Re: [PATCH/RFC] Re: 'perf test BPF' failing, libbpf regression wrt "basic API for BPF obj name"

From: Martin KaFai Lau
Date: Thu Nov 30 2017 - 17:10:11 EST


On Thu, Nov 30, 2017 at 04:00:42PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Thu, Nov 30, 2017 at 10:28:08AM -0800, Martin KaFai Lau escreveu:
> > On Thu, Nov 30, 2017 at 01:53:58PM -0300, Arnaldo Carvalho de Melo wrote:
> > > Em Thu, Nov 30, 2017 at 12:01:10AM -0300, Arnaldo Carvalho de Melo escreveu:
> > > > Em Wed, Nov 29, 2017 at 02:31:36PM -0800, Martin KaFai Lau escreveu:
> > > > > On Wed, Nov 29, 2017 at 06:15:43PM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > > Em Wed, Nov 29, 2017 at 01:07:34PM -0800, Martin KaFai Lau escreveu:
> > > > > > > On Tue, Nov 28, 2017 at 04:05:19PM -0300, Arnaldo Carvalho de Melo wrote:
> > > > > > > > [root@jouet ~]# perf test -v bpf
> > > > > > > > 39: BPF filter :
> > > > > > > > 39.1: Basic BPF filtering :
> > > > > > > > Kernel build dir is set to /lib/modules/4.14.0+/build
> > > > > > > [ ... ]
> > > > > > > > libbpf: failed to create map (name: 'flip_table'): Invalid argument
> > > > > > > > libbpf: failed to load object '[basic_bpf_test]'
> > > > > > > > bpf: load objects failed
> > > > > > > 88cda1c9da02 ("bpf: libbpf: Provide basic API support to specify BPF obj name")
> > > > > > > is introduced in 4.15.
> > > >
> > > > > > > I think the perf@xxxxxxxxxxx broke on older kernels like 4.14 because
> > > > > > > the new bpf prog/map name is only introduced since 4.15.
>
> > > > > > > The newer perf needs to be compatible with an older kernel?
>
> > > > > > Sure :-)
>
> > > > > Would the latest features introduced in perf/libbpf supposed to be
> > > > > available in the latest kernel only? What may be the reason that the
>
> > > > Yes, then the new perf binary should try to use the new stuff, if it
> > > > fails, use the old one, there is no requirement that one uses perf 4.14
> > > > in lockstep with the kernel 4.14 (or any other version), perf 4.15
> > > > should work with the 4.14 kernel as well as with 4.15 (or any other
> > > > future kernel), only limited by what it can grok up to when it was
> > > > released.
>
> > > So, see the patch below, that makes a 'perf test bpf' and my other test
> > > cases, including that one for probe_read_str() work again, it just
> > > fallbacks to a behaviour the older kernels can accept.
>
> > Thanks for the patch.
>
> > > We can improve it so that that EINVAL fallback happens only for
> > > MAP_CREATE, and probably we don't need to change the size arg, just zero
> > > the unused fields, but I haven't checked that.
> > >
> > > diff --git a/tools/lib/bpf/bpf.c b/tools/lib/bpf/bpf.c
> > > index 5128677e4117..3084f07c7c33 100644
> > > --- a/tools/lib/bpf/bpf.c
> > > +++ b/tools/lib/bpf/bpf.c
> > > @@ -19,6 +19,7 @@
> > > * License along with this program; if not, see <http://www.gnu.org/licenses>
> > > */
> > >
> > > +#include <errno.h>
> > > #include <stdlib.h>
> > > #include <memory.h>
> > > #include <unistd.h>
> > > @@ -53,10 +54,26 @@ static inline __u64 ptr_to_u64(const void *ptr)
> > > return (__u64) (unsigned long) ptr;
> > > }
> > >
> > > -static inline int sys_bpf(enum bpf_cmd cmd, union bpf_attr *attr,
> > > - unsigned int size)
> > > +static int sys_bpf(enum bpf_cmd cmd, union bpf_attr *attr, unsigned int size)
> > > {
> > > - return syscall(__NR_bpf, cmd, attr, size);
> > > + int err = syscall(__NR_bpf, cmd, attr, size);
> > > + if (err == -1 && (errno == EINVAL || errno == E2BIG)) {
> > I would add a check to the length of map_name/prog_name.
>
> Can you elaborate? What kind of check?
F.e. if map_name is not set (i.e. strlen is 0), there is no
need to retry.

>
> > > + const unsigned int old_union_size = offsetof(union bpf_attr, prog_name);
> > > + /*
> > > + * These were the ones that added fields after the old bpf_attr
> > > + * layout in commit 88cda1c9da02 ("bpf: libbpf: Provide basic
> > > + * API support to specify BPF obj name") so zero that out to
> > > + * pass the CHECK_ATTR() test in kernel/bpf/syscall.c in older
> > > + * kernels.
> > > + */
> > > + if (cmd == BPF_MAP_CREATE)
> > > + memset(&attr->map_name, 0, size - offsetof(union bpf_attr, map_name));
> > > + else
> > > + memset(&attr->prog_name, 0, size - old_union_size);
>
> > If bpf_attr is extended in the future, map_name/prog_name will still be
> > used as the anchor for backward compatibility instead of trial and error
> > attribute by attribute?
>
> Then you will first try the latest and greatest, if it fails, go the
> previous (like here), if it fails, etc.
>
> That or some sort of versioning to make sure the kernel and the tools
> can agree on a common set of functionality supported by both.
>
> Again, this is how perf_evsel__open() works in the perf case, try the
> latest and go fallbacking to the most recent set of features that could
> somehow service what is needed or disable some feature and warn the
> user, i.e. do the best you can with what you have.
>
> With this patch in place I was able to have what was working before
> 88cda1c9da02 working again.
>
> > Instead of sinking all future bpf_attr's backward compatibility
> > requirements to sys_bpf, I would push it up to its own BPF_* command
> > helper which has a better sense of its bpf_attr, i.e. push it up
> > to bpf_create_map_node() and bpf_load_program_name() in this case.
>
> Humm, we could try that approach, but the one in this patch seemed good
> enough.
>
> And after all if the first syscall() invokation, with the latest kernel
> and latest tooling will work, right?
>
> I think that we need as well bpf__strerror_METHOD() that can make sense
> of error returns per method (aka "command helper"), to tell the user
> when something can't be done why is that so: "you need a kernel >= 4.X",
> "this needs root privs", "too many maps, etc"
>
> Right now we have a generic, libbpf wide one:
>
> static const char *libbpf_strerror_table[NR_ERRNO] = {
> [ERRCODE_OFFSET(LIBELF)] = "Something wrong in libelf",
> [ERRCODE_OFFSET(FORMAT)] = "BPF object format invalid",
> [ERRCODE_OFFSET(KVERSION)] = "'version' section incorrect or lost",
> [ERRCODE_OFFSET(ENDIAN)] = "Endian mismatch",
> [ERRCODE_OFFSET(INTERNAL)] = "Internal error in libbpf",
> [ERRCODE_OFFSET(RELOC)] = "Relocation failed",
> [ERRCODE_OFFSET(VERIFY)] = "Kernel verifier blocks program loading",
> [ERRCODE_OFFSET(PROG2BIG)] = "Program too big",
> [ERRCODE_OFFSET(KVER)] = "Incorrect kernel version",
> [ERRCODE_OFFSET(PROGTYPE)] = "Kernel doesn't support this program type",
> };
>
> int libbpf_strerror(int err, char *buf, size_t size)
> {
>
> /* use the above or return strerror_r() for the errno usual range */
> }
>
> But as time goes by, features get added, ABI gets extended, and without
> a better kernel/user error reporting mechanism... Making sense of errno
> in the context of a specific method and looking at the system state
> looks like the best we can do to provide not-so-cryptic messages to the
> poor users, and we're all users :-)
>
> - Arnaldo
>
> > > + err = syscall(__NR_bpf, cmd, attr, old_union_size);
> > > + }
> > > + return err;
> > > }
> > >
> > > int bpf_create_map_node(enum bpf_map_type map_type, const char *name,