This patchset fixes a tailcall hierarchy issue with a better solution than v1[0].
v1 solution stores tail_call_cnt on the stack of bpf prog:
| STACK |
+---------+ RBP
| |
| |
| |
+--| tcc_ptr |
+->| tcc |
| rbx |
+---------+ RSP
v2 solution stores tail_call_cnt on the stack of bpf prog's caller:
| STACK |
| |
| rip |
+->| tcc |
| | rip |
| | rbp |
| +---------+ RBP
| | |
| | |
| | |
+--| tcc_ptr |
| rbx |
+---------+ RSP
With this change, it requires less instructions to resolve this issue.
For more resolving details, please read the following patches.
The issue is confirmed in the discussions of "bpf, x64: Fix tailcall infinite
loop"[1].
Currently, I only resolve this issue on x86. The ones on arm64, s390x and
loongarch are waiting to be resolved. So, the ci pipeline fails to run for this
issue fixing.
Hopefully, this issue on s390x and arm64 will be resolved soon.
v1 -> v2:
* address comments from Stanislav
* Separate moving emit_nops() as first patch.
Links:
[0] https://lore.kernel.org/bpf/20231005145814.83122-1-hffilwlqm@xxxxxxxxx/
[1] https://lore.kernel.org/bpf/6203dd01-789d-f02c-5293-def4c1b18aef@xxxxxxxxx/
Leon Hwang (4):
bpf, x64: Emit nops for X86_PATCH
bpf, x64: Fix tailcall hierarchy
bpf, x64: Load tail_call_cnt pointer
selftests/bpf: Add testcases for tailcall hierarchy fixing
arch/x86/net/bpf_jit_comp.c | 99 +++--
.../selftests/bpf/prog_tests/tailcalls.c | 418 ++++++++++++++++++
.../bpf/progs/tailcall_bpf2bpf_hierarchy1.c | 34 ++
.../bpf/progs/tailcall_bpf2bpf_hierarchy2.c | 55 +++
.../bpf/progs/tailcall_bpf2bpf_hierarchy3.c | 46 ++
5 files changed, 606 insertions(+), 46 deletions(-)
create mode 100644 tools/testing/selftests/bpf/progs/tailcall_bpf2bpf_hierarchy1.c
create mode 100644 tools/testing/selftests/bpf/progs/tailcall_bpf2bpf_hierarchy2.c
create mode 100644 tools/testing/selftests/bpf/progs/tailcall_bpf2bpf_hierarchy3.c
base-commit: 644b54d80d572438a815c05b1bab2b7871e1e5a1
--
2.41.0
