On Mon, 15 Aug 2022 at 23:26, Vincent Li <vincent.mc.li@xxxxxxxxx> wrote: > > On Mon, Aug 15, 2022 at 3:18 PM Quentin Monnet <quentin@xxxxxxxxxxxxx> wrote: > > > > Hi Vincent, > > > > On Mon, 15 Aug 2022 at 18:46, Vincent Li <vincent.mc.li@xxxxxxxxx> wrote: > > > > > > Hi, > > > > > > I compile and run kernel 5.18.0 in Centos 8 from bpf-next in my dev > > > machine, I also compiled bpftool from bpf-next on same machine, when > > > run bpftool on same machine, I got : > > > > > > ./bpftool feature probe > > > > > > Error: bug: failed to retrieve CAP_BPF status: Invalid argument > > > > > > where bpftool to retrieve CAP_BPF ? from running kernel or from somewhere else? > > > > Yes, bpftool calls cap_get_proc() to get the capabilities of the > > current process. From what I understand of your output, it looks like > > capget() returns CAP_BPF: I believe the "0x1c0" value at the end is > > (1<<(CAP_CHECKPOINT_RESTORE-32)) + (1<<(CAP_BPF-32)) + > > (1<<(CAP_PERFMON-32)). You could probably check this with a more > > recent version of strace. > > > > Then assuming you do retrieve CAP_BPF from capget(), I don't know why > > cap_get_flag() in bpftool fails to retrieve the capability state. It > > would be worth running bpftool in GDB to check what happens. The check > > in libcap is here [0] but I don't see where we would fail to provide > > valid arguments. Just in case, could you please let me know what > > version of libcap you're using when compiling bpftool? > > I think I installed libcap through centos distro > > [root@centos-dev ~]# rpm -qi libcap.x86_64 > > Name : libcap > > Version : 2.26 So we investigated this on Slack. The issue is related to libcap (and to how libcap is built on CentOS); it is fixed in libcap 2.30 and older. For the record, this is the commit that fixed it: https://git.kernel.org/pub/scm/libs/libcap/libcap.git/commit/?id=f1f62a748d7c67361e91e32d26abafbfb03eeee4 Before this, cap_get_flag() would compare its argument "value" (in our case, CAP_BPF == 39) with __CAP_BITS. This __CAP_BITS constant is defined in libcap/cap_names.h, generated by libcap/_makenames.c from the list in libcap/cap_names.list.h. The latter header file is itself generated in libcap/Makefile from the UAPI header at $(KERNEL_HEADERS)/linux/capability.h, which defaults to the local libcap/include/uapi/linux/capability.h. On your CentOS, the libcap version may have been compiled without setting KERNEL_HEADERS to make it point to the correct system UAPI header (or the header could be too old, but looking at it, it seems that it does have CAP_BPF), in which case it defaulted to libcap's version of the header, which in 2.26 stops at CAP_AUDIT (37). In that case, __CAP_BITS is worth 37 and is lower than CAP_BPF, the check in cap_get_flag() fails and we get -EINVAL. The commit referenced above changed the comparison for libcap 2.30+ to compare "value" with __CAP_MAXBITS == 64 instead, which works correctly. Thanks for the report and the shared debug session! Quentin
