to zero. To remove the current filter and replace it
> > with new functions to trace, have @reset be non zero.
> > 
> > Sometimes more than one function has the same name. To trace just a specific
> > function in this case, ftrace_set_filter_ip() can be used.
> > 
> >  ret = ftrace_set_filter_ip(&ops, ip, 0, 0);
> > 
> > Although the ip must be the address where the call to fentry or mcount is
> > located in the function.
> > 
> > If a glob is used to set the filter, to remove unwanted matches the
> > ftrace_set_notrace() can also be used.
> > 
> >   int ftrace_set_notrace(struct ftrace_ops *ops, unsigned char *buf,
> > 			 int len, int reset);
> > 
> > This takes the same parameters as ftrace_set_filter() but will add the
> > functions it finds to not be traced. This doesn't remove them from the
> > filter itself, but keeps them from being traced. If @reset is set,
> > the filter is cleaded but the functions that match @buf will still not  
>                   ^^^^^^^^
>                        cleared?
> 
> > be traced (the callback will not be called on those functions).  

Yep, I guess that stood out like a sore thumb.

-- Steve