Re: Per-CPU variables in modules and pahole

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Thu, Dec 10, 2020 at 8:43 AM Jiri Olsa <jolsa@redhat.com> wrote:
>
> On Wed, Dec 09, 2020 at 12:53:44PM -0800, Andrii Nakryiko wrote:
> > Hi,
> >
> > I'm working on supporting per-CPU symbols in BPF/libbpf, and the
> > prerequisite for that is BTF data for .data..percpu data section and
> > variables inside that.
> >
> > Turns out, pahole doesn't currently emit any BTF information for such
> > variables in kernel modules. And the reason why is quite confusing and
> > I can't figure it out myself, so was hoping someone else might be able
> > to help.
> >
> > To repro, you can take latest bpf-next tree and add this to
> > bpf_testmod/bpf_testmod.c inside selftests/bpf:
> >
> > $ git diff bpf_testmod/bpf_testmod.c
> >       diff --git
> > a/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c
> > b/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c
> > index 2df19d73ca49..b2086b798019 100644
> > --- a/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c
> > +++ b/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c
> > @@ -3,6 +3,7 @@
> >  #include <linux/error-injection.h>
> >  #include <linux/init.h>
> >  #include <linux/module.h>
> > +#include <linux/percpu-defs.h>
> >  #include <linux/sysfs.h>
> >  #include <linux/tracepoint.h>
> >  #include "bpf_testmod.h"
> > @@ -10,6 +11,10 @@
> >  #define CREATE_TRACE_POINTS
> >  #include "bpf_testmod-events.h"
> >
> > +DEFINE_PER_CPU(int, bpf_testmod_ksym_dummy1) = -1;
> > +DEFINE_PER_CPU(int, bpf_testmod_ksym_percpu) = 123;
> > +DEFINE_PER_CPU(int, bpf_testmod_ksym_dummy2) = -1;
> > +
> >  noinline ssize_t
> >  bpf_testmod_test_read(struct file *file, struct kobject *kobj,
> >                       struct bin_attribute *bin_attr,
> >
> > 1. So the very first issue (that I'm going to ignore for now) is that
> > if I just added bpf_testmod_ksym_percpu, it would get addr == 0 and
> > would be ignored by the current pahole logic. So we need to fix that
> > for modules. Adding dummy1 and dummy2 takes care of this for now,
> > bpf_testmod_ksym_percpu has offset 4.
>
> I removed that addr zero check in the modules changes but when
> collecting functions, but it's still there in collect_percpu_var

Hao had some reason to skip per-cpu variables with offset 0, maybe he
can comment on that before we change it.

>
> >
> > 2. Second issue is more interesting. Somehow, when pahole iterates
> > over DWARF variables, the address of bpf_testmod_ksym_percpu is
> > reported as 0x10e74, not 4. Which totally confuses pahole because
> > according to ELF symbols, bpf_testmod_ksym_percpu symbol has value 4.
> > I tracked this down to dwarf_getlocation() returning 10e74 as number
> > field in expr.
>
> in which place do you see that address? when I put displayed
> address from collect_percpu_var it shows 4

yes, ELF symbol's value is 4, but when iterating DWARF variables
(0x10e70 + 4) is returned. It does look like a special handling of
modules. I missed that libdw does some special things for specifically
modules. Further debugging yesterday showed that 0x10e70 roughly
corresponds to the offset of .data..per_cpu if you count all the
allocatable data sections that come before it. So I think you are
right. We should probably centralize the logic of kernel module
detection so that we can handle these module vs non-module differences
properly.

>
> not sure this is related but looks like similar issue I had to
> solve for modules functions, as described in the changelog:
> (not merged yet)
>
>     btf_encoder: Detect kernel module ftrace addresses
>
>     ...
>     There's one tricky point with kernel modules wrt Elf object,
>     which we get from dwfl_module_getelf function. This function
>     performs all possible relocations, including __mcount_loc
>     section.
>
>     So addrs array contains relocated values, which we need take
>     into account when we compare them to functions values which
>     are relative to their sections.
>     ...
>
> The 0x10e74 value could be relocated 4.. but it's me guessing,
> because not sure where you see that address exactly


It comes up in cu__encode_btf(), var->ip.addr is not 4, as we expect it to be.

>
> jirka
>