On Tue, Jan 25, 2022 at 04:23:26PM -0500, Steven Rostedt wrote:
> On Tue, 25 Jan 2022 12:57:24 -0800
> Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> > On Tue, Jan 25, 2022 at 11:31:54PM +0900, Masami Hiramatsu wrote:
> > > On Tue, 25 Jan 2022 22:27:32 +0900
> > > Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
> > >
> > > > > /*
> > > > > * struct trace_event_data_offsets_<call> {
> > > > > * u32 <item1>;
> > > > > * u32 <item2>;
> > > > > * [...]
> > > > > * };
> > > > > *
> > > > > * The __dynamic_array() macro will create each u32 <item>, this is
> > > > > * to keep the offset of each array from the beginning of the event.
> > > > > * The size of an array is also encoded, in the higher 16 bits of
> > > > > * <item>.
> > > > > */
> > > > >
> > > > > So, I think -Warray-bounds is refusing to see the destination as
> > > > > anything except a u32, but being accessed at 4 (sizeof(u32)) + 8
> > > > > (address && 0xffff) (?)
> > > >
> > > > Ah, I got it. Yes, that's right. __data_loc() will access the data
> > > > from the __entry, but the __rel_loc() points the same address from
> > > > the encoded field ("__rel_loc_foo" in this case) itself.
> > > > This is introduced for the user application event, which doesn't
> > > > know the actual __entry size because the __entry includes some
> > > > kernel internal defined fields.
> > > >
> > > > > But if this is true, I would imagine there would be plenty of other
> > > > > warnings? I'm currently stumped.
> > > >
> > > > That is because __rel_loc is used only in the sample code in the kernel
> > > > for testing. Other use-cases comes from user-space.
> > > > Hmm, can we skip this boundary check for this example?
> > >
> > > If the -Warray-bounds determines the destination array size from
> > > the type of given pointer, we can just change the macro as below;
> > >
> > > #define __get_rel_dynamic_array(field)
> > > ((void *)__entry + \
> > > offsetof(typeof(*__entry), __rel_loc_##field) + \
> > > sizeof(__entry->__rel_loc_##field) + \
> > > (__entry->__rel_loc_##field & 0xffff))
> > >
> > > This must works same as __get_dynamic_array() macro.
> > >
> > > Could you try this patch?
> > >
> > > From 2982ba01367ec1f746a4f128512436e5325a7f9d Mon Sep 17 00:00:00 2001
> > > From: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
> > > Date: Tue, 25 Jan 2022 23:19:30 +0900
> > > Subject: [PATCH] tracing: Avoid -Warray-bounds warning for __rel_loc macro
> > >
> > > Since -Warray-bounds checks the destination size from the
> > > type of given pointer, __assign_rel_str() macro gets warned
> > > because it passes the pointer to the 'u32' field instead of
> > > 'trace_event_raw_*' data structure.
> > > Pass the data address calculated from the 'trace_event_raw_*'
> > > instead of 'u32' __rel_loc field.
> > >
> > > Signed-off-by: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
> > > Reported-by: Stephen Rothwell <sfr@xxxxxxxxxxxxxxxx>
> > > Cc: Kees Cook <keescook@xxxxxxxxxxxx>
> > > ---
> > > include/trace/trace_events.h | 7 ++++---
> > > 1 file changed, 4 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/include/trace/trace_events.h b/include/trace/trace_events.h
> > > index 8c6f7c433518..65d927e059d3 100644
> > > --- a/include/trace/trace_events.h
> > > +++ b/include/trace/trace_events.h
> > > @@ -318,9 +318,10 @@ TRACE_MAKE_SYSTEM_STR();
> > > #define __get_str(field) ((char *)__get_dynamic_array(field))
> > >
> > > #undef __get_rel_dynamic_array
> > > -#define __get_rel_dynamic_array(field) \
> > > - ((void *)(&__entry->__rel_loc_##field) + \
> > > - sizeof(__entry->__rel_loc_##field) + \
> > > +#define __get_rel_dynamic_array(field) \
> > > + ((void *)__entry + \
> > > + offsetof(typeof(*__entry), __rel_loc_##field) + \
> > > + sizeof(__entry->__rel_loc_##field) + \
> > > (__entry->__rel_loc_##field & 0xffff))
> > >
> > > #undef __get_rel_dynamic_array_len
> >
> > This patch doesn't silence the warning, but now that I see the shape of
> > things more clearly, let me see if I can find the right combo.
>
> Hmm, could the zero size array cause an issues here. That is, does this
> help?
>
> diff --git a/include/trace/trace_events.h b/include/trace/trace_events.h
> index 65d927e059d3..3d29919045af 100644
> --- a/include/trace/trace_events.h
> +++ b/include/trace/trace_events.h
> @@ -128,7 +128,7 @@ TRACE_MAKE_SYSTEM_STR();
> struct trace_event_raw_##name { \
> struct trace_entry ent; \
> tstruct \
> - char __data[0]; \
> + char __data[]; \
> }; \
> \
> static struct trace_event_class event_class_##name;
I changed this too, just to future-proof it, and in an attempt to base
the address off of __data[] (which turned out not to be needed).
It turns out that there was still a version of the __get_rel_dyanmic_array
in perf.h that was the actual culprit.
--
Kees Cook
