Pu Lehui <pulehui@xxxxxxxxxxxxxxx> writes:
BPF trampoline is the critical infrastructure of the bpf
subsystem, acting as a mediator between kernel functions
and BPF programs. Numerous important features, such as
using ebpf program for zero overhead kernel introspection,
rely on this key component. We can't wait to support bpf
trampoline on RV64. Since RV64 does not support ftrace
direct call yet, the current RV64 bpf trampoline is only
used in bpf context.
As most of riscv cpu support unaligned memory accesses,
we temporarily use patch [1] to facilitate testing. The
test results are as follow, and test_verifier with no
new failure ceses.
- fexit_bpf2bpf:OK
- dummy_st_ops:OK
- xdp_bpf2bpf:OK
[1] https://lore.kernel.org/linux-riscv/20210916130855.4054926-2-chenhuang5@xxxxxxxxxx/
v1:
- Remove the logic of bpf_arch_text_poke supported for
kernel functions. (Kuohai and Björn)
- Extend patch_text for multiple instructions. (Björn)
- Fix OOB issue when image too big. (Björn)
This series is ready to go in as is.
@Palmer I'd like to take this series via the bpf-next tree (as usual),
but note that there are some non-BPF changes as well, related to text
poking.
@Lehui I'd like to see two follow-up patches:
1. Enable kfunc for RV64, by adding:
| bool bpf_jit_supports_kfunc_call(void)
| {
| return true;
| }
2. Remove the checkpatch warning on patch 4:
| WARNING: kfree(NULL) is safe and this check is probably not required
| #313: FILE: arch/riscv/net/bpf_jit_comp64.c:984:
| + if (branches_off)
| + kfree(branches_off);
For the series:
Tested-by: Björn Töpel <bjorn@xxxxxxxxxxxx>
Acked-by: Björn Töpel <bjorn@xxxxxxxxxxxx>