[RFC PATCH v1] powerpc: Enable KFENCE for PPC32
Marco Elver
elver at google.com
Fri Mar 5 20:14:23 AEDT 2021
On Fri, 5 Mar 2021 at 09:23, Christophe Leroy
<christophe.leroy at csgroup.eu> wrote:
>
>
>
> Le 05/03/2021 à 08:50, Marco Elver a écrit :
> > On Fri, Mar 05, 2021 at 04:01PM +1100, Michael Ellerman wrote:
> >> Marco Elver <elver at google.com> writes:
> >>> On Thu, Mar 04, 2021 at 12:48PM +0100, Christophe Leroy wrote:
> >>>> Le 04/03/2021 à 12:31, Marco Elver a écrit :
> >>>>> On Thu, 4 Mar 2021 at 12:23, Christophe Leroy
> >>>>> <christophe.leroy at csgroup.eu> wrote:
> >>>>>> Le 03/03/2021 à 11:56, Marco Elver a écrit :
> >>>>>>>
> >>>>>>> Somewhat tangentially, I also note that e.g. show_regs(regs) (which
> >>>>>>> was printed along the KFENCE report above) didn't include the top
> >>>>>>> frame in the "Call Trace", so this assumption is definitely not
> >>>>>>> isolated to KFENCE.
> >>>>>>>
> >>>>>>
> >>>>>> Now, I have tested PPC64 (with the patch I sent yesterday to modify save_stack_trace_regs()
> >>>>>> applied), and I get many failures. Any idea ?
> >>>>>>
> >>>>>> [ 17.653751][ T58] ==================================================================
> >>>>>> [ 17.654379][ T58] BUG: KFENCE: invalid free in .kfence_guarded_free+0x2e4/0x530
> >>>>>> [ 17.654379][ T58]
> >>>>>> [ 17.654831][ T58] Invalid free of 0xc00000003c9c0000 (in kfence-#77):
> >>>>>> [ 17.655358][ T58] .kfence_guarded_free+0x2e4/0x530
> >>>>>> [ 17.655775][ T58] .__slab_free+0x320/0x5a0
> >>>>>> [ 17.656039][ T58] .test_double_free+0xe0/0x198
> >>>>>> [ 17.656308][ T58] .kunit_try_run_case+0x80/0x110
> >>>>>> [ 17.656523][ T58] .kunit_generic_run_threadfn_adapter+0x38/0x50
> >>>>>> [ 17.657161][ T58] .kthread+0x18c/0x1a0
> >>>>>> [ 17.659148][ T58] .ret_from_kernel_thread+0x58/0x70
> >>>>>> [ 17.659869][ T58]
> >>> [...]
> >>>>>
> >>>>> Looks like something is prepending '.' to function names. We expect
> >>>>> the function name to appear as-is, e.g. "kfence_guarded_free",
> >>>>> "test_double_free", etc.
> >>>>>
> >>>>> Is there something special on ppc64, where the '.' is some convention?
> >>>>>
> >>>>
> >>>> I think so, see https://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi.html#FUNC-DES
> >>>>
> >>>> Also see commit https://github.com/linuxppc/linux/commit/02424d896
> >>>
> >>> Thanks -- could you try the below patch? You'll need to define
> >>> ARCH_FUNC_PREFIX accordingly.
> >>>
> >>> We think, since there are only very few architectures that add a prefix,
> >>> requiring <asm/kfence.h> to define something like ARCH_FUNC_PREFIX is
> >>> the simplest option. Let me know if this works for you.
> >>>
> >>> There an alternative option, which is to dynamically figure out the
> >>> prefix, but if this simpler option is fine with you, we'd prefer it.
> >>
> >> We have rediscovered this problem in basically every tracing / debugging
> >> feature added in the last 20 years :)
> >>
> >> I think the simplest solution is the one tools/perf/util/symbol.c uses,
> >> which is to just skip a leading '.'.
> >>
> >> Does that work?
> >>
> >> diff --git a/mm/kfence/report.c b/mm/kfence/report.c
> >> index ab83d5a59bb1..67b49dc54b38 100644
> >> --- a/mm/kfence/report.c
> >> +++ b/mm/kfence/report.c
> >> @@ -67,6 +67,9 @@ static int get_stack_skipnr(const unsigned long stack_entries[], int num_entries
> >> for (skipnr = 0; skipnr < num_entries; skipnr++) {
> >> int len = scnprintf(buf, sizeof(buf), "%ps", (void *)stack_entries[skipnr]);
> >>
> >> + if (buf[0] == '.')
> >> + buf++;
> >> +
> >
> > Unfortunately this does not work, since buf is an array. We'd need an
> > offset, and it should be determined outside the loop. I had a solution
> > like this, but it turned out quite complex (see below). And since most
> > architectures do not require this, decided that the safest option is to
> > use the macro approach with ARCH_FUNC_PREFIX, for which Christophe
> > already prepared a patch and tested:
> > https://lore.kernel.org/linux-mm/20210304144000.1148590-1-elver@google.com/
> > https://lkml.kernel.org/r/afaec81a551ef15345cb7d7563b3fac3d7041c3a.1614868445.git.christophe.leroy@csgroup.eu
> >
> > Since KFENCE requires <asm/kfence.h> anyway, we'd prefer this approach
> > (vs. dynamically detecting).
> >
> > Thanks,
> > -- Marco
> >
>
> What about
Sure something like that would work. But I explicitly did *not* want
to hard-code the '.' in non-arch code.
The choice is between:
1. ARCH_FUNC_PREFIX (as a matter of fact, the ARCH_FUNC_PREFIX patch
is already in -mm). Perhaps we could optimize it further, by checking
ARCH_FUNC_PREFIX in buf, and advancing buf like you propose, but I'm
not sure it's worth worrying about.
2. The dynamic solution that I proposed that does not use a hard-coded
'.' (or some variation thereof).
Please tell me which solution you prefer, 1 or 2 -- I'd like to stop
bikeshedding here. If there's a compelling argument for hard-coding
the '.' in non-arch code, please clarify, but otherwise I'd like to
keep arch-specific things out of generic code.
Thanks.
> diff --git a/mm/kfence/report.c b/mm/kfence/report.c
> index 519f037720f5..5e196625fb34 100644
> --- a/mm/kfence/report.c
> +++ b/mm/kfence/report.c
> @@ -43,7 +43,7 @@ static void seq_con_printf(struct seq_file *seq, const char *fmt, ...)
> static int get_stack_skipnr(const unsigned long stack_entries[], int num_entries,
> const enum kfence_error_type *type)
> {
> - char buf[64];
> + char _buf[64];
> int skipnr, fallback = 0;
>
> if (type) {
> @@ -65,7 +65,11 @@ static int get_stack_skipnr(const unsigned long stack_entries[], int num_entries
> }
>
> for (skipnr = 0; skipnr < num_entries; skipnr++) {
> - int len = scnprintf(buf, sizeof(buf), "%ps", (void *)stack_entries[skipnr]);
> + char *buf = _buf;
> + int len = scnprintf(_buf, sizeof(_buf), "%ps", (void *)stack_entries[skipnr]);
> +
> + if (_buf[0] == '.')
> + buf++, len--;
>
> if (str_has_prefix(buf, "kfence_") || str_has_prefix(buf, "__kfence_") ||
> !strncmp(buf, "__slab_free", len)) {
> ---
>
> Christophe
More information about the Linuxppc-dev
mailing list