Re: [PATCH v2 8/8] perf record: Directly bail out for compat case

From: Leo Yan
Date: Wed Jun 09 2021 - 04:58:05 EST


Hi Adrian,

On Wed, Jun 09, 2021 at 11:23:25AM +0300, Adrian Hunter wrote:

[...]

> >> I was thinking we would separate out the compat case:
> >>
> >> #if BITS_PER_LONG == 32
> >> if (kernel_is_64_bit)
> >> return compat_auxtrace_mmap__[read_head/write_tail]()
> >> #endif
> >>
> >> So the non-compat cases would not be affected.
> >
> > Because I don't want to introduce the complexity for read/write head
> > and tail, and we also need to handle the same issue for the perf ring
> > buffer. So how about below change?
> >
> > The main idea for below change is it allows the perf to run normally
> > on the compat mode and exitly if detects the buffer head is close to
> > the low 32-bit's overflow: when detect the low 32-bit value is bigger
> > than 0xf0000000 (so we have 256MiB margin to the overflow), it reports
> > error and exit.
> >
> > diff --git a/tools/perf/util/auxtrace.c b/tools/perf/util/auxtrace.c
> > index 1b4091a3b508..2a9965bfeab4 100644
> > --- a/tools/perf/util/auxtrace.c
> > +++ b/tools/perf/util/auxtrace.c
> > @@ -1693,6 +1693,14 @@ static int __auxtrace_mmap__read(struct mmap *map,
> > pr_debug3("auxtrace idx %d old %#"PRIx64" head %#"PRIx64" diff %#"PRIx64"\n",
> > mm->idx, old, head, head - old);
> >
> > +#ifdef BITS_PER_LONG == 32
> > + if (kernel_is_64bit() && head >= 0xf0000000) {
>
> You are assuming the head never increases by more than 256MiB which
> means you should limit the buffer size to 256MiB maximum.
>
> To me this seems a bit too far from an ideal solution.
>
> I would have thought separating out the compat case makes things
> simpler to understand.

Agreed. I will follow up the suggestions to add compat variants for
accessing AUX head and tail, and will distinguish compat case with
global env variable for 64-bit kernel.

After get ready, will send out for review. Thanks a lot for suggestions!

Leo