linux.git/kernel/bpf/verifier.c, branch v6.4-rc2 Linux kernel source tree bpf: Add __rcu_read_{lock,unlock} into btf id deny list 2023-04-24T21:16:01+00:00 Yafang Shao laoar.shao@gmail.com 2023-04-24T16:11:03+00:00 a0c109dcafb15b8bee187c49fb746779374f60f0 The tracing recursion prevention mechanism must be protected by rcu, that leaves __rcu_read_{lock,unlock} unprotected by this mechanism. If we trace them, the recursion will happen. Let's add them into the btf id deny list. When CONFIG_PREEMPT_RCU is enabled, it can be reproduced with a simple bpf program as such: SEC("fentry/__rcu_read_lock") int fentry_run() { return 0; } Signed-off-by: Yafang Shao <laoar.shao@gmail.com> Link: https://lore.kernel.org/r/20230424161104.3737-2-laoar.shao@gmail.com Signed-off-by: Alexei Starovoitov <ast@kernel.org> 
The tracing recursion prevention mechanism must be protected by rcu, that
leaves __rcu_read_{lock,unlock} unprotected by this mechanism. If we trace
them, the recursion will happen. Let's add them into the btf id deny list.

When CONFIG_PREEMPT_RCU is enabled, it can be reproduced with a simple bpf
program as such:
  SEC("fentry/__rcu_read_lock")
  int fentry_run()
  {
      return 0;
  }

Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
Link: https://lore.kernel.org/r/20230424161104.3737-2-laoar.shao@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
bpf: Disable bpf_refcount_acquire kfunc calls until race conditions are fixed 2023-04-24T21:02:11+00:00 Dave Marchevsky davemarchevsky@fb.com 2023-04-24T20:43:21+00:00 7deca5eae83389ca40ac1b1bde96e4af17cca84f As reported by Kumar in [0], the shared ownership implementation for BPF programs has some race conditions which need to be addressed before it can safely be used. This patch does so in a minimal way instead of ripping out shared ownership entirely, as proper fixes for the issues raised will follow ASAP, at which point this patch's commit can be reverted to re-enable shared ownership. The patch removes the ability to call bpf_refcount_acquire_impl from BPF programs. Programs can only bump refcount and obtain a new owning reference using this kfunc, so removing the ability to call it effectively disables shared ownership. Instead of changing success / failure expectations for bpf_refcount-related selftests, this patch just disables them from running for now. [0]: https://lore.kernel.org/bpf/d7hyspcow5wtjcmw4fugdgyp3fwhljwuscp3xyut5qnwivyeru@ysdq543otzv2/ Reported-by: Kumar Kartikeya Dwivedi <memxor@gmail.com> Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com> Link: https://lore.kernel.org/r/20230424204321.2680232-1-davemarchevsky@fb.com Signed-off-by: Alexei Starovoitov <ast@kernel.org> 
As reported by Kumar in [0], the shared ownership implementation for BPF
programs has some race conditions which need to be addressed before it
can safely be used. This patch does so in a minimal way instead of
ripping out shared ownership entirely, as proper fixes for the issues
raised will follow ASAP, at which point this patch's commit can be
reverted to re-enable shared ownership.

The patch removes the ability to call bpf_refcount_acquire_impl from BPF
programs. Programs can only bump refcount and obtain a new owning
reference using this kfunc, so removing the ability to call it
effectively disables shared ownership.

Instead of changing success / failure expectations for
bpf_refcount-related selftests, this patch just disables them from
running for now.

  [0]: https://lore.kernel.org/bpf/d7hyspcow5wtjcmw4fugdgyp3fwhljwuscp3xyut5qnwivyeru@ysdq543otzv2/

Reported-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230424204321.2680232-1-davemarchevsky@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Merge tag 'for-netdev' of https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next 2023-04-22T03:32:37+00:00 Jakub Kicinski kuba@kernel.org 2023-04-22T03:32:36+00:00 9a82cdc28f47f5e70f657a25c6e5bf16cf306848 Daniel Borkmann says: ==================== pull-request: bpf-next 2023-04-21 We've added 71 non-merge commits during the last 8 day(s) which contain a total of 116 files changed, 13397 insertions(+), 8896 deletions(-). The main changes are: 1) Add a new BPF netfilter program type and minimal support to hook BPF programs to netfilter hooks such as prerouting or forward, from Florian Westphal. 2) Fix race between btf_put and btf_idr walk which caused a deadlock, from Alexei Starovoitov. 3) Second big batch to migrate test_verifier unit tests into test_progs for ease of readability and debugging, from Eduard Zingerman. 4) Add support for refcounted local kptrs to the verifier for allowing shared ownership, useful for adding a node to both the BPF list and rbtree, from Dave Marchevsky. 5) Migrate bpf_for(), bpf_for_each() and bpf_repeat() macros from BPF selftests into libbpf-provided bpf_helpers.h header and improve kfunc handling, from Andrii Nakryiko. 6) Support 64-bit pointers to kfuncs needed for archs like s390x, from Ilya Leoshkevich. 7) Support BPF progs under getsockopt with a NULL optval, from Stanislav Fomichev. 8) Improve verifier u32 scalar equality checking in order to enable LLVM transformations which earlier had to be disabled specifically for BPF backend, from Yonghong Song. 9) Extend bpftool's struct_ops object loading to support links, from Kui-Feng Lee. 10) Add xsk selftest follow-up fixes for hugepage allocated umem, from Magnus Karlsson. 11) Support BPF redirects from tc BPF to ifb devices, from Daniel Borkmann. 12) Add BPF support for integer type when accessing variable length arrays, from Feng Zhou. * tag 'for-netdev' of https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next: (71 commits) selftests/bpf: verifier/value_ptr_arith converted to inline assembly selftests/bpf: verifier/value_illegal_alu converted to inline assembly selftests/bpf: verifier/unpriv converted to inline assembly selftests/bpf: verifier/subreg converted to inline assembly selftests/bpf: verifier/spin_lock converted to inline assembly selftests/bpf: verifier/sock converted to inline assembly selftests/bpf: verifier/search_pruning converted to inline assembly selftests/bpf: verifier/runtime_jit converted to inline assembly selftests/bpf: verifier/regalloc converted to inline assembly selftests/bpf: verifier/ref_tracking converted to inline assembly selftests/bpf: verifier/map_ptr_mixing converted to inline assembly selftests/bpf: verifier/map_in_map converted to inline assembly selftests/bpf: verifier/lwt converted to inline assembly selftests/bpf: verifier/loops1 converted to inline assembly selftests/bpf: verifier/jeq_infer_not_null converted to inline assembly selftests/bpf: verifier/direct_packet_access converted to inline assembly selftests/bpf: verifier/d_path converted to inline assembly selftests/bpf: verifier/ctx converted to inline assembly selftests/bpf: verifier/btf_ctx_access converted to inline assembly selftests/bpf: verifier/bpf_get_stack converted to inline assembly ... ==================== Link: https://lore.kernel.org/r/20230421211035.9111-1-daniel@iogearbox.net Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
Daniel Borkmann says:

====================
pull-request: bpf-next 2023-04-21

We've added 71 non-merge commits during the last 8 day(s) which contain
a total of 116 files changed, 13397 insertions(+), 8896 deletions(-).

The main changes are:

1) Add a new BPF netfilter program type and minimal support to hook
   BPF programs to netfilter hooks such as prerouting or forward,
   from Florian Westphal.

2) Fix race between btf_put and btf_idr walk which caused a deadlock,
   from Alexei Starovoitov.

3) Second big batch to migrate test_verifier unit tests into test_progs
   for ease of readability and debugging, from Eduard Zingerman.

4) Add support for refcounted local kptrs to the verifier for allowing
   shared ownership, useful for adding a node to both the BPF list and
   rbtree, from Dave Marchevsky.

5) Migrate bpf_for(), bpf_for_each() and bpf_repeat() macros from BPF
  selftests into libbpf-provided bpf_helpers.h header and improve
  kfunc handling, from Andrii Nakryiko.

6) Support 64-bit pointers to kfuncs needed for archs like s390x,
   from Ilya Leoshkevich.

7) Support BPF progs under getsockopt with a NULL optval,
   from Stanislav Fomichev.

8) Improve verifier u32 scalar equality checking in order to enable
   LLVM transformations which earlier had to be disabled specifically
   for BPF backend, from Yonghong Song.

9) Extend bpftool's struct_ops object loading to support links,
   from Kui-Feng Lee.

10) Add xsk selftest follow-up fixes for hugepage allocated umem,
    from Magnus Karlsson.

11) Support BPF redirects from tc BPF to ifb devices,
    from Daniel Borkmann.

12) Add BPF support for integer type when accessing variable length
    arrays, from Feng Zhou.

* tag 'for-netdev' of https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next: (71 commits)
  selftests/bpf: verifier/value_ptr_arith converted to inline assembly
  selftests/bpf: verifier/value_illegal_alu converted to inline assembly
  selftests/bpf: verifier/unpriv converted to inline assembly
  selftests/bpf: verifier/subreg converted to inline assembly
  selftests/bpf: verifier/spin_lock converted to inline assembly
  selftests/bpf: verifier/sock converted to inline assembly
  selftests/bpf: verifier/search_pruning converted to inline assembly
  selftests/bpf: verifier/runtime_jit converted to inline assembly
  selftests/bpf: verifier/regalloc converted to inline assembly
  selftests/bpf: verifier/ref_tracking converted to inline assembly
  selftests/bpf: verifier/map_ptr_mixing converted to inline assembly
  selftests/bpf: verifier/map_in_map converted to inline assembly
  selftests/bpf: verifier/lwt converted to inline assembly
  selftests/bpf: verifier/loops1 converted to inline assembly
  selftests/bpf: verifier/jeq_infer_not_null converted to inline assembly
  selftests/bpf: verifier/direct_packet_access converted to inline assembly
  selftests/bpf: verifier/d_path converted to inline assembly
  selftests/bpf: verifier/ctx converted to inline assembly
  selftests/bpf: verifier/btf_ctx_access converted to inline assembly
  selftests/bpf: verifier/bpf_get_stack converted to inline assembly
  ...
====================

Link: https://lore.kernel.org/r/20230421211035.9111-1-daniel@iogearbox.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
bpf: minimal support for programs hooked into netfilter framework 2023-04-21T18:34:14+00:00 Florian Westphal fw@strlen.de 2023-04-21T17:02:55+00:00 fd9c663b9ad67dedfc9a3fd3429ddd3e83782b4d This adds minimal support for BPF_PROG_TYPE_NETFILTER bpf programs that will be invoked via the NF_HOOK() points in the ip stack. Invocation incurs an indirect call. This is not a necessity: Its possible to add 'DEFINE_BPF_DISPATCHER(nf_progs)' and handle the program invocation with the same method already done for xdp progs. This isn't done here to keep the size of this chunk down. Verifier restricts verdicts to either DROP or ACCEPT. Signed-off-by: Florian Westphal <fw@strlen.de> Link: https://lore.kernel.org/r/20230421170300.24115-3-fw@strlen.de Signed-off-by: Alexei Starovoitov <ast@kernel.org> 
This adds minimal support for BPF_PROG_TYPE_NETFILTER bpf programs
that will be invoked via the NF_HOOK() points in the ip stack.

Invocation incurs an indirect call.  This is not a necessity: Its
possible to add 'DEFINE_BPF_DISPATCHER(nf_progs)' and handle the
program invocation with the same method already done for xdp progs.

This isn't done here to keep the size of this chunk down.

Verifier restricts verdicts to either DROP or ACCEPT.

Signed-off-by: Florian Westphal <fw@strlen.de>
Link: https://lore.kernel.org/r/20230421170300.24115-3-fw@strlen.de
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net 2023-04-20T23:29:51+00:00 Jakub Kicinski kuba@kernel.org 2023-04-20T23:27:33+00:00 681c5b51dc6b8ff1ec05555243eccf64a08cb2fd Adjacent changes: net/mptcp/protocol.h 63740448a32e ("mptcp: fix accept vs worker race") 2a6a870e44dd ("mptcp: stops worker on unaccepted sockets at listener close") ddb1a072f858 ("mptcp: move first subflow allocation at mpc access time") Signed-off-by: Jakub Kicinski <kuba@kernel.org> 
Adjacent changes:

net/mptcp/protocol.h
  63740448a32e ("mptcp: fix accept vs worker race")
  2a6a870e44dd ("mptcp: stops worker on unaccepted sockets at listener close")
  ddb1a072f858 ("mptcp: move first subflow allocation at mpc access time")

Signed-off-by: Jakub Kicinski <kuba@kernel.org>
bpf: Fix incorrect verifier pruning due to missing register precision taints 2023-04-19T17:18:18+00:00 Daniel Borkmann daniel@iogearbox.net 2023-04-11T15:24:13+00:00 71b547f561247897a0a14f3082730156c0533fed Juan Jose et al reported an issue found via fuzzing where the verifier's pruning logic prematurely marks a program path as safe. Consider the following program: 0: (b7) r6 = 1024 1: (b7) r7 = 0 2: (b7) r8 = 0 3: (b7) r9 = -2147483648 4: (97) r6 %= 1025 5: (05) goto pc+0 6: (bd) if r6 <= r9 goto pc+2 7: (97) r6 %= 1 8: (b7) r9 = 0 9: (bd) if r6 <= r9 goto pc+1 10: (b7) r6 = 0 11: (b7) r0 = 0 12: (63) *(u32 *)(r10 -4) = r0 13: (18) r4 = 0xffff888103693400 // map_ptr(ks=4,vs=48) 15: (bf) r1 = r4 16: (bf) r2 = r10 17: (07) r2 += -4 18: (85) call bpf_map_lookup_elem#1 19: (55) if r0 != 0x0 goto pc+1 20: (95) exit 21: (77) r6 >>= 10 22: (27) r6 *= 8192 23: (bf) r1 = r0 24: (0f) r0 += r6 25: (79) r3 = *(u64 *)(r0 +0) 26: (7b) *(u64 *)(r1 +0) = r3 27: (95) exit The verifier treats this as safe, leading to oob read/write access due to an incorrect verifier conclusion: func#0 @0 0: R1=ctx(off=0,imm=0) R10=fp0 0: (b7) r6 = 1024 ; R6_w=1024 1: (b7) r7 = 0 ; R7_w=0 2: (b7) r8 = 0 ; R8_w=0 3: (b7) r9 = -2147483648 ; R9_w=-2147483648 4: (97) r6 %= 1025 ; R6_w=scalar() 5: (05) goto pc+0 6: (bd) if r6 <= r9 goto pc+2 ; R6_w=scalar(umin=18446744071562067969,var_off=(0xffffffff00000000; 0xffffffff)) R9_w=-2147483648 7: (97) r6 %= 1 ; R6_w=scalar() 8: (b7) r9 = 0 ; R9=0 9: (bd) if r6 <= r9 goto pc+1 ; R6=scalar(umin=1) R9=0 10: (b7) r6 = 0 ; R6_w=0 11: (b7) r0 = 0 ; R0_w=0 12: (63) *(u32 *)(r10 -4) = r0 last_idx 12 first_idx 9 regs=1 stack=0 before 11: (b7) r0 = 0 13: R0_w=0 R10=fp0 fp-8=0000???? 13: (18) r4 = 0xffff8ad3886c2a00 ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 15: (bf) r1 = r4 ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 16: (bf) r2 = r10 ; R2_w=fp0 R10=fp0 17: (07) r2 += -4 ; R2_w=fp-4 18: (85) call bpf_map_lookup_elem#1 ; R0=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0) 19: (55) if r0 != 0x0 goto pc+1 ; R0=0 20: (95) exit from 19 to 21: R0=map_value(off=0,ks=4,vs=48,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm???? 21: (77) r6 >>= 10 ; R6_w=0 22: (27) r6 *= 8192 ; R6_w=0 23: (bf) r1 = r0 ; R0=map_value(off=0,ks=4,vs=48,imm=0) R1_w=map_value(off=0,ks=4,vs=48,imm=0) 24: (0f) r0 += r6 last_idx 24 first_idx 19 regs=40 stack=0 before 23: (bf) r1 = r0 regs=40 stack=0 before 22: (27) r6 *= 8192 regs=40 stack=0 before 21: (77) r6 >>= 10 regs=40 stack=0 before 19: (55) if r0 != 0x0 goto pc+1 parent didn't have regs=40 stack=0 marks: R0_rw=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0) R6_rw=P0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm???? last_idx 18 first_idx 9 regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1 regs=40 stack=0 before 17: (07) r2 += -4 regs=40 stack=0 before 16: (bf) r2 = r10 regs=40 stack=0 before 15: (bf) r1 = r4 regs=40 stack=0 before 13: (18) r4 = 0xffff8ad3886c2a00 regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0 regs=40 stack=0 before 11: (b7) r0 = 0 regs=40 stack=0 before 10: (b7) r6 = 0 25: (79) r3 = *(u64 *)(r0 +0) ; R0_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar() 26: (7b) *(u64 *)(r1 +0) = r3 ; R1_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar() 27: (95) exit from 9 to 11: R1=ctx(off=0,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0 11: (b7) r0 = 0 ; R0_w=0 12: (63) *(u32 *)(r10 -4) = r0 last_idx 12 first_idx 11 regs=1 stack=0 before 11: (b7) r0 = 0 13: R0_w=0 R10=fp0 fp-8=0000???? 13: (18) r4 = 0xffff8ad3886c2a00 ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 15: (bf) r1 = r4 ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 16: (bf) r2 = r10 ; R2_w=fp0 R10=fp0 17: (07) r2 += -4 ; R2_w=fp-4 18: (85) call bpf_map_lookup_elem#1 frame 0: propagating r6 last_idx 19 first_idx 11 regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1 regs=40 stack=0 before 17: (07) r2 += -4 regs=40 stack=0 before 16: (bf) r2 = r10 regs=40 stack=0 before 15: (bf) r1 = r4 regs=40 stack=0 before 13: (18) r4 = 0xffff8ad3886c2a00 regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0 regs=40 stack=0 before 11: (b7) r0 = 0 parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_r=P0 R7=0 R8=0 R9=0 R10=fp0 last_idx 9 first_idx 9 regs=40 stack=0 before 9: (bd) if r6 <= r9 goto pc+1 parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_rw=Pscalar() R7_w=0 R8_w=0 R9_rw=0 R10=fp0 last_idx 8 first_idx 0 regs=40 stack=0 before 8: (b7) r9 = 0 regs=40 stack=0 before 7: (97) r6 %= 1 regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2 regs=40 stack=0 before 5: (05) goto pc+0 regs=40 stack=0 before 4: (97) r6 %= 1025 regs=40 stack=0 before 3: (b7) r9 = -2147483648 regs=40 stack=0 before 2: (b7) r8 = 0 regs=40 stack=0 before 1: (b7) r7 = 0 regs=40 stack=0 before 0: (b7) r6 = 1024 19: safe frame 0: propagating r6 last_idx 9 first_idx 0 regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2 regs=40 stack=0 before 5: (05) goto pc+0 regs=40 stack=0 before 4: (97) r6 %= 1025 regs=40 stack=0 before 3: (b7) r9 = -2147483648 regs=40 stack=0 before 2: (b7) r8 = 0 regs=40 stack=0 before 1: (b7) r7 = 0 regs=40 stack=0 before 0: (b7) r6 = 1024 from 6 to 9: safe verification time 110 usec stack depth 4 processed 36 insns (limit 1000000) max_states_per_insn 0 total_states 3 peak_states 3 mark_read 2 The verifier considers this program as safe by mistakenly pruning unsafe code paths. In the above func#0, code lines 0-10 are of interest. In line 0-3 registers r6 to r9 are initialized with known scalar values. In line 4 the register r6 is reset to an unknown scalar given the verifier does not track modulo operations. Due to this, the verifier can also not determine precisely which branches in line 6 and 9 are taken, therefore it needs to explore them both. As can be seen, the verifier starts with exploring the false/fall-through paths first. The 'from 19 to 21' path has both r6=0 and r9=0 and the pointer arithmetic on r0 += r6 is therefore considered safe. Given the arithmetic, r6 is correctly marked for precision tracking where backtracking kicks in where it walks back the current path all the way where r6 was set to 0 in the fall-through branch. Next, the pruning logics pops the path 'from 9 to 11' from the stack. Also here, the state of the registers is the same, that is, r6=0 and r9=0, so that at line 19 the path can be pruned as it is considered safe. It is interesting to note that the conditional in line 9 turned r6 into a more precise state, that is, in the fall-through path at the beginning of line 10, it is R6=scalar(umin=1), and in the branch-taken path (which is analyzed here) at the beginning of line 11, r6 turned into a known const r6=0 as r9=0 prior to that and therefore (unsigned) r6 <= 0 concludes that r6 must be 0 (**): [...] ; R6_w=scalar() 9: (bd) if r6 <= r9 goto pc+1 ; R6=scalar(umin=1) R9=0 [...] from 9 to 11: R1=ctx(off=0,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0 [...] The next path is 'from 6 to 9'. The verifier considers the old and current state equivalent, and therefore prunes the search incorrectly. Looking into the two states which are being compared by the pruning logic at line 9, the old state consists of R6_rwD=Pscalar() R9_rwD=0 R10=fp0 and the new state consists of R1=ctx(off=0,imm=0) R6_w=scalar(umax=18446744071562067968) R7_w=0 R8_w=0 R9_w=-2147483648 R10=fp0. While r6 had the reg->precise flag correctly set in the old state, r9 did not. Both r6'es are considered as equivalent given the old one is a superset of the current, more precise one, however, r9's actual values (0 vs 0x80000000) mismatch. Given the old r9 did not have reg->precise flag set, the verifier does not consider the register as contributing to the precision state of r6, and therefore it considered both r9 states as equivalent. However, for this specific pruned path (which is also the actual path taken at runtime), register r6 will be 0x400 and r9 0x80000000 when reaching line 21, thus oob-accessing the map. The purpose of precision tracking is to initially mark registers (including spilled ones) as imprecise to help verifier's pruning logic finding equivalent states it can then prune if they don't contribute to the program's safety aspects. For example, if registers are used for pointer arithmetic or to pass constant length to a helper, then the verifier sets reg->precise flag and backtracks the BPF program instruction sequence and chain of verifier states to ensure that the given register or stack slot including their dependencies are marked as precisely tracked scalar. This also includes any other registers and slots that contribute to a tracked state of given registers/stack slot. This backtracking relies on recorded jmp_history and is able to traverse entire chain of parent states. This process ends only when all the necessary registers/slots and their transitive dependencies are marked as precise. The backtrack_insn() is called from the current instruction up to the first instruction, and its purpose is to compute a bitmask of registers and stack slots that need precision tracking in the parent's verifier state. For example, if a current instruction is r6 = r7, then r6 needs precision after this instruction and r7 needs precision before this instruction, that is, in the parent state. Hence for the latter r7 is marked and r6 unmarked. For the class of jmp/jmp32 instructions, backtrack_insn() today only looks at call and exit instructions and for all other conditionals the masks remain as-is. However, in the given situation register r6 has a dependency on r9 (as described above in **), so also that one needs to be marked for precision tracking. In other words, if an imprecise register influences a precise one, then the imprecise register should also be marked precise. Meaning, in the parent state both dest and src register need to be tracked for precision and therefore the marking must be more conservative by setting reg->precise flag for both. The precision propagation needs to cover both for the conditional: if the src reg was marked but not the dst reg and vice versa. After the fix the program is correctly rejected: func#0 @0 0: R1=ctx(off=0,imm=0) R10=fp0 0: (b7) r6 = 1024 ; R6_w=1024 1: (b7) r7 = 0 ; R7_w=0 2: (b7) r8 = 0 ; R8_w=0 3: (b7) r9 = -2147483648 ; R9_w=-2147483648 4: (97) r6 %= 1025 ; R6_w=scalar() 5: (05) goto pc+0 6: (bd) if r6 <= r9 goto pc+2 ; R6_w=scalar(umin=18446744071562067969,var_off=(0xffffffff80000000; 0x7fffffff),u32_min=-2147483648) R9_w=-2147483648 7: (97) r6 %= 1 ; R6_w=scalar() 8: (b7) r9 = 0 ; R9=0 9: (bd) if r6 <= r9 goto pc+1 ; R6=scalar(umin=1) R9=0 10: (b7) r6 = 0 ; R6_w=0 11: (b7) r0 = 0 ; R0_w=0 12: (63) *(u32 *)(r10 -4) = r0 last_idx 12 first_idx 9 regs=1 stack=0 before 11: (b7) r0 = 0 13: R0_w=0 R10=fp0 fp-8=0000???? 13: (18) r4 = 0xffff9290dc5bfe00 ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 15: (bf) r1 = r4 ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 16: (bf) r2 = r10 ; R2_w=fp0 R10=fp0 17: (07) r2 += -4 ; R2_w=fp-4 18: (85) call bpf_map_lookup_elem#1 ; R0=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0) 19: (55) if r0 != 0x0 goto pc+1 ; R0=0 20: (95) exit from 19 to 21: R0=map_value(off=0,ks=4,vs=48,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm???? 21: (77) r6 >>= 10 ; R6_w=0 22: (27) r6 *= 8192 ; R6_w=0 23: (bf) r1 = r0 ; R0=map_value(off=0,ks=4,vs=48,imm=0) R1_w=map_value(off=0,ks=4,vs=48,imm=0) 24: (0f) r0 += r6 last_idx 24 first_idx 19 regs=40 stack=0 before 23: (bf) r1 = r0 regs=40 stack=0 before 22: (27) r6 *= 8192 regs=40 stack=0 before 21: (77) r6 >>= 10 regs=40 stack=0 before 19: (55) if r0 != 0x0 goto pc+1 parent didn't have regs=40 stack=0 marks: R0_rw=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0) R6_rw=P0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm???? last_idx 18 first_idx 9 regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1 regs=40 stack=0 before 17: (07) r2 += -4 regs=40 stack=0 before 16: (bf) r2 = r10 regs=40 stack=0 before 15: (bf) r1 = r4 regs=40 stack=0 before 13: (18) r4 = 0xffff9290dc5bfe00 regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0 regs=40 stack=0 before 11: (b7) r0 = 0 regs=40 stack=0 before 10: (b7) r6 = 0 25: (79) r3 = *(u64 *)(r0 +0) ; R0_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar() 26: (7b) *(u64 *)(r1 +0) = r3 ; R1_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar() 27: (95) exit from 9 to 11: R1=ctx(off=0,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0 11: (b7) r0 = 0 ; R0_w=0 12: (63) *(u32 *)(r10 -4) = r0 last_idx 12 first_idx 11 regs=1 stack=0 before 11: (b7) r0 = 0 13: R0_w=0 R10=fp0 fp-8=0000???? 13: (18) r4 = 0xffff9290dc5bfe00 ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 15: (bf) r1 = r4 ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 16: (bf) r2 = r10 ; R2_w=fp0 R10=fp0 17: (07) r2 += -4 ; R2_w=fp-4 18: (85) call bpf_map_lookup_elem#1 frame 0: propagating r6 last_idx 19 first_idx 11 regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1 regs=40 stack=0 before 17: (07) r2 += -4 regs=40 stack=0 before 16: (bf) r2 = r10 regs=40 stack=0 before 15: (bf) r1 = r4 regs=40 stack=0 before 13: (18) r4 = 0xffff9290dc5bfe00 regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0 regs=40 stack=0 before 11: (b7) r0 = 0 parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_r=P0 R7=0 R8=0 R9=0 R10=fp0 last_idx 9 first_idx 9 regs=40 stack=0 before 9: (bd) if r6 <= r9 goto pc+1 parent didn't have regs=240 stack=0 marks: R1=ctx(off=0,imm=0) R6_rw=Pscalar() R7_w=0 R8_w=0 R9_rw=P0 R10=fp0 last_idx 8 first_idx 0 regs=240 stack=0 before 8: (b7) r9 = 0 regs=40 stack=0 before 7: (97) r6 %= 1 regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2 regs=240 stack=0 before 5: (05) goto pc+0 regs=240 stack=0 before 4: (97) r6 %= 1025 regs=240 stack=0 before 3: (b7) r9 = -2147483648 regs=40 stack=0 before 2: (b7) r8 = 0 regs=40 stack=0 before 1: (b7) r7 = 0 regs=40 stack=0 before 0: (b7) r6 = 1024 19: safe from 6 to 9: R1=ctx(off=0,imm=0) R6_w=scalar(umax=18446744071562067968) R7_w=0 R8_w=0 R9_w=-2147483648 R10=fp0 9: (bd) if r6 <= r9 goto pc+1 last_idx 9 first_idx 0 regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2 regs=240 stack=0 before 5: (05) goto pc+0 regs=240 stack=0 before 4: (97) r6 %= 1025 regs=240 stack=0 before 3: (b7) r9 = -2147483648 regs=40 stack=0 before 2: (b7) r8 = 0 regs=40 stack=0 before 1: (b7) r7 = 0 regs=40 stack=0 before 0: (b7) r6 = 1024 last_idx 9 first_idx 0 regs=200 stack=0 before 6: (bd) if r6 <= r9 goto pc+2 regs=240 stack=0 before 5: (05) goto pc+0 regs=240 stack=0 before 4: (97) r6 %= 1025 regs=240 stack=0 before 3: (b7) r9 = -2147483648 regs=40 stack=0 before 2: (b7) r8 = 0 regs=40 stack=0 before 1: (b7) r7 = 0 regs=40 stack=0 before 0: (b7) r6 = 1024 11: R6=scalar(umax=18446744071562067968) R9=-2147483648 11: (b7) r0 = 0 ; R0_w=0 12: (63) *(u32 *)(r10 -4) = r0 last_idx 12 first_idx 11 regs=1 stack=0 before 11: (b7) r0 = 0 13: R0_w=0 R10=fp0 fp-8=0000???? 13: (18) r4 = 0xffff9290dc5bfe00 ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 15: (bf) r1 = r4 ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0) 16: (bf) r2 = r10 ; R2_w=fp0 R10=fp0 17: (07) r2 += -4 ; R2_w=fp-4 18: (85) call bpf_map_lookup_elem#1 ; R0_w=map_value_or_null(id=3,off=0,ks=4,vs=48,imm=0) 19: (55) if r0 != 0x0 goto pc+1 ; R0_w=0 20: (95) exit from 19 to 21: R0=map_value(off=0,ks=4,vs=48,imm=0) R6=scalar(umax=18446744071562067968) R7=0 R8=0 R9=-2147483648 R10=fp0 fp-8=mmmm???? 21: (77) r6 >>= 10 ; R6_w=scalar(umax=18014398507384832,var_off=(0x0; 0x3fffffffffffff)) 22: (27) r6 *= 8192 ; R6_w=scalar(smax=9223372036854767616,umax=18446744073709543424,var_off=(0x0; 0xffffffffffffe000),s32_max=2147475456,u32_max=-8192) 23: (bf) r1 = r0 ; R0=map_value(off=0,ks=4,vs=48,imm=0) R1_w=map_value(off=0,ks=4,vs=48,imm=0) 24: (0f) r0 += r6 last_idx 24 first_idx 21 regs=40 stack=0 before 23: (bf) r1 = r0 regs=40 stack=0 before 22: (27) r6 *= 8192 regs=40 stack=0 before 21: (77) r6 >>= 10 parent didn't have regs=40 stack=0 marks: R0_rw=map_value(off=0,ks=4,vs=48,imm=0) R6_r=Pscalar(umax=18446744071562067968) R7=0 R8=0 R9=-2147483648 R10=fp0 fp-8=mmmm???? last_idx 19 first_idx 11 regs=40 stack=0 before 19: (55) if r0 != 0x0 goto pc+1 regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1 regs=40 stack=0 before 17: (07) r2 += -4 regs=40 stack=0 before 16: (bf) r2 = r10 regs=40 stack=0 before 15: (bf) r1 = r4 regs=40 stack=0 before 13: (18) r4 = 0xffff9290dc5bfe00 regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0 regs=40 stack=0 before 11: (b7) r0 = 0 parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_rw=Pscalar(umax=18446744071562067968) R7_w=0 R8_w=0 R9_w=-2147483648 R10=fp0 last_idx 9 first_idx 0 regs=40 stack=0 before 9: (bd) if r6 <= r9 goto pc+1 regs=240 stack=0 before 6: (bd) if r6 <= r9 goto pc+2 regs=240 stack=0 before 5: (05) goto pc+0 regs=240 stack=0 before 4: (97) r6 %= 1025 regs=240 stack=0 before 3: (b7) r9 = -2147483648 regs=40 stack=0 before 2: (b7) r8 = 0 regs=40 stack=0 before 1: (b7) r7 = 0 regs=40 stack=0 before 0: (b7) r6 = 1024 math between map_value pointer and register with unbounded min value is not allowed verification time 886 usec stack depth 4 processed 49 insns (limit 1000000) max_states_per_insn 1 total_states 5 peak_states 5 mark_read 2 Fixes: b5dc0163d8fd ("bpf: precise scalar_value tracking") Reported-by: Juan Jose Lopez Jaimez <jjlopezjaimez@google.com> Reported-by: Meador Inge <meadori@google.com> Reported-by: Simon Scannell <simonscannell@google.com> Reported-by: Nenad Stojanovski <thenenadx@google.com> Signed-off-by: Daniel Borkmann <daniel@iogearbox.net> Co-developed-by: Andrii Nakryiko <andrii@kernel.org> Signed-off-by: Andrii Nakryiko <andrii@kernel.org> Reviewed-by: John Fastabend <john.fastabend@gmail.com> Reviewed-by: Juan Jose Lopez Jaimez <jjlopezjaimez@google.com> Reviewed-by: Meador Inge <meadori@google.com> Reviewed-by: Simon Scannell <simonscannell@google.com> 
Juan Jose et al reported an issue found via fuzzing where the verifier's
pruning logic prematurely marks a program path as safe.

Consider the following program:

   0: (b7) r6 = 1024
   1: (b7) r7 = 0
   2: (b7) r8 = 0
   3: (b7) r9 = -2147483648
   4: (97) r6 %= 1025
   5: (05) goto pc+0
   6: (bd) if r6 <= r9 goto pc+2
   7: (97) r6 %= 1
   8: (b7) r9 = 0
   9: (bd) if r6 <= r9 goto pc+1
  10: (b7) r6 = 0
  11: (b7) r0 = 0
  12: (63) *(u32 *)(r10 -4) = r0
  13: (18) r4 = 0xffff888103693400 // map_ptr(ks=4,vs=48)
  15: (bf) r1 = r4
  16: (bf) r2 = r10
  17: (07) r2 += -4
  18: (85) call bpf_map_lookup_elem#1
  19: (55) if r0 != 0x0 goto pc+1
  20: (95) exit
  21: (77) r6 >>= 10
  22: (27) r6 *= 8192
  23: (bf) r1 = r0
  24: (0f) r0 += r6
  25: (79) r3 = *(u64 *)(r0 +0)
  26: (7b) *(u64 *)(r1 +0) = r3
  27: (95) exit

The verifier treats this as safe, leading to oob read/write access due
to an incorrect verifier conclusion:

  func#0 @0
  0: R1=ctx(off=0,imm=0) R10=fp0
  0: (b7) r6 = 1024                     ; R6_w=1024
  1: (b7) r7 = 0                        ; R7_w=0
  2: (b7) r8 = 0                        ; R8_w=0
  3: (b7) r9 = -2147483648              ; R9_w=-2147483648
  4: (97) r6 %= 1025                    ; R6_w=scalar()
  5: (05) goto pc+0
  6: (bd) if r6 <= r9 goto pc+2         ; R6_w=scalar(umin=18446744071562067969,var_off=(0xffffffff00000000; 0xffffffff)) R9_w=-2147483648
  7: (97) r6 %= 1                       ; R6_w=scalar()
  8: (b7) r9 = 0                        ; R9=0
  9: (bd) if r6 <= r9 goto pc+1         ; R6=scalar(umin=1) R9=0
  10: (b7) r6 = 0                       ; R6_w=0
  11: (b7) r0 = 0                       ; R0_w=0
  12: (63) *(u32 *)(r10 -4) = r0
  last_idx 12 first_idx 9
  regs=1 stack=0 before 11: (b7) r0 = 0
  13: R0_w=0 R10=fp0 fp-8=0000????
  13: (18) r4 = 0xffff8ad3886c2a00      ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  15: (bf) r1 = r4                      ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  16: (bf) r2 = r10                     ; R2_w=fp0 R10=fp0
  17: (07) r2 += -4                     ; R2_w=fp-4
  18: (85) call bpf_map_lookup_elem#1   ; R0=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0)
  19: (55) if r0 != 0x0 goto pc+1       ; R0=0
  20: (95) exit

  from 19 to 21: R0=map_value(off=0,ks=4,vs=48,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm????
  21: (77) r6 >>= 10                    ; R6_w=0
  22: (27) r6 *= 8192                   ; R6_w=0
  23: (bf) r1 = r0                      ; R0=map_value(off=0,ks=4,vs=48,imm=0) R1_w=map_value(off=0,ks=4,vs=48,imm=0)
  24: (0f) r0 += r6
  last_idx 24 first_idx 19
  regs=40 stack=0 before 23: (bf) r1 = r0
  regs=40 stack=0 before 22: (27) r6 *= 8192
  regs=40 stack=0 before 21: (77) r6 >>= 10
  regs=40 stack=0 before 19: (55) if r0 != 0x0 goto pc+1
  parent didn't have regs=40 stack=0 marks: R0_rw=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0) R6_rw=P0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm????
  last_idx 18 first_idx 9
  regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1
  regs=40 stack=0 before 17: (07) r2 += -4
  regs=40 stack=0 before 16: (bf) r2 = r10
  regs=40 stack=0 before 15: (bf) r1 = r4
  regs=40 stack=0 before 13: (18) r4 = 0xffff8ad3886c2a00
  regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0
  regs=40 stack=0 before 11: (b7) r0 = 0
  regs=40 stack=0 before 10: (b7) r6 = 0
  25: (79) r3 = *(u64 *)(r0 +0)         ; R0_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar()
  26: (7b) *(u64 *)(r1 +0) = r3         ; R1_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar()
  27: (95) exit

  from 9 to 11: R1=ctx(off=0,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0
  11: (b7) r0 = 0                       ; R0_w=0
  12: (63) *(u32 *)(r10 -4) = r0
  last_idx 12 first_idx 11
  regs=1 stack=0 before 11: (b7) r0 = 0
  13: R0_w=0 R10=fp0 fp-8=0000????
  13: (18) r4 = 0xffff8ad3886c2a00      ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  15: (bf) r1 = r4                      ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  16: (bf) r2 = r10                     ; R2_w=fp0 R10=fp0
  17: (07) r2 += -4                     ; R2_w=fp-4
  18: (85) call bpf_map_lookup_elem#1
  frame 0: propagating r6
  last_idx 19 first_idx 11
  regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1
  regs=40 stack=0 before 17: (07) r2 += -4
  regs=40 stack=0 before 16: (bf) r2 = r10
  regs=40 stack=0 before 15: (bf) r1 = r4
  regs=40 stack=0 before 13: (18) r4 = 0xffff8ad3886c2a00
  regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0
  regs=40 stack=0 before 11: (b7) r0 = 0
  parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_r=P0 R7=0 R8=0 R9=0 R10=fp0
  last_idx 9 first_idx 9
  regs=40 stack=0 before 9: (bd) if r6 <= r9 goto pc+1
  parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_rw=Pscalar() R7_w=0 R8_w=0 R9_rw=0 R10=fp0
  last_idx 8 first_idx 0
  regs=40 stack=0 before 8: (b7) r9 = 0
  regs=40 stack=0 before 7: (97) r6 %= 1
  regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2
  regs=40 stack=0 before 5: (05) goto pc+0
  regs=40 stack=0 before 4: (97) r6 %= 1025
  regs=40 stack=0 before 3: (b7) r9 = -2147483648
  regs=40 stack=0 before 2: (b7) r8 = 0
  regs=40 stack=0 before 1: (b7) r7 = 0
  regs=40 stack=0 before 0: (b7) r6 = 1024
  19: safe
  frame 0: propagating r6
  last_idx 9 first_idx 0
  regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2
  regs=40 stack=0 before 5: (05) goto pc+0
  regs=40 stack=0 before 4: (97) r6 %= 1025
  regs=40 stack=0 before 3: (b7) r9 = -2147483648
  regs=40 stack=0 before 2: (b7) r8 = 0
  regs=40 stack=0 before 1: (b7) r7 = 0
  regs=40 stack=0 before 0: (b7) r6 = 1024

  from 6 to 9: safe
  verification time 110 usec
  stack depth 4
  processed 36 insns (limit 1000000) max_states_per_insn 0 total_states 3 peak_states 3 mark_read 2

The verifier considers this program as safe by mistakenly pruning unsafe
code paths. In the above func#0, code lines 0-10 are of interest. In line
0-3 registers r6 to r9 are initialized with known scalar values. In line 4
the register r6 is reset to an unknown scalar given the verifier does not
track modulo operations. Due to this, the verifier can also not determine
precisely which branches in line 6 and 9 are taken, therefore it needs to
explore them both.

As can be seen, the verifier starts with exploring the false/fall-through
paths first. The 'from 19 to 21' path has both r6=0 and r9=0 and the pointer
arithmetic on r0 += r6 is therefore considered safe. Given the arithmetic,
r6 is correctly marked for precision tracking where backtracking kicks in
where it walks back the current path all the way where r6 was set to 0 in
the fall-through branch.

Next, the pruning logics pops the path 'from 9 to 11' from the stack. Also
here, the state of the registers is the same, that is, r6=0 and r9=0, so
that at line 19 the path can be pruned as it is considered safe. It is
interesting to note that the conditional in line 9 turned r6 into a more
precise state, that is, in the fall-through path at the beginning of line
10, it is R6=scalar(umin=1), and in the branch-taken path (which is analyzed
here) at the beginning of line 11, r6 turned into a known const r6=0 as
r9=0 prior to that and therefore (unsigned) r6 <= 0 concludes that r6 must
be 0 (**):

  [...]                                 ; R6_w=scalar()
  9: (bd) if r6 <= r9 goto pc+1         ; R6=scalar(umin=1) R9=0
  [...]

  from 9 to 11: R1=ctx(off=0,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0
  [...]

The next path is 'from 6 to 9'. The verifier considers the old and current
state equivalent, and therefore prunes the search incorrectly. Looking into
the two states which are being compared by the pruning logic at line 9, the
old state consists of R6_rwD=Pscalar() R9_rwD=0 R10=fp0 and the new state
consists of R1=ctx(off=0,imm=0) R6_w=scalar(umax=18446744071562067968)
R7_w=0 R8_w=0 R9_w=-2147483648 R10=fp0. While r6 had the reg->precise flag
correctly set in the old state, r9 did not. Both r6'es are considered as
equivalent given the old one is a superset of the current, more precise one,
however, r9's actual values (0 vs 0x80000000) mismatch. Given the old r9
did not have reg->precise flag set, the verifier does not consider the
register as contributing to the precision state of r6, and therefore it
considered both r9 states as equivalent. However, for this specific pruned
path (which is also the actual path taken at runtime), register r6 will be
0x400 and r9 0x80000000 when reaching line 21, thus oob-accessing the map.

The purpose of precision tracking is to initially mark registers (including
spilled ones) as imprecise to help verifier's pruning logic finding equivalent
states it can then prune if they don't contribute to the program's safety
aspects. For example, if registers are used for pointer arithmetic or to pass
constant length to a helper, then the verifier sets reg->precise flag and
backtracks the BPF program instruction sequence and chain of verifier states
to ensure that the given register or stack slot including their dependencies
are marked as precisely tracked scalar. This also includes any other registers
and slots that contribute to a tracked state of given registers/stack slot.
This backtracking relies on recorded jmp_history and is able to traverse
entire chain of parent states. This process ends only when all the necessary
registers/slots and their transitive dependencies are marked as precise.

The backtrack_insn() is called from the current instruction up to the first
instruction, and its purpose is to compute a bitmask of registers and stack
slots that need precision tracking in the parent's verifier state. For example,
if a current instruction is r6 = r7, then r6 needs precision after this
instruction and r7 needs precision before this instruction, that is, in the
parent state. Hence for the latter r7 is marked and r6 unmarked.

For the class of jmp/jmp32 instructions, backtrack_insn() today only looks
at call and exit instructions and for all other conditionals the masks
remain as-is. However, in the given situation register r6 has a dependency
on r9 (as described above in **), so also that one needs to be marked for
precision tracking. In other words, if an imprecise register influences a
precise one, then the imprecise register should also be marked precise.
Meaning, in the parent state both dest and src register need to be tracked
for precision and therefore the marking must be more conservative by setting
reg->precise flag for both. The precision propagation needs to cover both
for the conditional: if the src reg was marked but not the dst reg and vice
versa.

After the fix the program is correctly rejected:

  func#0 @0
  0: R1=ctx(off=0,imm=0) R10=fp0
  0: (b7) r6 = 1024                     ; R6_w=1024
  1: (b7) r7 = 0                        ; R7_w=0
  2: (b7) r8 = 0                        ; R8_w=0
  3: (b7) r9 = -2147483648              ; R9_w=-2147483648
  4: (97) r6 %= 1025                    ; R6_w=scalar()
  5: (05) goto pc+0
  6: (bd) if r6 <= r9 goto pc+2         ; R6_w=scalar(umin=18446744071562067969,var_off=(0xffffffff80000000; 0x7fffffff),u32_min=-2147483648) R9_w=-2147483648
  7: (97) r6 %= 1                       ; R6_w=scalar()
  8: (b7) r9 = 0                        ; R9=0
  9: (bd) if r6 <= r9 goto pc+1         ; R6=scalar(umin=1) R9=0
  10: (b7) r6 = 0                       ; R6_w=0
  11: (b7) r0 = 0                       ; R0_w=0
  12: (63) *(u32 *)(r10 -4) = r0
  last_idx 12 first_idx 9
  regs=1 stack=0 before 11: (b7) r0 = 0
  13: R0_w=0 R10=fp0 fp-8=0000????
  13: (18) r4 = 0xffff9290dc5bfe00      ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  15: (bf) r1 = r4                      ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  16: (bf) r2 = r10                     ; R2_w=fp0 R10=fp0
  17: (07) r2 += -4                     ; R2_w=fp-4
  18: (85) call bpf_map_lookup_elem#1   ; R0=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0)
  19: (55) if r0 != 0x0 goto pc+1       ; R0=0
  20: (95) exit

  from 19 to 21: R0=map_value(off=0,ks=4,vs=48,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm????
  21: (77) r6 >>= 10                    ; R6_w=0
  22: (27) r6 *= 8192                   ; R6_w=0
  23: (bf) r1 = r0                      ; R0=map_value(off=0,ks=4,vs=48,imm=0) R1_w=map_value(off=0,ks=4,vs=48,imm=0)
  24: (0f) r0 += r6
  last_idx 24 first_idx 19
  regs=40 stack=0 before 23: (bf) r1 = r0
  regs=40 stack=0 before 22: (27) r6 *= 8192
  regs=40 stack=0 before 21: (77) r6 >>= 10
  regs=40 stack=0 before 19: (55) if r0 != 0x0 goto pc+1
  parent didn't have regs=40 stack=0 marks: R0_rw=map_value_or_null(id=1,off=0,ks=4,vs=48,imm=0) R6_rw=P0 R7=0 R8=0 R9=0 R10=fp0 fp-8=mmmm????
  last_idx 18 first_idx 9
  regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1
  regs=40 stack=0 before 17: (07) r2 += -4
  regs=40 stack=0 before 16: (bf) r2 = r10
  regs=40 stack=0 before 15: (bf) r1 = r4
  regs=40 stack=0 before 13: (18) r4 = 0xffff9290dc5bfe00
  regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0
  regs=40 stack=0 before 11: (b7) r0 = 0
  regs=40 stack=0 before 10: (b7) r6 = 0
  25: (79) r3 = *(u64 *)(r0 +0)         ; R0_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar()
  26: (7b) *(u64 *)(r1 +0) = r3         ; R1_w=map_value(off=0,ks=4,vs=48,imm=0) R3_w=scalar()
  27: (95) exit

  from 9 to 11: R1=ctx(off=0,imm=0) R6=0 R7=0 R8=0 R9=0 R10=fp0
  11: (b7) r0 = 0                       ; R0_w=0
  12: (63) *(u32 *)(r10 -4) = r0
  last_idx 12 first_idx 11
  regs=1 stack=0 before 11: (b7) r0 = 0
  13: R0_w=0 R10=fp0 fp-8=0000????
  13: (18) r4 = 0xffff9290dc5bfe00      ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  15: (bf) r1 = r4                      ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  16: (bf) r2 = r10                     ; R2_w=fp0 R10=fp0
  17: (07) r2 += -4                     ; R2_w=fp-4
  18: (85) call bpf_map_lookup_elem#1
  frame 0: propagating r6
  last_idx 19 first_idx 11
  regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1
  regs=40 stack=0 before 17: (07) r2 += -4
  regs=40 stack=0 before 16: (bf) r2 = r10
  regs=40 stack=0 before 15: (bf) r1 = r4
  regs=40 stack=0 before 13: (18) r4 = 0xffff9290dc5bfe00
  regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0
  regs=40 stack=0 before 11: (b7) r0 = 0
  parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_r=P0 R7=0 R8=0 R9=0 R10=fp0
  last_idx 9 first_idx 9
  regs=40 stack=0 before 9: (bd) if r6 <= r9 goto pc+1
  parent didn't have regs=240 stack=0 marks: R1=ctx(off=0,imm=0) R6_rw=Pscalar() R7_w=0 R8_w=0 R9_rw=P0 R10=fp0
  last_idx 8 first_idx 0
  regs=240 stack=0 before 8: (b7) r9 = 0
  regs=40 stack=0 before 7: (97) r6 %= 1
  regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2
  regs=240 stack=0 before 5: (05) goto pc+0
  regs=240 stack=0 before 4: (97) r6 %= 1025
  regs=240 stack=0 before 3: (b7) r9 = -2147483648
  regs=40 stack=0 before 2: (b7) r8 = 0
  regs=40 stack=0 before 1: (b7) r7 = 0
  regs=40 stack=0 before 0: (b7) r6 = 1024
  19: safe

  from 6 to 9: R1=ctx(off=0,imm=0) R6_w=scalar(umax=18446744071562067968) R7_w=0 R8_w=0 R9_w=-2147483648 R10=fp0
  9: (bd) if r6 <= r9 goto pc+1
  last_idx 9 first_idx 0
  regs=40 stack=0 before 6: (bd) if r6 <= r9 goto pc+2
  regs=240 stack=0 before 5: (05) goto pc+0
  regs=240 stack=0 before 4: (97) r6 %= 1025
  regs=240 stack=0 before 3: (b7) r9 = -2147483648
  regs=40 stack=0 before 2: (b7) r8 = 0
  regs=40 stack=0 before 1: (b7) r7 = 0
  regs=40 stack=0 before 0: (b7) r6 = 1024
  last_idx 9 first_idx 0
  regs=200 stack=0 before 6: (bd) if r6 <= r9 goto pc+2
  regs=240 stack=0 before 5: (05) goto pc+0
  regs=240 stack=0 before 4: (97) r6 %= 1025
  regs=240 stack=0 before 3: (b7) r9 = -2147483648
  regs=40 stack=0 before 2: (b7) r8 = 0
  regs=40 stack=0 before 1: (b7) r7 = 0
  regs=40 stack=0 before 0: (b7) r6 = 1024
  11: R6=scalar(umax=18446744071562067968) R9=-2147483648
  11: (b7) r0 = 0                       ; R0_w=0
  12: (63) *(u32 *)(r10 -4) = r0
  last_idx 12 first_idx 11
  regs=1 stack=0 before 11: (b7) r0 = 0
  13: R0_w=0 R10=fp0 fp-8=0000????
  13: (18) r4 = 0xffff9290dc5bfe00      ; R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  15: (bf) r1 = r4                      ; R1_w=map_ptr(off=0,ks=4,vs=48,imm=0) R4_w=map_ptr(off=0,ks=4,vs=48,imm=0)
  16: (bf) r2 = r10                     ; R2_w=fp0 R10=fp0
  17: (07) r2 += -4                     ; R2_w=fp-4
  18: (85) call bpf_map_lookup_elem#1   ; R0_w=map_value_or_null(id=3,off=0,ks=4,vs=48,imm=0)
  19: (55) if r0 != 0x0 goto pc+1       ; R0_w=0
  20: (95) exit

  from 19 to 21: R0=map_value(off=0,ks=4,vs=48,imm=0) R6=scalar(umax=18446744071562067968) R7=0 R8=0 R9=-2147483648 R10=fp0 fp-8=mmmm????
  21: (77) r6 >>= 10                    ; R6_w=scalar(umax=18014398507384832,var_off=(0x0; 0x3fffffffffffff))
  22: (27) r6 *= 8192                   ; R6_w=scalar(smax=9223372036854767616,umax=18446744073709543424,var_off=(0x0; 0xffffffffffffe000),s32_max=2147475456,u32_max=-8192)
  23: (bf) r1 = r0                      ; R0=map_value(off=0,ks=4,vs=48,imm=0) R1_w=map_value(off=0,ks=4,vs=48,imm=0)
  24: (0f) r0 += r6
  last_idx 24 first_idx 21
  regs=40 stack=0 before 23: (bf) r1 = r0
  regs=40 stack=0 before 22: (27) r6 *= 8192
  regs=40 stack=0 before 21: (77) r6 >>= 10
  parent didn't have regs=40 stack=0 marks: R0_rw=map_value(off=0,ks=4,vs=48,imm=0) R6_r=Pscalar(umax=18446744071562067968) R7=0 R8=0 R9=-2147483648 R10=fp0 fp-8=mmmm????
  last_idx 19 first_idx 11
  regs=40 stack=0 before 19: (55) if r0 != 0x0 goto pc+1
  regs=40 stack=0 before 18: (85) call bpf_map_lookup_elem#1
  regs=40 stack=0 before 17: (07) r2 += -4
  regs=40 stack=0 before 16: (bf) r2 = r10
  regs=40 stack=0 before 15: (bf) r1 = r4
  regs=40 stack=0 before 13: (18) r4 = 0xffff9290dc5bfe00
  regs=40 stack=0 before 12: (63) *(u32 *)(r10 -4) = r0
  regs=40 stack=0 before 11: (b7) r0 = 0
  parent didn't have regs=40 stack=0 marks: R1=ctx(off=0,imm=0) R6_rw=Pscalar(umax=18446744071562067968) R7_w=0 R8_w=0 R9_w=-2147483648 R10=fp0
  last_idx 9 first_idx 0
  regs=40 stack=0 before 9: (bd) if r6 <= r9 goto pc+1
  regs=240 stack=0 before 6: (bd) if r6 <= r9 goto pc+2
  regs=240 stack=0 before 5: (05) goto pc+0
  regs=240 stack=0 before 4: (97) r6 %= 1025
  regs=240 stack=0 before 3: (b7) r9 = -2147483648
  regs=40 stack=0 before 2: (b7) r8 = 0
  regs=40 stack=0 before 1: (b7) r7 = 0
  regs=40 stack=0 before 0: (b7) r6 = 1024
  math between map_value pointer and register with unbounded min value is not allowed
  verification time 886 usec
  stack depth 4
  processed 49 insns (limit 1000000) max_states_per_insn 1 total_states 5 peak_states 5 mark_read 2

Fixes: b5dc0163d8fd ("bpf: precise scalar_value tracking")
Reported-by: Juan Jose Lopez Jaimez <jjlopezjaimez@google.com>
Reported-by: Meador Inge <meadori@google.com>
Reported-by: Simon Scannell <simonscannell@google.com>
Reported-by: Nenad Stojanovski <thenenadx@google.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Co-developed-by: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Reviewed-by: John Fastabend <john.fastabend@gmail.com>
Reviewed-by: Juan Jose Lopez Jaimez <jjlopezjaimez@google.com>
Reviewed-by: Meador Inge <meadori@google.com>
Reviewed-by: Simon Scannell <simonscannell@google.com>
bpf: Improve verifier u32 scalar equality checking 2023-04-17T22:50:02+00:00 Yonghong Song yhs@fb.com 2023-04-17T22:21:34+00:00 3be49f79555ee975acb4ad8b5de06fa4351264aa In [1], I tried to remove bpf-specific codes to prevent certain llvm optimizations, and add llvm TTI (target transform info) hooks to prevent those optimizations. During this process, I found if I enable llvm SimplifyCFG:shouldFoldTwoEntryPHINode transformation, I will hit the following verification failure with selftests: ... 8: (18) r1 = 0xffffc900001b2230 ; R1_w=map_value(off=560,ks=4,vs=564,imm=0) 10: (61) r1 = *(u32 *)(r1 +0) ; R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC) 11: (79) r2 = *(u64 *)(r6 +152) ; R2_w=scalar() R6=ctx(off=0,imm=0) ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC) 12: (55) if r2 != 0xb9fbeef goto pc+10 ; R2_w=195018479 13: (bc) w2 = w1 ; R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) R2_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) ; if (test < __NR_TESTS) 14: (a6) if w1 < 0x9 goto pc+1 16: R0=2 R1_w=scalar(umax=8,var_off=(0x0; 0xf)) R2_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) R6=ctx(off=0,imm=0) R10=fp0 ; 16: (27) r2 *= 28 ; R2_w=scalar(umax=120259084260,var_off=(0x0; 0x1ffffffffc),s32_max=2147483644,u32_max=-4) 17: (18) r3 = 0xffffc900001b2118 ; R3_w=map_value(off=280,ks=4,vs=564,imm=0) 19: (0f) r3 += r2 ; R2_w=scalar(umax=120259084260,var_off=(0x0; 0x1ffffffffc),s32_max=2147483644,u32_max=-4) R3_w=map_value(off=280,ks=4,vs=564,umax=120259084260,var_off=(0x0; 0x1ffffffffc),s32_max=2147483644,u32_max=-4) 20: (61) r2 = *(u32 *)(r3 +0) R3 unbounded memory access, make sure to bounds check any such access processed 97 insns (limit 1000000) max_states_per_insn 1 total_states 10 peak_states 10 mark_read 6 -- END PROG LOAD LOG -- libbpf: prog 'ingress_fwdns_prio100': failed to load: -13 libbpf: failed to load object 'test_tc_dtime' libbpf: failed to load BPF skeleton 'test_tc_dtime': -13 ... At insn 14, with condition 'w1 < 9', register r1 is changed from an arbitrary u32 value to `scalar(umax=8,var_off=(0x0; 0xf))`. Register r2, however, remains as an arbitrary u32 value. Current verifier won't claim r1/r2 equality if the previous mov is alu32 ('w2 = w1'). If r1 upper 32bit value is not 0, we indeed cannot clamin r1/r2 equality after 'w2 = w1'. But in this particular case, we know r1 upper 32bit value is 0, so it is safe to claim r1/r2 equality. This patch exactly did this. For a 32bit subreg mov, if the src register upper 32bit is 0, it is okay to claim equality between src and dst registers. With this patch, the above verification sequence becomes ... 8: (18) r1 = 0xffffc9000048e230 ; R1_w=map_value(off=560,ks=4,vs=564,imm=0) 10: (61) r1 = *(u32 *)(r1 +0) ; R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC) 11: (79) r2 = *(u64 *)(r6 +152) ; R2_w=scalar() R6=ctx(off=0,imm=0) ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC) 12: (55) if r2 != 0xb9fbeef goto pc+10 ; R2_w=195018479 13: (bc) w2 = w1 ; R1_w=scalar(id=6,umax=4294967295,var_off=(0x0; 0xffffffff)) R2_w=scalar(id=6,umax=4294967295,var_off=(0x0; 0xffffffff)) ; if (test < __NR_TESTS) 14: (a6) if w1 < 0x9 goto pc+1 ; R1_w=scalar(id=6,umin=9,umax=4294967295,var_off=(0x0; 0xffffffff)) ... from 14 to 16: R0=2 R1_w=scalar(id=6,umax=8,var_off=(0x0; 0xf)) R2_w=scalar(id=6,umax=8,var_off=(0x0; 0xf)) R6=ctx(off=0,imm=0) R10=fp0 16: (27) r2 *= 28 ; R2_w=scalar(umax=224,var_off=(0x0; 0xfc)) 17: (18) r3 = 0xffffc9000048e118 ; R3_w=map_value(off=280,ks=4,vs=564,imm=0) 19: (0f) r3 += r2 20: (61) r2 = *(u32 *)(r3 +0) ; R2_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) R3_w=map_value(off=280,ks=4,vs=564,umax=224,var_off=(0x0; 0xfc),s32_max=252,u32_max=252) ... and eventually the bpf program can be verified successfully. [1] https://reviews.llvm.org/D147968 Signed-off-by: Yonghong Song <yhs@fb.com> Link: https://lore.kernel.org/r/20230417222134.359714-1-yhs@fb.com Signed-off-by: Alexei Starovoitov <ast@kernel.org> 
In [1], I tried to remove bpf-specific codes to prevent certain
llvm optimizations, and add llvm TTI (target transform info) hooks
to prevent those optimizations. During this process, I found
if I enable llvm SimplifyCFG:shouldFoldTwoEntryPHINode
transformation, I will hit the following verification failure with selftests:

  ...
  8: (18) r1 = 0xffffc900001b2230       ; R1_w=map_value(off=560,ks=4,vs=564,imm=0)
  10: (61) r1 = *(u32 *)(r1 +0)         ; R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff))
  ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC)
  11: (79) r2 = *(u64 *)(r6 +152)       ; R2_w=scalar() R6=ctx(off=0,imm=0)
  ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC)
  12: (55) if r2 != 0xb9fbeef goto pc+10        ; R2_w=195018479
  13: (bc) w2 = w1                      ; R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) R2_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff))
  ; if (test < __NR_TESTS)
  14: (a6) if w1 < 0x9 goto pc+1 16: R0=2 R1_w=scalar(umax=8,var_off=(0x0; 0xf)) R2_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) R6=ctx(off=0,imm=0) R10=fp0
  ;
  16: (27) r2 *= 28                     ; R2_w=scalar(umax=120259084260,var_off=(0x0; 0x1ffffffffc),s32_max=2147483644,u32_max=-4)
  17: (18) r3 = 0xffffc900001b2118      ; R3_w=map_value(off=280,ks=4,vs=564,imm=0)
  19: (0f) r3 += r2                     ; R2_w=scalar(umax=120259084260,var_off=(0x0; 0x1ffffffffc),s32_max=2147483644,u32_max=-4) R3_w=map_value(off=280,ks=4,vs=564,umax=120259084260,var_off=(0x0; 0x1ffffffffc),s32_max=2147483644,u32_max=-4)
  20: (61) r2 = *(u32 *)(r3 +0)
  R3 unbounded memory access, make sure to bounds check any such access
  processed 97 insns (limit 1000000) max_states_per_insn 1 total_states 10 peak_states 10 mark_read 6
  -- END PROG LOAD LOG --
  libbpf: prog 'ingress_fwdns_prio100': failed to load: -13
  libbpf: failed to load object 'test_tc_dtime'
  libbpf: failed to load BPF skeleton 'test_tc_dtime': -13
  ...

At insn 14, with condition 'w1 < 9', register r1 is changed from an arbitrary
u32 value to `scalar(umax=8,var_off=(0x0; 0xf))`. Register r2, however, remains
as an arbitrary u32 value. Current verifier won't claim r1/r2 equality if
the previous mov is alu32 ('w2 = w1').

If r1 upper 32bit value is not 0, we indeed cannot clamin r1/r2 equality
after 'w2 = w1'. But in this particular case, we know r1 upper 32bit value
is 0, so it is safe to claim r1/r2 equality. This patch exactly did this.
For a 32bit subreg mov, if the src register upper 32bit is 0,
it is okay to claim equality between src and dst registers.

With this patch, the above verification sequence becomes

  ...
  8: (18) r1 = 0xffffc9000048e230       ; R1_w=map_value(off=560,ks=4,vs=564,imm=0)
  10: (61) r1 = *(u32 *)(r1 +0)         ; R1_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff))
  ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC)
  11: (79) r2 = *(u64 *)(r6 +152)       ; R2_w=scalar() R6=ctx(off=0,imm=0)
  ; if (skb->tstamp == EGRESS_ENDHOST_MAGIC)
  12: (55) if r2 != 0xb9fbeef goto pc+10        ; R2_w=195018479
  13: (bc) w2 = w1                      ; R1_w=scalar(id=6,umax=4294967295,var_off=(0x0; 0xffffffff)) R2_w=scalar(id=6,umax=4294967295,var_off=(0x0; 0xffffffff))
  ; if (test < __NR_TESTS)
  14: (a6) if w1 < 0x9 goto pc+1        ; R1_w=scalar(id=6,umin=9,umax=4294967295,var_off=(0x0; 0xffffffff))
  ...
  from 14 to 16: R0=2 R1_w=scalar(id=6,umax=8,var_off=(0x0; 0xf)) R2_w=scalar(id=6,umax=8,var_off=(0x0; 0xf)) R6=ctx(off=0,imm=0) R10=fp0
  16: (27) r2 *= 28                     ; R2_w=scalar(umax=224,var_off=(0x0; 0xfc))
  17: (18) r3 = 0xffffc9000048e118      ; R3_w=map_value(off=280,ks=4,vs=564,imm=0)
  19: (0f) r3 += r2
  20: (61) r2 = *(u32 *)(r3 +0)         ; R2_w=scalar(umax=4294967295,var_off=(0x0; 0xffffffff)) R3_w=map_value(off=280,ks=4,vs=564,umax=224,var_off=(0x0; 0xfc),s32_max=252,u32_max=252)
  ...

and eventually the bpf program can be verified successfully.

  [1] https://reviews.llvm.org/D147968

Signed-off-by: Yonghong Song <yhs@fb.com>
Link: https://lore.kernel.org/r/20230417222134.359714-1-yhs@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
bpf: Remove KF_KPTR_GET kfunc flag 2023-04-16T15:51:24+00:00 David Vernet void@manifault.com 2023-04-16T08:49:27+00:00 7b4ddf3920d247c2949073b9c274301c8131332a We've managed to improve the UX for kptrs significantly over the last 9 months. All of the existing use cases which previously had KF_KPTR_GET kfuncs (struct bpf_cpumask *, struct task_struct *, and struct cgroup *) have all been updated to be synchronized using RCU. In other words, their KF_KPTR_GET kfuncs have been removed in favor of KF_RCU | KF_ACQUIRE kfuncs, with the pointers themselves also being readable from maps in an RCU read region thanks to the types being RCU safe. While KF_KPTR_GET was a logical starting point for kptrs, it's become clear that they're not the correct abstraction. KF_KPTR_GET is a flag that essentially does nothing other than enforcing that the argument to a function is a pointer to a referenced kptr map value. At first glance, that's a useful thing to guarantee to a kfunc. It gives kfuncs the ability to try and acquire a reference on that kptr without requiring the BPF prog to do something like this: struct kptr_type *in_map, *new = NULL; in_map = bpf_kptr_xchg(&map->value, NULL); if (in_map) { new = bpf_kptr_type_acquire(in_map); in_map = bpf_kptr_xchg(&map->value, in_map); if (in_map) bpf_kptr_type_release(in_map); } That's clearly a pretty ugly (and racy) UX, and if using KF_KPTR_GET is the only alternative, it's better than nothing. However, the problem with any KF_KPTR_GET kfunc lies in the fact that it always requires some kind of synchronization in order to safely do an opportunistic acquire of the kptr in the map. This is because a BPF program running on another CPU could do a bpf_kptr_xchg() on that map value, and free the kptr after it's been read by the KF_KPTR_GET kfunc. For example, the now-removed bpf_task_kptr_get() kfunc did the following: struct task_struct *bpf_task_kptr_get(struct task_struct **pp) { struct task_struct *p; rcu_read_lock(); p = READ_ONCE(*pp); /* If p is non-NULL, it could still be freed by another CPU, * so we have to do an opportunistic refcount_inc_not_zero() * and return NULL if the task will be freed after the * current RCU read region. */ |f (p && !refcount_inc_not_zero(&p->rcu_users)) p = NULL; rcu_read_unlock(); return p; } In other words, the kfunc uses RCU to ensure that the task remains valid after it's been peeked from the map. However, this is completely redundant with just defining a KF_RCU kfunc that itself does a refcount_inc_not_zero(), which is exactly what bpf_task_acquire() now does. So, the question of whether KF_KPTR_GET is useful is actually, "Are there any synchronization mechanisms / safety flags that are required by certain kptrs, but which are not provided by the verifier to kfuncs?" The answer to that question today is "No", because every kptr we currently care about is RCU protected. Even if the answer ever became "yes", the proper way to support that referenced kptr type would be to add support for whatever synchronization mechanism it requires in the verifier, rather than giving kfuncs a flag that says, "Here's a pointer to a referenced kptr in a map, do whatever you need to do." With all that said -- so as to allow us to consolidate the kfunc API, and simplify the verifier a bit, this patch removes KF_KPTR_GET, and all relevant logic from the verifier. Signed-off-by: David Vernet <void@manifault.com> Link: https://lore.kernel.org/r/20230416084928.326135-3-void@manifault.com Signed-off-by: Alexei Starovoitov <ast@kernel.org> 
We've managed to improve the UX for kptrs significantly over the last 9
months. All of the existing use cases which previously had KF_KPTR_GET
kfuncs (struct bpf_cpumask *, struct task_struct *, and struct cgroup *)
have all been updated to be synchronized using RCU. In other words,
their KF_KPTR_GET kfuncs have been removed in favor of KF_RCU |
KF_ACQUIRE kfuncs, with the pointers themselves also being readable from
maps in an RCU read region thanks to the types being RCU safe.

While KF_KPTR_GET was a logical starting point for kptrs, it's become
clear that they're not the correct abstraction. KF_KPTR_GET is a flag
that essentially does nothing other than enforcing that the argument to
a function is a pointer to a referenced kptr map value. At first glance,
that's a useful thing to guarantee to a kfunc. It gives kfuncs the
ability to try and acquire a reference on that kptr without requiring
the BPF prog to do something like this:

struct kptr_type *in_map, *new = NULL;

in_map = bpf_kptr_xchg(&map->value, NULL);
if (in_map) {
        new = bpf_kptr_type_acquire(in_map);
        in_map = bpf_kptr_xchg(&map->value, in_map);
        if (in_map)
                bpf_kptr_type_release(in_map);
}

That's clearly a pretty ugly (and racy) UX, and if using KF_KPTR_GET is
the only alternative, it's better than nothing. However, the problem
with any KF_KPTR_GET kfunc lies in the fact that it always requires some
kind of synchronization in order to safely do an opportunistic acquire
of the kptr in the map. This is because a BPF program running on another
CPU could do a bpf_kptr_xchg() on that map value, and free the kptr
after it's been read by the KF_KPTR_GET kfunc. For example, the
now-removed bpf_task_kptr_get() kfunc did the following:

struct task_struct *bpf_task_kptr_get(struct task_struct **pp)
{
            struct task_struct *p;

        rcu_read_lock();
        p = READ_ONCE(*pp);
        /* If p is non-NULL, it could still be freed by another CPU,
         * so we have to do an opportunistic refcount_inc_not_zero()
         * and return NULL if the task will be freed after the
         * current RCU read region.
         */
        |f (p && !refcount_inc_not_zero(&p->rcu_users))
                p = NULL;
        rcu_read_unlock();

        return p;
}

In other words, the kfunc uses RCU to ensure that the task remains valid
after it's been peeked from the map. However, this is completely
redundant with just defining a KF_RCU kfunc that itself does a
refcount_inc_not_zero(), which is exactly what bpf_task_acquire() now
does.

So, the question of whether KF_KPTR_GET is useful is actually, "Are
there any synchronization mechanisms / safety flags that are required by
certain kptrs, but which are not provided by the verifier to kfuncs?"
The answer to that question today is "No", because every kptr we
currently care about is RCU protected.

Even if the answer ever became "yes", the proper way to support that
referenced kptr type would be to add support for whatever
synchronization mechanism it requires in the verifier, rather than
giving kfuncs a flag that says, "Here's a pointer to a referenced kptr
in a map, do whatever you need to do."

With all that said -- so as to allow us to consolidate the kfunc API,
and simplify the verifier a bit, this patch removes KF_KPTR_GET, and all
relevant logic from the verifier.

Signed-off-by: David Vernet <void@manifault.com>
Link: https://lore.kernel.org/r/20230416084928.326135-3-void@manifault.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
bpf: Migrate bpf_rbtree_remove to possibly fail 2023-04-16T00:36:50+00:00 Dave Marchevsky davemarchevsky@fb.com 2023-04-15T20:18:09+00:00 404ad75a36fb1a1008e9fe803aa7d0212df9e240 This patch modifies bpf_rbtree_remove to account for possible failure due to the input rb_node already not being in any collection. The function can now return NULL, and does when the aforementioned scenario occurs. As before, on successful removal an owning reference to the removed node is returned. Adding KF_RET_NULL to bpf_rbtree_remove's kfunc flags - now KF_RET_NULL | KF_ACQUIRE - provides the desired verifier semantics: * retval must be checked for NULL before use * if NULL, retval's ref_obj_id is released * retval is a "maybe acquired" owning ref, not a non-owning ref, so it will live past end of critical section (bpf_spin_unlock), and thus can be checked for NULL after the end of the CS BPF programs must add checks ============================ This does change bpf_rbtree_remove's verifier behavior. BPF program writers will need to add NULL checks to their programs, but the resulting UX looks natural: bpf_spin_lock(&glock); n = bpf_rbtree_first(&ghead); if (!n) { /* ... */} res = bpf_rbtree_remove(&ghead, &n->node); bpf_spin_unlock(&glock); if (!res) /* Newly-added check after this patch */ return 1; n = container_of(res, /* ... */); /* Do something else with n */ bpf_obj_drop(n); return 0; The "if (!res)" check above is the only addition necessary for the above program to pass verification after this patch. bpf_rbtree_remove no longer clobbers non-owning refs ==================================================== An issue arises when bpf_rbtree_remove fails, though. Consider this example: struct node_data { long key; struct bpf_list_node l; struct bpf_rb_node r; struct bpf_refcount ref; }; long failed_sum; void bpf_prog() { struct node_data *n = bpf_obj_new(/* ... */); struct bpf_rb_node *res; n->key = 10; bpf_spin_lock(&glock); bpf_list_push_back(&some_list, &n->l); /* n is now a non-owning ref */ res = bpf_rbtree_remove(&some_tree, &n->r, /* ... */); if (!res) failed_sum += n->key; /* not possible */ bpf_spin_unlock(&glock); /* if (res) { do something useful and drop } ... */ } The bpf_rbtree_remove in this example will always fail. Similarly to bpf_spin_unlock, bpf_rbtree_remove is a non-owning reference invalidation point. The verifier clobbers all non-owning refs after a bpf_rbtree_remove call, so the "failed_sum += n->key" line will fail verification, and in fact there's no good way to get information about the node which failed to add after the invalidation. This patch removes non-owning reference invalidation from bpf_rbtree_remove to allow the above usecase to pass verification. The logic for why this is now possible is as follows: Before this series, bpf_rbtree_add couldn't fail and thus assumed that its input, a non-owning reference, was in the tree. But it's easy to construct an example where two non-owning references pointing to the same underlying memory are acquired and passed to rbtree_remove one after another (see rbtree_api_release_aliasing in selftests/bpf/progs/rbtree_fail.c). So it was necessary to clobber non-owning refs to prevent this case and, more generally, to enforce "non-owning ref is definitely in some collection" invariant. This series removes that invariant and the failure / runtime checking added in this patch provide a clean way to deal with the aliasing issue - just fail to remove. Because the aliasing issue prevented by clobbering non-owning refs is no longer an issue, this patch removes the invalidate_non_owning_refs call from verifier handling of bpf_rbtree_remove. Note that bpf_spin_unlock - the other caller of invalidate_non_owning_refs - clobbers non-owning refs for a different reason, so its clobbering behavior remains unchanged. No BPF program changes are necessary for programs to remain valid as a result of this clobbering change. A valid program before this patch passed verification with its non-owning refs having shorter (or equal) lifetimes due to more aggressive clobbering. Also, update existing tests to check bpf_rbtree_remove retval for NULL where necessary, and move rbtree_api_release_aliasing from progs/rbtree_fail.c to progs/rbtree.c since it's now expected to pass verification. Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com> Link: https://lore.kernel.org/r/20230415201811.343116-8-davemarchevsky@fb.com Signed-off-by: Alexei Starovoitov <ast@kernel.org> 
This patch modifies bpf_rbtree_remove to account for possible failure
due to the input rb_node already not being in any collection.
The function can now return NULL, and does when the aforementioned
scenario occurs. As before, on successful removal an owning reference to
the removed node is returned.

Adding KF_RET_NULL to bpf_rbtree_remove's kfunc flags - now KF_RET_NULL |
KF_ACQUIRE - provides the desired verifier semantics:

  * retval must be checked for NULL before use
  * if NULL, retval's ref_obj_id is released
  * retval is a "maybe acquired" owning ref, not a non-owning ref,
    so it will live past end of critical section (bpf_spin_unlock), and
    thus can be checked for NULL after the end of the CS

BPF programs must add checks
============================

This does change bpf_rbtree_remove's verifier behavior. BPF program
writers will need to add NULL checks to their programs, but the
resulting UX looks natural:

  bpf_spin_lock(&glock);

  n = bpf_rbtree_first(&ghead);
  if (!n) { /* ... */}
  res = bpf_rbtree_remove(&ghead, &n->node);

  bpf_spin_unlock(&glock);

  if (!res)  /* Newly-added check after this patch */
    return 1;

  n = container_of(res, /* ... */);
  /* Do something else with n */
  bpf_obj_drop(n);
  return 0;

The "if (!res)" check above is the only addition necessary for the above
program to pass verification after this patch.

bpf_rbtree_remove no longer clobbers non-owning refs
====================================================

An issue arises when bpf_rbtree_remove fails, though. Consider this
example:

  struct node_data {
    long key;
    struct bpf_list_node l;
    struct bpf_rb_node r;
    struct bpf_refcount ref;
  };

  long failed_sum;

  void bpf_prog()
  {
    struct node_data *n = bpf_obj_new(/* ... */);
    struct bpf_rb_node *res;
    n->key = 10;

    bpf_spin_lock(&glock);

    bpf_list_push_back(&some_list, &n->l); /* n is now a non-owning ref */
    res = bpf_rbtree_remove(&some_tree, &n->r, /* ... */);
    if (!res)
      failed_sum += n->key;  /* not possible */

    bpf_spin_unlock(&glock);
    /* if (res) { do something useful and drop } ... */
  }

The bpf_rbtree_remove in this example will always fail. Similarly to
bpf_spin_unlock, bpf_rbtree_remove is a non-owning reference
invalidation point. The verifier clobbers all non-owning refs after a
bpf_rbtree_remove call, so the "failed_sum += n->key" line will fail
verification, and in fact there's no good way to get information about
the node which failed to add after the invalidation. This patch removes
non-owning reference invalidation from bpf_rbtree_remove to allow the
above usecase to pass verification. The logic for why this is now
possible is as follows:

Before this series, bpf_rbtree_add couldn't fail and thus assumed that
its input, a non-owning reference, was in the tree. But it's easy to
construct an example where two non-owning references pointing to the same
underlying memory are acquired and passed to rbtree_remove one after
another (see rbtree_api_release_aliasing in
selftests/bpf/progs/rbtree_fail.c).

So it was necessary to clobber non-owning refs to prevent this
case and, more generally, to enforce "non-owning ref is definitely
in some collection" invariant. This series removes that invariant and
the failure / runtime checking added in this patch provide a clean way
to deal with the aliasing issue - just fail to remove.

Because the aliasing issue prevented by clobbering non-owning refs is no
longer an issue, this patch removes the invalidate_non_owning_refs
call from verifier handling of bpf_rbtree_remove. Note that
bpf_spin_unlock - the other caller of invalidate_non_owning_refs -
clobbers non-owning refs for a different reason, so its clobbering
behavior remains unchanged.

No BPF program changes are necessary for programs to remain valid as a
result of this clobbering change. A valid program before this patch
passed verification with its non-owning refs having shorter (or equal)
lifetimes due to more aggressive clobbering.

Also, update existing tests to check bpf_rbtree_remove retval for NULL
where necessary, and move rbtree_api_release_aliasing from
progs/rbtree_fail.c to progs/rbtree.c since it's now expected to pass
verification.

Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230415201811.343116-8-davemarchevsky@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
bpf: Migrate bpf_rbtree_add and bpf_list_push_{front,back} to possibly fail 2023-04-16T00:36:50+00:00 Dave Marchevsky davemarchevsky@fb.com 2023-04-15T20:18:07+00:00 d2dcc67df910dd85253a701b6a5b747f955d28f5 Consider this code snippet: struct node { long key; bpf_list_node l; bpf_rb_node r; bpf_refcount ref; } int some_bpf_prog(void *ctx) { struct node *n = bpf_obj_new(/*...*/), *m; bpf_spin_lock(&glock); bpf_rbtree_add(&some_tree, &n->r, /* ... */); m = bpf_refcount_acquire(n); bpf_rbtree_add(&other_tree, &m->r, /* ... */); bpf_spin_unlock(&glock); /* ... */ } After bpf_refcount_acquire, n and m point to the same underlying memory, and that node's bpf_rb_node field is being used by the some_tree insert, so overwriting it as a result of the second insert is an error. In order to properly support refcounted nodes, the rbtree and list insert functions must be allowed to fail. This patch adds such support. The kfuncs bpf_rbtree_add, bpf_list_push_{front,back} are modified to return an int indicating success/failure, with 0 -> success, nonzero -> failure. bpf_obj_drop on failure ======================= Currently the only reason an insert can fail is the example above: the bpf_{list,rb}_node is already in use. When such a failure occurs, the insert kfuncs will bpf_obj_drop the input node. This allows the insert operations to logically fail without changing their verifier owning ref behavior, namely the unconditional release_reference of the input owning ref. With insert that always succeeds, ownership of the node is always passed to the collection, since the node always ends up in the collection. With a possibly-failed insert w/ bpf_obj_drop, ownership of the node is always passed either to the collection (success), or to bpf_obj_drop (failure). Regardless, it's correct to continue unconditionally releasing the input owning ref, as something is always taking ownership from the calling program on insert. Keeping owning ref behavior unchanged results in a nice default UX for insert functions that can fail. If the program's reaction to a failed insert is "fine, just get rid of this owning ref for me and let me go on with my business", then there's no reason to check for failure since that's default behavior. e.g.: long important_failures = 0; int some_bpf_prog(void *ctx) { struct node *n, *m, *o; /* all bpf_obj_new'd */ bpf_spin_lock(&glock); bpf_rbtree_add(&some_tree, &n->node, /* ... */); bpf_rbtree_add(&some_tree, &m->node, /* ... */); if (bpf_rbtree_add(&some_tree, &o->node, /* ... */)) { important_failures++; } bpf_spin_unlock(&glock); } If we instead chose to pass ownership back to the program on failed insert - by returning NULL on success or an owning ref on failure - programs would always have to do something with the returned ref on failure. The most likely action is probably "I'll just get rid of this owning ref and go about my business", which ideally would look like: if (n = bpf_rbtree_add(&some_tree, &n->node, /* ... */)) bpf_obj_drop(n); But bpf_obj_drop isn't allowed in a critical section and inserts must occur within one, so in reality error handling would become a hard-to-parse mess. For refcounted nodes, we can replicate the "pass ownership back to program on failure" logic with this patch's semantics, albeit in an ugly way: struct node *n = bpf_obj_new(/* ... */), *m; bpf_spin_lock(&glock); m = bpf_refcount_acquire(n); if (bpf_rbtree_add(&some_tree, &n->node, /* ... */)) { /* Do something with m */ } bpf_spin_unlock(&glock); bpf_obj_drop(m); bpf_refcount_acquire is used to simulate "return owning ref on failure". This should be an uncommon occurrence, though. Addition of two verifier-fixup'd args to collection inserts =========================================================== The actual bpf_obj_drop kfunc is bpf_obj_drop_impl(void *, struct btf_struct_meta *), with bpf_obj_drop macro populating the second arg with 0 and the verifier later filling in the arg during insn fixup. Because bpf_rbtree_add and bpf_list_push_{front,back} now might do bpf_obj_drop, these kfuncs need a btf_struct_meta parameter that can be passed to bpf_obj_drop_impl. Similarly, because the 'node' param to those insert functions is the bpf_{list,rb}_node within the node type, and bpf_obj_drop expects a pointer to the beginning of the node, the insert functions need to be able to find the beginning of the node struct. A second verifier-populated param is necessary: the offset of {list,rb}_node within the node type. These two new params allow the insert kfuncs to correctly call __bpf_obj_drop_impl: beginning_of_node = bpf_rb_node_ptr - offset if (already_inserted) __bpf_obj_drop_impl(beginning_of_node, btf_struct_meta->record); Similarly to other kfuncs with "hidden" verifier-populated params, the insert functions are renamed with _impl prefix and a macro is provided for common usage. For example, bpf_rbtree_add kfunc is now bpf_rbtree_add_impl and bpf_rbtree_add is now a macro which sets "hidden" args to 0. Due to the two new args BPF progs will need to be recompiled to work with the new _impl kfuncs. This patch also rewrites the "hidden argument" explanation to more directly say why the BPF program writer doesn't need to populate the arguments with anything meaningful. How does this new logic affect non-owning references? ===================================================== Currently, non-owning refs are valid until the end of the critical section in which they're created. We can make this guarantee because, if a non-owning ref exists, the referent was added to some collection. The collection will drop() its nodes when it goes away, but it can't go away while our program is accessing it, so that's not a problem. If the referent is removed from the collection in the same CS that it was added in, it can't be bpf_obj_drop'd until after CS end. Those are the only two ways to free the referent's memory and neither can happen until after the non-owning ref's lifetime ends. On first glance, having these collection insert functions potentially bpf_obj_drop their input seems like it breaks the "can't be bpf_obj_drop'd until after CS end" line of reasoning. But we care about the memory not being _freed_ until end of CS end, and a previous patch in the series modified bpf_obj_drop such that it doesn't free refcounted nodes until refcount == 0. So the statement can be more accurately rewritten as "can't be free'd until after CS end". We can prove that this rewritten statement holds for any non-owning reference produced by collection insert functions: * If the input to the insert function is _not_ refcounted * We have an owning reference to the input, and can conclude it isn't in any collection * Inserting a node in a collection turns owning refs into non-owning, and since our input type isn't refcounted, there's no way to obtain additional owning refs to the same underlying memory * Because our node isn't in any collection, the insert operation cannot fail, so bpf_obj_drop will not execute * If bpf_obj_drop is guaranteed not to execute, there's no risk of memory being free'd * Otherwise, the input to the insert function is refcounted * If the insert operation fails due to the node's list_head or rb_root already being in some collection, there was some previous successful insert which passed refcount to the collection * We have an owning reference to the input, it must have been acquired via bpf_refcount_acquire, which bumped the refcount * refcount must be >= 2 since there's a valid owning reference and the node is already in a collection * Insert triggering bpf_obj_drop will decr refcount to >= 1, never resulting in a free So although we may do bpf_obj_drop during the critical section, this will never result in memory being free'd, and no changes to non-owning ref logic are needed in this patch. Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com> Link: https://lore.kernel.org/r/20230415201811.343116-6-davemarchevsky@fb.com Signed-off-by: Alexei Starovoitov <ast@kernel.org> 
Consider this code snippet:

  struct node {
    long key;
    bpf_list_node l;
    bpf_rb_node r;
    bpf_refcount ref;
  }

  int some_bpf_prog(void *ctx)
  {
    struct node *n = bpf_obj_new(/*...*/), *m;

    bpf_spin_lock(&glock);

    bpf_rbtree_add(&some_tree, &n->r, /* ... */);
    m = bpf_refcount_acquire(n);
    bpf_rbtree_add(&other_tree, &m->r, /* ... */);

    bpf_spin_unlock(&glock);

    /* ... */
  }

After bpf_refcount_acquire, n and m point to the same underlying memory,
and that node's bpf_rb_node field is being used by the some_tree insert,
so overwriting it as a result of the second insert is an error. In order
to properly support refcounted nodes, the rbtree and list insert
functions must be allowed to fail. This patch adds such support.

The kfuncs bpf_rbtree_add, bpf_list_push_{front,back} are modified to
return an int indicating success/failure, with 0 -> success, nonzero ->
failure.

bpf_obj_drop on failure
=======================

Currently the only reason an insert can fail is the example above: the
bpf_{list,rb}_node is already in use. When such a failure occurs, the
insert kfuncs will bpf_obj_drop the input node. This allows the insert
operations to logically fail without changing their verifier owning ref
behavior, namely the unconditional release_reference of the input
owning ref.

With insert that always succeeds, ownership of the node is always passed
to the collection, since the node always ends up in the collection.

With a possibly-failed insert w/ bpf_obj_drop, ownership of the node
is always passed either to the collection (success), or to bpf_obj_drop
(failure). Regardless, it's correct to continue unconditionally
releasing the input owning ref, as something is always taking ownership
from the calling program on insert.

Keeping owning ref behavior unchanged results in a nice default UX for
insert functions that can fail. If the program's reaction to a failed
insert is "fine, just get rid of this owning ref for me and let me go
on with my business", then there's no reason to check for failure since
that's default behavior. e.g.:

  long important_failures = 0;

  int some_bpf_prog(void *ctx)
  {
    struct node *n, *m, *o; /* all bpf_obj_new'd */

    bpf_spin_lock(&glock);
    bpf_rbtree_add(&some_tree, &n->node, /* ... */);
    bpf_rbtree_add(&some_tree, &m->node, /* ... */);
    if (bpf_rbtree_add(&some_tree, &o->node, /* ... */)) {
      important_failures++;
    }
    bpf_spin_unlock(&glock);
  }

If we instead chose to pass ownership back to the program on failed
insert - by returning NULL on success or an owning ref on failure -
programs would always have to do something with the returned ref on
failure. The most likely action is probably "I'll just get rid of this
owning ref and go about my business", which ideally would look like:

  if (n = bpf_rbtree_add(&some_tree, &n->node, /* ... */))
    bpf_obj_drop(n);

But bpf_obj_drop isn't allowed in a critical section and inserts must
occur within one, so in reality error handling would become a
hard-to-parse mess.

For refcounted nodes, we can replicate the "pass ownership back to
program on failure" logic with this patch's semantics, albeit in an ugly
way:

  struct node *n = bpf_obj_new(/* ... */), *m;

  bpf_spin_lock(&glock);

  m = bpf_refcount_acquire(n);
  if (bpf_rbtree_add(&some_tree, &n->node, /* ... */)) {
    /* Do something with m */
  }

  bpf_spin_unlock(&glock);
  bpf_obj_drop(m);

bpf_refcount_acquire is used to simulate "return owning ref on failure".
This should be an uncommon occurrence, though.

Addition of two verifier-fixup'd args to collection inserts
===========================================================

The actual bpf_obj_drop kfunc is
bpf_obj_drop_impl(void *, struct btf_struct_meta *), with bpf_obj_drop
macro populating the second arg with 0 and the verifier later filling in
the arg during insn fixup.

Because bpf_rbtree_add and bpf_list_push_{front,back} now might do
bpf_obj_drop, these kfuncs need a btf_struct_meta parameter that can be
passed to bpf_obj_drop_impl.

Similarly, because the 'node' param to those insert functions is the
bpf_{list,rb}_node within the node type, and bpf_obj_drop expects a
pointer to the beginning of the node, the insert functions need to be
able to find the beginning of the node struct. A second
verifier-populated param is necessary: the offset of {list,rb}_node within the
node type.

These two new params allow the insert kfuncs to correctly call
__bpf_obj_drop_impl:

  beginning_of_node = bpf_rb_node_ptr - offset
  if (already_inserted)
    __bpf_obj_drop_impl(beginning_of_node, btf_struct_meta->record);

Similarly to other kfuncs with "hidden" verifier-populated params, the
insert functions are renamed with _impl prefix and a macro is provided
for common usage. For example, bpf_rbtree_add kfunc is now
bpf_rbtree_add_impl and bpf_rbtree_add is now a macro which sets
"hidden" args to 0.

Due to the two new args BPF progs will need to be recompiled to work
with the new _impl kfuncs.

This patch also rewrites the "hidden argument" explanation to more
directly say why the BPF program writer doesn't need to populate the
arguments with anything meaningful.

How does this new logic affect non-owning references?
=====================================================

Currently, non-owning refs are valid until the end of the critical
section in which they're created. We can make this guarantee because, if
a non-owning ref exists, the referent was added to some collection. The
collection will drop() its nodes when it goes away, but it can't go away
while our program is accessing it, so that's not a problem. If the
referent is removed from the collection in the same CS that it was added
in, it can't be bpf_obj_drop'd until after CS end. Those are the only
two ways to free the referent's memory and neither can happen until
after the non-owning ref's lifetime ends.

On first glance, having these collection insert functions potentially
bpf_obj_drop their input seems like it breaks the "can't be
bpf_obj_drop'd until after CS end" line of reasoning. But we care about
the memory not being _freed_ until end of CS end, and a previous patch
in the series modified bpf_obj_drop such that it doesn't free refcounted
nodes until refcount == 0. So the statement can be more accurately
rewritten as "can't be free'd until after CS end".

We can prove that this rewritten statement holds for any non-owning
reference produced by collection insert functions:

* If the input to the insert function is _not_ refcounted
  * We have an owning reference to the input, and can conclude it isn't
    in any collection
    * Inserting a node in a collection turns owning refs into
      non-owning, and since our input type isn't refcounted, there's no
      way to obtain additional owning refs to the same underlying
      memory
  * Because our node isn't in any collection, the insert operation
    cannot fail, so bpf_obj_drop will not execute
  * If bpf_obj_drop is guaranteed not to execute, there's no risk of
    memory being free'd

* Otherwise, the input to the insert function is refcounted
  * If the insert operation fails due to the node's list_head or rb_root
    already being in some collection, there was some previous successful
    insert which passed refcount to the collection
  * We have an owning reference to the input, it must have been
    acquired via bpf_refcount_acquire, which bumped the refcount
  * refcount must be >= 2 since there's a valid owning reference and the
    node is already in a collection
  * Insert triggering bpf_obj_drop will decr refcount to >= 1, never
    resulting in a free

So although we may do bpf_obj_drop during the critical section, this
will never result in memory being free'd, and no changes to non-owning
ref logic are needed in this patch.

Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230415201811.343116-6-davemarchevsky@fb.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>