On Thu, Mar 05, 2026 at 01:45:44PM +0100, Viktor Malik wrote:
> On 3/3/26 15:58, Saket Kumar Bhaskar wrote:
> > On Fri, Feb 27, 2026 at 09:25:02AM +0100, Viktor Malik wrote:
> >> It may happen that mm is already released, which leads to kernel panic.
> >> This adds the NULL check for current->mm, similarly to 20afc60f892d
> >> ("x86, perf: Check that current->mm is alive before getting user
> >> callchain").
> >>
> >> I was getting this panic when running a profiling BPF program
> >> (profile.py from bcc-tools):
> >>
> >> [26215.051935] Kernel attempted to read user page (588) - exploit
> >> attempt? (uid: 0)
> >> [26215.051950] BUG: Kernel NULL pointer dereference on read at
> >> 0x00000588
> >> [26215.051952] Faulting instruction address: 0xc00000000020fac0
> >> [26215.051957] Oops: Kernel access of bad area, sig: 11 [#1]
> >> [...]
> >> [26215.052049] Call Trace:
> >> [26215.052050] [c000000061da6d30] [c00000000020fc10]
> >> perf_callchain_user_64+0x2d0/0x490 (unreliable)
> >> [26215.052054] [c000000061da6dc0] [c00000000020f92c]
> >> perf_callchain_user+0x1c/0x30
> >> [26215.052057] [c000000061da6de0] [c0000000005ab2a0]
> >> get_perf_callchain+0x100/0x360
> >> [26215.052063] [c000000061da6e70] [c000000000573bc8]
> >> bpf_get_stackid+0x88/0xf0
> >> [26215.052067] [c000000061da6ea0] [c008000000042258]
> >> bpf_prog_16d4ab9ab662f669_do_perf_event+0xf8/0x274
> >> [...]
> >>
> >> Fixes: 20002ded4d93 ("perf_counter: powerpc: Add callchain support")
> >> Signed-off-by: Viktor Malik <[email protected]>
> >> ---
> >> arch/powerpc/perf/callchain_32.c | 3 +++
> >> arch/powerpc/perf/callchain_64.c | 3 +++
> >> 2 files changed, 6 insertions(+)
> >>
> >> diff --git a/arch/powerpc/perf/callchain_32.c
> >> b/arch/powerpc/perf/callchain_32.c
> >> index ddcc2d8aa64a..b46e21679566 100644
> >> --- a/arch/powerpc/perf/callchain_32.c
> >> +++ b/arch/powerpc/perf/callchain_32.c
> >> @@ -144,6 +144,9 @@ void perf_callchain_user_32(struct
> >> perf_callchain_entry_ctx *entry,
> >> sp = regs->gpr[1];
> >> perf_callchain_store(entry, next_ip);
> >>
> >> + if (!current->mm)
> >> + return;
> >> +
> >> while (entry->nr < entry->max_stack) {
> >> fp = (unsigned int __user *) (unsigned long) sp;
> >> if (invalid_user_sp(sp) || read_user_stack_32(fp, &next_sp))
> >> diff --git a/arch/powerpc/perf/callchain_64.c
> >> b/arch/powerpc/perf/callchain_64.c
> >> index 115d1c105e8a..eaaadd6fa81b 100644
> >> --- a/arch/powerpc/perf/callchain_64.c
> >> +++ b/arch/powerpc/perf/callchain_64.c
> >> @@ -79,6 +79,9 @@ void perf_callchain_user_64(struct
> >> perf_callchain_entry_ctx *entry,
> >> sp = regs->gpr[1];
> >> perf_callchain_store(entry, next_ip);
> >>
> >> + if (!current->mm)
> >> + return;
> >> +
> >> while (entry->nr < entry->max_stack) {
> >> fp = (unsigned long __user *) sp;
> >> if (invalid_user_sp(sp) || read_user_stack_64(fp, &next_sp))
> >> --
> >> 2.53.0
> >>
> > Sorry, I missed adding cc list for the last conversation so adding this for
> > reference:
> >
> >> Wouldn't be good if we check this in perf_callchain_user() as it will
> >> cover both cases.
> >
> > to which Viktor replied:
> > I considered it but in that case, we'd also miss the top-level stack
> > frame (the perf_callchain_store call above). Other arches include it so
> > I followed the behavior for powerpc.
> >
> > Viktor, agreed with your first point. I have another concern:
> >
> > I was hitting this issue with stacktrace_build_id_nmi in bpf and
> > applied this patch
> > https://lore.kernel.org/bpf/[email protected]/T/#mf901967ebe77506f1bd6e3d876c2a85824d9519d
> >
> > Wondering if the above generic fix is working do we need to add this
> > check in powerpc specific code?
>
> I tried to apply that patch series but, unfortunately, keep getting the
> panic when running the BCC profile tool.
>
> Also, looking at the patch, it seems that it would only solve the issue
> when perf_callchain_user is called from a BPF context, however, I assume
> that it may be called from other contexts, too.
>
> Since perf_callchain_user_{32,64} are dereferencing current->mm while
> walking the stack, I think that an explicit protection against
> current->mm being NULL makes sense here, even in the presence of the
> above patch. Especially since other arches have it, too.
>
> Viktor
>
Ok that looks convincing then, another thing is that, how about moving
perf_callchain_store
to perf_callchain_user and checking current->mm == NULL there for both
perf_callchain_user_32/64.
next_ip, lr and sp can be passed to perf_callchain_user_32/64.
Thanks,
Saket
