On 4/29/18 4:20 PM, Alexei Starovoitov wrote:
On Sun, Apr 29, 2018 at 03:06:31PM -0700, Yonghong Song wrote:Commit d5a00528b58c ("syscalls/core, syscalls/x86: Rename struct pt_regs-based sys_*() to __x64_sys_*()") renamed a lot of syscall function sys_*() to __x64_sys_*(). This caused several kprobe based samples/bpf tests failing. This patch fixed the problem by using __x64_sys_*(), instead of sys_*(), in bpf program SEC annotations if the target arch is __TARGET_ARCH_x86. Fixes: d5a00528b58c ("syscalls/core, syscalls/x86: Rename struct pt_regs-based sys_*() to __x64_sys_*()") Signed-off-by: Yonghong Song <[email protected]> --- samples/bpf/map_perf_test_kern.c | 32 +++++++++++++++++++++++ samples/bpf/test_current_task_under_cgroup_kern.c | 4 +++ samples/bpf/test_map_in_map_kern.c | 4 +++ samples/bpf/test_probe_write_user_kern.c | 4 +++ samples/bpf/trace_output_kern.c | 4 +++ samples/bpf/tracex2_kern.c | 4 +++ 6 files changed, 52 insertions(+) diff --git a/samples/bpf/map_perf_test_kern.c b/samples/bpf/map_perf_test_kern.c index 2b2ffb9..79f4320 100644 --- a/samples/bpf/map_perf_test_kern.c +++ b/samples/bpf/map_perf_test_kern.c @@ -95,7 +95,11 @@ struct bpf_map_def SEC("maps") lru_hash_lookup_map = { .max_entries = MAX_ENTRIES, };+#ifdef __TARGET_ARCH_x86+SEC("kprobe/__x64_sys_getuid") +#else SEC("kprobe/sys_getuid") +#endifI think it would be better to hack bpf_load.c to add __x64_ automatically when it matches "sys_" in the beginning of kprobe name.
I thought this before but there a few outliers for the particular kernel configuration I have for latest bpf-next: drivers/video/fbdev/core/sysfillrect.c: void sys_fillrect(struct fb_info *p, const struct fb_fillrect *rect) drivers/video/fbdev/core/syscopyarea.c: void sys_copyarea(struct fb_info *p, const struct fb_copyarea *area) drivers/video/fbdev/core/sysimgblt.c: void sys_imageblit(struct fb_info *p, const struct fb_image *image) I am not sure whether any other outliers for other configurations or in the future somebody could introduces a kernel function sys_ but not a syscall. That is why I took the approach as in the patch.
