Re: [PATCH 0/2] tracing: Detect unsafe dereferencing of pointers from trace events

[Steven Rostedt <rostedt@goodmis.org>]

On Fri, 26 Feb 2021 14:21:00 -0800
Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Fri, Feb 26, 2021 at 11:07 AM Steven Rostedt <rostedt@goodmis.org> wrote:
> >
> > The first patch scans the print fmts of the trace events looking for
> > dereferencing pointers from %p*, and making sure that they refer back
> > to the trace event itself.
> >
> > The second patch handles strings "%s" [..]  
> 
> Doing this at runtime really feels like the wrong thing to do.
> 
> It won't even protect us from what happened - people like me and
> Andrew won't even run those tracepoints in the first place, so we
> won't notice.
> 
> It really would be much better in every respect to have this done by
> checkpatch, I think.

And after fixing the parsing to not trigger false positives, an
allyesconfig boot found this:

event cdns3_gadget_giveback has unsafe dereference of argument 11
print_fmt: "%s: req: %p, req buff %p, length: %u/%u %s%s%s, status: %d, trb: [start:%d, end:%d: virt addr %pa], flags:%x SID: %u", __get_str(name), REC->req, REC->buf,
 REC->actual, REC->length, REC->zero ? "Z" : "z", REC->short_not_ok ? "S" : "s", REC->no_interrupt ? "I" : "i", REC->status, REC->start_trb, REC->end_trb, REC->start_trb_addr, REC->flags, RE
C->stream_id

(as the above is from a trace event class, it triggered for every event
in that class).

As it looks like it uses %pa which IIUC from the printk code, it
dereferences the pointer to find it's virtual address. The event has
this as the field:

                __field(struct cdns3_trb *, start_trb_addr)

Assigns it with:

                __entry->start_trb_addr = req->trb;

And prints that with %pa, which will dereference pointer at the time of
reading, where the address in question may no longer be around. That
looks to me as a potential bug.

[ Cc'd the people responsible for that code. ]

-- Steve