On Thu, Nov 05, 2020 at 08:19:47PM -0800, Alexei Starovoitov wrote:
> > Subject: [PATCH] Update perf ring buffer to prevent corruption from
> > bpf_perf_output_event()
$Subject is broken, it lacks subsystem prefix.
> >
> > The bpf_perf_output_event() helper takes a sample size parameter of u64, but
> > the underlying perf ring buffer uses a u16 internally. This 64KB maximum size
> > has to also accommodate a variable sized header. Failure to observe this
> > restriction can result in corruption of the perf ring buffer as samples
> > overlap.
> >
> > Track the sample size and return -E2BIG if too big to fit into the u16
> > size parameter.
> >
> > Signed-off-by: Kevin Sheldrake <kevin.sheldrake@xxxxxxxxxxxxx>
> > ---
> > include/linux/perf_event.h | 2 +-
> > kernel/events/core.c | 40 ++++++++++++++++++++++++++--------------
> > 2 files changed, 27 insertions(+), 15 deletions(-)
> >
> > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> > index 0c19d27..b9802e5 100644
> > --- a/include/linux/perf_event.h
> > +++ b/include/linux/perf_event.h
> > @@ -1060,7 +1060,7 @@ extern void perf_output_sample(struct perf_output_handle *handle,
> > struct perf_event_header *header,
> > struct perf_sample_data *data,
> > struct perf_event *event);
> > -extern void perf_prepare_sample(struct perf_event_header *header,
> > +extern int perf_prepare_sample(struct perf_event_header *header,
> > struct perf_sample_data *data,
> > struct perf_event *event,
> > struct pt_regs *regs);
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index da467e1..c6c4a3c 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -7016,15 +7016,17 @@ perf_callchain(struct perf_event *event, struct pt_regs *regs)
> > return callchain ?: &__empty_callchain;
> > }
> >
> > -void perf_prepare_sample(struct perf_event_header *header,
> > +int perf_prepare_sample(struct perf_event_header *header,
> > struct perf_sample_data *data,
> > struct perf_event *event,
> > struct pt_regs *regs)
please re-align things.
> > {
> > u64 sample_type = event->attr.sample_type;
> > + u32 header_size = header->size;
> > +
> >
> > header->type = PERF_RECORD_SAMPLE;
> > - header->size = sizeof(*header) + event->header_size;
> > + header_size = sizeof(*header) + event->header_size;
> >
> > header->misc = 0;
> > header->misc |= perf_misc_flags(regs);
> > @@ -7042,7 +7044,7 @@ void perf_prepare_sample(struct perf_event_header *header,
> >
> > size += data->callchain->nr;
> >
> > - header->size += size * sizeof(u64);
> > + header_size += size * sizeof(u64);
> > }
> >
> > if (sample_type & PERF_SAMPLE_RAW) {
> > @@ -7067,7 +7069,7 @@ void perf_prepare_sample(struct perf_event_header *header,
> > size = sizeof(u64);
> > }
> >
> > - header->size += size;
> > + header_size += size;
> > }
AFAICT perf_raw_frag::size is a u32, so the above addition can already
fully overflow. Best is probably to make header_size u64 and delay that
until the final tally below.
> >
> > if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
> > @@ -7162,14 +7164,20 @@ void perf_prepare_sample(struct perf_event_header *header,
> > * Make sure this doesn't happen by using up to U16_MAX bytes
> > * per sample in total (rounded down to 8 byte boundary).
> > */
> > - size = min_t(size_t, U16_MAX - header->size,
> > + size = min_t(size_t, U16_MAX - header_size,
> > event->attr.aux_sample_size);
> > size = rounddown(size, 8);
> > size = perf_prepare_sample_aux(event, data, size);
> >
> > - WARN_ON_ONCE(size + header->size > U16_MAX);
> > - header->size += size;
> > + WARN_ON_ONCE(size + header_size > U16_MAX);
> > + header_size += size;
> > }
> > +
> > + if (header_size > U16_MAX)
> > + return -E2BIG;
> > +
> > + header->size = header_size;
> > +
> > /*
> > * If you're adding more sample types here, you likely need to do
> > * something about the overflowing header::size, like repurpose the
> > @@ -7179,6 +7187,8 @@ void perf_prepare_sample(struct perf_event_header *header,
> > * do here next.
> > */
> > WARN_ON_ONCE(header->size & 7);
> > +
> > + return 0;
> > }
> >
> > static __always_inline int
> > @@ -7196,7 +7206,9 @@ __perf_event_output(struct perf_event *event,
> > /* protect the callchain buffers */
> > rcu_read_lock();
> >
> > - perf_prepare_sample(&header, data, event, regs);
> > + err = perf_prepare_sample(&header, data, event, regs);
> > + if (err)
> > + goto exit;
This is wrong I think. The thing is that when output_begin() below
returns an error, there either is no buffer (in which case we can't do
anything much at all) or it will have incremented rb->lost.
This OTOH will completely fail to report the loss. The error case here
is to immediately try and emit a RECORD_LOST event, but then please also
consider these patches:
https://lkml.kernel.org/r/20201030151345.540479897@xxxxxxxxxxxxx
(which I'll be pushing into tip/perf/urgent soonish)
> >
> > err = output_begin(&handle, event, header.size);
> > if (err)
> > --
> > 2.7.4
> >
