Re: [PATCH v1 04/12] dyndbg: add 2 trace-events: pr_debug, dev_dbg

[Łukasz Bartosik]

pt., 10 lis 2023 o 20:20 <jim.cromie@xxxxxxxxx> napisał(a):
>
> On Fri, Nov 10, 2023 at 7:51 AM Łukasz Bartosik <lb@xxxxxxxxxxxx> wrote:
> >
> > wt., 7 lis 2023 o 00:55 Steven Rostedt <rostedt@xxxxxxxxxxx> napisał(a):
> > >
> > > On Fri,  3 Nov 2023 14:10:03 +0100
> > > Łukasz Bartosik <lb@xxxxxxxxxxxx> wrote:
> > >
> > > > +/* capture pr_debug() callsite descriptor and message */
> > > > +TRACE_EVENT(prdbg,
> > > > +         TP_PROTO(const struct _ddebug *desc, const char *text, size_t len),
> > > > +
> > > > +         TP_ARGS(desc, text, len),
> > > > +
> > > > +         TP_STRUCT__entry(
> > > > +                     __field(const struct _ddebug *, desc)
> > > > +                     __dynamic_array(char, msg, len + 1)
> > > > +                 ),
> > > > +
> > > > +         TP_fast_assign(
> > > > +                     __entry->desc = desc;
> > > > +                     /*
> > > > +                      * Each trace entry is printed in a new line.
> > > > +                      * If the msg finishes with '\n', cut it off
> > > > +                      * to avoid blank lines in the trace.
> > > > +                      */
> > > > +                     if (len > 0 && (text[len - 1] == '\n'))
> > > > +                             len -= 1;
> > > > +
> > > > +                     memcpy(__get_str(msg), text, len);
> > > > +                     __get_str(msg)[len] = 0;
> > > > +                 ),
> > > > +
> > >
> > >
> > > > +         TP_printk("%s.%s %s", __entry->desc->modname,
> > > > +                   __entry->desc->function, __get_str(msg))
> > > > +);
> > > > +
> > >
> > > That TP_printk() is dangerous. How do you know __entry->desc still exists
> > > when reading the buffer?
> > >
> > > Is the struct _ddebug permanent? Can it be freed? If so, the above can
> > > easily cause a crash.
> > >
> >
> > I assume that we're talking here about the scenario where TP prdbg is
> > called and before TP_printk runs _ddebug pointer
> > becomes invalid, is that correct ? If so then I believe this also
> > applied to __dynamic_pr_debug and other dyndbg functions because there
> > is also potential for _ddebug pointer to become invalid (in case of
> > rrmod) before a function dereferences it.
> >
> > Would it be acceptable to increase reference count of a module and
> > hold it until at least one callsite in that module is enabled ?
> > This  would ensure that passed pointer to a _ddebug struct is valid.
> >
>
> Im not understanding you, but I dont think its on-point -
>
> a loadable module might write lots to trace-log, and each trace-entry
> would have the descriptor address, with which it could deref and print 3 fields.
> Then rmmod happens, all the module mem is freed, and reused for someth9ing else.
>
> then someone cats trace, and the descriptor addrs are used to render
> the tracelog.
> BOOM.
>

Jim, thanks for educating me on this one. I completely missed the fact
that TP_printk is delayed until, for example as you mentioned cat is
run on a trace.

I'll remove passing of _ddebug ptr to trace_prdbg and trace_devdbg
functions. Probably also passing of dev ptr can removed from
trace_devdbg.

>
>
> > > -- Steve