bpf, x64: implement retpoline for tail call
authorDaniel Borkmann <daniel@iogearbox.net>
Thu, 8 Mar 2018 15:17:34 +0000 (16:17 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 11 Mar 2018 15:19:47 +0000 (16:19 +0100)
commitea1c4ebe282d6bb6afca4a42bfbfb933c86b264c
tree316ad1544c7398736dda990b247920db1eed1231
parentbb97a65730f975743b2d95b5cecadc5a254c0159
bpf, x64: implement retpoline for tail call

[ upstream commit a493a87f38cfa48caaa95c9347be2d914c6fdf29 ]

Implement a retpoline [0] for the BPF tail call JIT'ing that converts
the indirect jump via jmp %rax that is used to make the long jump into
another JITed BPF image. Since this is subject to speculative execution,
we need to control the transient instruction sequence here as well
when CONFIG_RETPOLINE is set, and direct it into a pause + lfence loop.
The latter aligns also with what gcc / clang emits (e.g. [1]).

JIT dump after patch:

  # bpftool p d x i 1
   0: (18) r2 = map[id:1]
   2: (b7) r3 = 0
   3: (85) call bpf_tail_call#12
   4: (b7) r0 = 2
   5: (95) exit

With CONFIG_RETPOLINE:

  # bpftool p d j i 1
  [...]
  33: cmp    %edx,0x24(%rsi)
  36: jbe    0x0000000000000072  |*
  38: mov    0x24(%rbp),%eax
  3e: cmp    $0x20,%eax
  41: ja     0x0000000000000072  |
  43: add    $0x1,%eax
  46: mov    %eax,0x24(%rbp)
  4c: mov    0x90(%rsi,%rdx,8),%rax
  54: test   %rax,%rax
  57: je     0x0000000000000072  |
  59: mov    0x28(%rax),%rax
  5d: add    $0x25,%rax
  61: callq  0x000000000000006d  |+
  66: pause                      |
  68: lfence                     |
  6b: jmp    0x0000000000000066  |
  6d: mov    %rax,(%rsp)         |
  71: retq                       |
  72: mov    $0x2,%eax
  [...]

  * relative fall-through jumps in error case
  + retpoline for indirect jump

Without CONFIG_RETPOLINE:

  # bpftool p d j i 1
  [...]
  33: cmp    %edx,0x24(%rsi)
  36: jbe    0x0000000000000063  |*
  38: mov    0x24(%rbp),%eax
  3e: cmp    $0x20,%eax
  41: ja     0x0000000000000063  |
  43: add    $0x1,%eax
  46: mov    %eax,0x24(%rbp)
  4c: mov    0x90(%rsi,%rdx,8),%rax
  54: test   %rax,%rax
  57: je     0x0000000000000063  |
  59: mov    0x28(%rax),%rax
  5d: add    $0x25,%rax
  61: jmpq   *%rax               |-
  63: mov    $0x2,%eax
  [...]

  * relative fall-through jumps in error case
  - plain indirect jump as before

  [0] https://support.google.com/faqs/answer/7625886
  [1] https://github.com/gcc-mirror/gcc/commit/a31e654fa107be968b802786d747e962c2fcdb2b

Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
arch/x86/include/asm/nospec-branch.h
arch/x86/net/bpf_jit_comp.c