Re: [GIT PULL] tracing: Updates for 5.2

From: Linus Torvalds
Date: Wed May 15 2019 - 19:34:03 EST

On Wed, May 15, 2019 at 10:36 AM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> The major changes in this tracing update includes:

This is not directly related to this pull request, but newer versions
of gcc hate your trace_iterator clearing trick.

This code:

/* reset all but tr, trace, and overruns */
memset(&iter.seq, 0,
sizeof(struct trace_iterator) -
offsetof(struct trace_iterator, seq));

not only has a completely misleading comment (it resets a lot more
than the comment states), but modern gcc looks at that code and says
"oh, you're passing it a pointer to 'iter.seq', but then clearing a
lot more than a 'trace_seq'":

In function âmemsetâ,
inlined from âftrace_dumpâ at kernel/trace/trace.c:8914:3:
/include/linux/string.h:344:9: warning: â__builtin_memsetâ offset
[8505, 8560] from the object at âiterâ is out of the bounds of
referenced subobject âseqâ with type âstruct trace_seqâ at offset 4368
344 | return __builtin_memset(p, c, size);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~

It's a somewhat annoying warning because the code itself is
technically correct, but at the same time, I think the gcc warning is
reasonable. You *are* passing it a 'struct trace_seq' pointer, and
then you're clearing a whole lot more than that.

One option is to just rewrite it something like

const unsigned int offset = offsetof(struct trace_iterator, seq);
memset(offset+(void *)&iter, 0, sizeof(iter) - offset);

which should compile cleanly - because now you're doing the memset on
a part of the much bigger 'iter' structure, not on one member (and
overflowing that one member).

Another option might be to separate the zeroed part of the structure
into a sub-structure of its own, and then just use

memset(&iter.sub, 0, sizeof(iter.sub));

but then you'd obviously have to change all the uses of the sub-fields..