All of lore.kernel.org
 help / color / mirror / Atom feed
* [RFC PATCH 6.18.y 0/1] bpf: reject NULL arguments to four kfuncs
@ 2026-11-18 10:17 asatsuyu.liu
  2026-11-18 10:17 ` [RFC PATCH 6.18.y 1/1] bpf: Reject " asatsuyu.liu
  2026-11-18 10:40 ` [RFC PATCH 6.18.y 0/1] bpf: reject " Greg KH
  0 siblings, 2 replies; 4+ messages in thread
From: asatsuyu.liu @ 2026-11-18 10:17 UTC (permalink / raw)
  To: stable; +Cc: bpf, ast, daniel, puranjay, vadim.fedorenko, Yi Liu

From: Yi Liu <asatsuyu.liu@gmail.com>

Hi,

6.18.64 still accepts NULL and PTR_MAYBE_NULL arguments for four kfuncs
whose implementations dereference them.  Mainline fixed this class in
v7.0 with

  1a5c01d2508a ("bpf: Make KF_TRUSTED_ARGS the default for all kfuncs")

but that commit is a broad verifier behavior change, it was not
tagged for stable, and it is not in 6.18.y or 6.19.y.  6.18.64 is the
current 6.18.y tip; 6.19.14 is the current 6.19.y tip.

The four deterministic reproducers on 6.18.64 are:

  1. bpf_copy_from_user_str(dst=NULL, dst__sz=8, src=NULL,
                            BPF_F_PAD_ZEROS)
     -> memset(NULL, 0, 8) in bpf_copy_from_user_str
        RIP: memset_orig+0x72, call trace bpf_copy_from_user_str+0x69

  2. bpf_copy_from_user_task_str(dst=NULL, dst__sz=16, src=NULL,
                                 tsk=current, BPF_F_PAD_ZEROS)
     -> copy_remote_vm_str() writes to dst
        RIP: copy_remote_vm_str+0x45,
        call trace bpf_copy_from_user_task_str+0x2f

  3. bpf_crypto_ctx_create(params, params__sz=0, err=NULL)
     -> *err = -EINVAL
        RIP: bpf_crypto_ctx_create+0x1f8
        faulting insn: c7 45 00 ea ff ff ff  (movl $-22, 0(%rbp))

  4. bpf_rbtree_remove(root, node=bpf_rbtree_first(root))
     with an empty tree; first() returns NULL and remove() reads
     node_internal->owner
        RIP: bpf_rbtree_remove+0xb, address 0x18
        faulting insn: 48 8b 46 18  (mov 0x18(%rsi), %rax, RSI=0)

All four are LOAD_OK on 6.18.64; the four safe controls (non-NULL
dst/err, valid add+remove) load and run.

The 6.18.64 kernel is built from commit fd7cfd39c885 and the QEMU
tests use oops=panic panic=1.

This is a local DoS from a context that can load the relevant
program type.  It requires CAP_BPF/CAP_SYS_ADMIN (the tc/rbtree
reproducer also needs CAP_NET_ADMIN), so it is not a privilege
escalation and I have not demonstrated an unprivileged path.  With
panic_on_oops it turns into a kernel panic.

The following patch adds KF_TRUSTED_ARGS to the four kfuncs and
teaches the pre-v7.0 NULL guard to skip __ign arguments.  The latter is
needed because bpf_copy_from_user_str/task_str intentionally allow
unsafe_ptr__ign == NULL with BPF_F_PAD_ZEROS; without the skip, the
legitimate safe controls fail to load.  The diffstat is:

  kernel/bpf/crypto.c   | 2 +-
  kernel/bpf/helpers.c  | 6 +++---
  kernel/bpf/verifier.c | 1 +

Tested on a patched 6.18.64 kernel:

  test                           unpatched 6.18.64       patched
  ---------------------------------------------------------------------
  copy_str_null_dst              LOAD_OK + oops          LOAD_FAIL -EACCES (arg0)
  copy_task_str_null_dst         LOAD_OK + oops          LOAD_FAIL -EACCES (arg0)
  crypto_null_err                LOAD_OK + oops          LOAD_FAIL -EACCES (arg2)
  rbtree_null_first              LOAD_OK + oops          LOAD_FAIL -EACCES (arg1)
  copy_str_safe (NULL src)       LOAD_OK, RUN_DONE       LOAD_OK, RUN_DONE
  copy_task_str_safe (NULL src)  LOAD_OK, RUN_DONE       LOAD_OK, RUN_DONE
  crypto_safe                    LOAD_OK, RUN_DONE       LOAD_OK, RUN_DONE
  rbtree_valid_remove            LOAD_OK, RUN_DONE       LOAD_OK, RUN_DONE
  rbtree_safe                    LOAD_OK, RUN_DONE       LOAD_OK, RUN_DONE
  copy_str_zero_null             LOAD_OK, RUN_DONE       LOAD_FAIL -EACCES

One behavior change: dst == NULL with dst__sz == 0 is now rejected.
I confirmed that mainline 7.4.0-rc3+ rejects this too; its
copy_str_safe/copy_task_str_safe/rbtree_valid_remove controls still load
and run.  If 6.18.y must preserve that case, the
copy helpers can instead get an implementation-side

  if (!dst)
      return -EINVAL;

after the existing dst__sz == 0 early return.  That preserves the
zero-length semantics but leaves the verifier accepting NULL; it only
prevents the oops at runtime.

Fixes tags for a split submission:

  Fixes: 65ab5ac4df01 ("bpf: Add bpf_copy_from_user_str kfunc")
  Fixes: f0f8a5b58f78 ("bpf: Add bpf_copy_from_user_task_str() kfunc")
  Fixes: 3e1c6f35409f ("bpf: make common crypto API for TC/XDP programs")
  Fixes: bd1279ae8a69 ("bpf: Add bpf_rbtree_{add,remove,first} kfuncs")

Thanks,
Yi Liu


Yi Liu (1):
  bpf: Reject NULL arguments to four kfuncs

 kernel/bpf/crypto.c   | 2 +-
 kernel/bpf/helpers.c  | 6 +++---
 kernel/bpf/verifier.c | 1 +
 3 files changed, 5 insertions(+), 4 deletions(-)

-- 
2.34.1


^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2026-11-18 10:40 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-18 10:17 [RFC PATCH 6.18.y 0/1] bpf: reject NULL arguments to four kfuncs asatsuyu.liu
2026-11-18 10:17 ` [RFC PATCH 6.18.y 1/1] bpf: Reject " asatsuyu.liu
2026-11-18 10:40   ` sashiko-bot
2026-11-18 10:40 ` [RFC PATCH 6.18.y 0/1] bpf: reject " Greg KH

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.