On 8/1/23 1:36 PM, Dave Marchevsky wrote:
This series is the third of three (or more) followups to address issues
in the bpf_refcount shared ownership implementation discovered by Kumar.
This series addresses the use-after-free scenario described in [0]. The
first followup series ([1]) also attempted to address the same
use-after-free, but only got rid of the splat without addressing the
underlying issue. After this series the underyling issue is fixed and
bpf_refcount_acquire can be re-enabled.
The main fix here is migration of bpf_obj_drop to use
bpf_mem_free_rcu. To understand why this fixes the issue, let us consider
the example interleaving provided by Kumar in [0]:
CPU 0 CPU 1
n = bpf_obj_new
lock(lock1)
bpf_rbtree_add(rbtree1, n)
m = bpf_rbtree_acquire(n)
unlock(lock1)
kptr_xchg(map, m) // move to map
// at this point, refcount = 2
m = kptr_xchg(map, NULL)
lock(lock2)
lock(lock1) bpf_rbtree_add(rbtree2, m)
On the right column: bpf_rbtree_add(rbtree1, m) ?
p = bpf_rbtree_first(rbtree1) if (!RB_EMPTY_NODE) bpf_obj_drop_impl(m) // A
bpf_rbtree_remove(rbtree1, p)
unlock(lock1)
bpf_obj_drop(p) // B
bpf_refcount_acquire(m) // use-after-free
...
Before this series, bpf_obj_drop returns memory to the allocator using
bpf_mem_free. At this point (B in the example) there might be some
non-owning references to that memory which the verifier believes are valid,
but where the underlying memory was reused for some other allocation.
Commit 7793fc3babe9 ("bpf: Make bpf_refcount_acquire fallible for
non-owning refs") attempted to fix this by doing refcount_inc_non_zero
on refcount_acquire in instead of refcount_inc under the assumption that
preventing erroneous incr-on-0 would be sufficient. This isn't true,
though: refcount_inc_non_zero must *check* if the refcount is zero, and
the memory it's checking could have been reused, so the check may look
at and incr random reused bytes.
If we wait to reuse this memory until all non-owning refs that could
point to it are gone, there is no possibility of this scenario
happening. Migrating bpf_obj_drop to use bpf_mem_free_rcu for refcounted
nodes accomplishes this.
For such nodes, the validity of their underlying memory is now tied to
RCU Tasks Trace critical section. This matches MEM_RCU trustedness
expectations, so the series takes the opportunity to more explicitly
mark this trustedness state.
The functional effects of trustedness changes here are rather small.
This is largely due to local kptrs having separate verifier handling -
with implicit trustedness assumptions - than arbitrary kptrs.
Regardless, let's take the opportunity to move towards a world where
trustedness is more explictly handled.
Summary of patch contents, with sub-bullets being leading questions and
comments I think are worth reviewer attention:
* Patches 1 and 2 are moreso documententation - and
enforcement, in patch 1's case - of existing semantics / expectations
* Patch 3 changes bpf_obj_drop behavior for refcounted nodes such that
their underlying memory is not reused until RCU grace period elapses
* Perhaps it makes sense to move to mem_free_rcu for _all_
non-owning refs in the future, not just refcounted. This might
allow custom non-owning ref lifetime + invalidation logic to be
entirely subsumed by MEM_RCU handling. IMO this needs a bit more
thought and should be tackled outside of a fix series, so it's not
attempted here.
* Patch 4 re-enables bpf_refcount_acquire as changes in patch 3 fix
the remaining use-after-free
* One might expect this patch to be last in the series, or last
before selftest changes. Patches 5 and 6 don't change
verification or runtime behavior for existing BPF progs, though.
* Patch 5 brings the verifier's understanding of refcounted node
trustedness in line with Patch 4's changes
* Patch 6 allows some bpf_spin_{lock, unlock} calls in sleepable
progs. Marked RFC for a few reasons:
* bpf_spin_{lock,unlock} haven't been usable in sleepable progs
since before the introduction of bpf linked list and rbtree. As
such this feels more like a new feature that may not belong in
this fixes series.
* If we do want to allow bpf_spin_{lock,unlock} in sleepable progs,
Alexei has expressed a preference for that do be done as part of a
broader set of changes to verifier determination of where those
helpers can be called, and what can be called inside the spin_lock
CS.
* I'm unsure whether preemption needs to be disabled in this patch
as well.
* Patch 7 adds tests
[0]: https://lore.kernel.org/bpf/atfviesiidev4hu53hzravmtlau3wdodm2vqs7rd7tnwft34e3@xktodqeqevir/
[1]: https://lore.kernel.org/bpf/20230602022647.1571784-1-davemarchevsky@xxxxxx/
Dave Marchevsky (7):
bpf: Ensure kptr_struct_meta is non-NULL for collection insert and
refcount_acquire
bpf: Consider non-owning refs trusted
bpf: Use bpf_mem_free_rcu when bpf_obj_dropping refcounted nodes
bpf: Reenable bpf_refcount_acquire
bpf: Consider non-owning refs to refcounted nodes RCU protected
[RFC] bpf: Allow bpf_spin_{lock,unlock} in sleepable prog's RCU CS
selftests/bpf: Add tests for rbtree API interaction in sleepable progs
include/linux/bpf.h | 3 +-
include/linux/bpf_verifier.h | 2 +-
kernel/bpf/helpers.c | 6 ++-
kernel/bpf/verifier.c | 45 ++++++++++++-----
.../bpf/prog_tests/refcounted_kptr.c | 26 ++++++++++
.../selftests/bpf/progs/refcounted_kptr.c | 37 ++++++++++++++
.../bpf/progs/refcounted_kptr_fail.c | 48 +++++++++++++++++++
7 files changed, 153 insertions(+), 14 deletions(-)
