Re: [Patch bpf-next] bpf: clear per_cpu pointers in bpf_prog_clone_create()

On Wed, Feb 17, 2021 at 2:01 PM Daniel Borkmann [off-list ref] wrote:

On 2/17/21 4:58 AM, Cong Wang wrote:

quoted

From: Cong Wang <redacted>

Pretty much similar to commit 1336c662474e
("bpf: Clear per_cpu pointers during bpf_prog_realloc") we also need to
clear these two percpu pointers in bpf_prog_clone_create(), otherwise
would get a double free:

  BUG: kernel NULL pointer dereference, address: 0000000000000000
  #PF: supervisor read access in kernel mode
  #PF: error_code(0x0000) - not-present page
  PGD 0 P4D 0
  Oops: 0000 [#1] SMP PTI
  CPU: 13 PID: 8140 Comm: kworker/13:247 Kdump: loaded Tainted: G         W   OE
 5.11.0-rc4.bm.1-amd64+ #1
  Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
  test_bpf: #1 TXA
  Workqueue: events bpf_prog_free_deferred
  RIP: 0010:percpu_ref_get_many.constprop.97+0x42/0xf0
  Code: [...]
  RSP: 0018:ffffa6bce1f9bda0 EFLAGS: 00010002
  RAX: 0000000000000001 RBX: 0000000000000000 RCX: 00000000021dfc7b
  RDX: ffffffffae2eeb90 RSI: 867f92637e338da5 RDI: 0000000000000046
  RBP: ffffa6bce1f9bda8 R08: 0000000000000000 R09: 0000000000000001
  R10: 0000000000000046 R11: 0000000000000000 R12: 0000000000000280
  R13: 0000000000000000 R14: 0000000000000000 R15: ffff9b5f3ffdedc0
  FS:   0000000000000000(0000) GS:ffff9b5f2fb40000(0000) knlGS:0000000000000000
  CS:   0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 0000000000000000 CR3: 000000027c36c002 CR4: 00000000003706e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
  refill_obj_stock+0x5e/0xd0
  free_percpu+0xee/0x550
  __bpf_prog_free+0x4d/0x60
  process_one_work+0x26a/0x590
  worker_thread+0x3c/0x390
  ? process_one_work+0x590/0x590
  kthread+0x130/0x150
  ? kthread_park+0x80/0x80
  ret_from_fork+0x1f/0x30

This bug is 100% reproducible with test_kmod.sh.

Reported-by: Jiang Wang <redacted>
Fixes: 700d4796ef59 ("bpf: Optimize program stats")
Fixes: ca06f55b9002 ("bpf: Add per-program recursion prevention mechanism")
Cc: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Cong Wang <redacted>
---
  kernel/bpf/core.c | 2 ++
  1 file changed, 2 insertions(+)

diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
index 0ae015ad1e05..b0c11532e535 100644
--- a/kernel/bpf/core.c
+++ b/kernel/bpf/core.c

@@ -1103,6 +1103,8 @@ static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
               * this still needs to be adapted.
               */
              memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
+             fp_other->stats = NULL;
+             fp_other->active = NULL;
      }

      return fp;

This is not correct. I presume if you enable blinding and stats, then this will still

Well, at least I ran all BPF selftests and found no crash. (Before my patch, the
crash happened 100%.)

crash. The proper way to fix it is to NULL these pointers in bpf_prog_clone_free()
since the clone can be promoted as the actual prog and the prog ptr released instead.

Not sure if I understand your point, but what I cleared is fp_other,
which is the original, not the clone. And of course, the original would
be overriden:

        tmp = bpf_jit_blind_constants(prog);
        if (IS_ERR(tmp))
                return orig_prog;
        if (tmp != prog) {
                tmp_blinded = true;
                prog = tmp;  // <=== HERE
        }

I think this is precisely why the crash does not happen after my patch.

However, it does seem to me patching bpf_prog_clone_free() is better,
as there would be no assumption on using the original. All I want to
say here is that both ways could fix the crash, which one is better is
arguable.

Thanks.