Re: [PATCH 00/14] bpf: Add trampoline and dispatcher to /proc/kallsyms

On Tue, Feb 11, 2020 at 04:32:23PM -0300, Arnaldo Carvalho de Melo wrote:

Em Mon, Feb 10, 2020 at 05:17:51PM +0100, Jiri Olsa escreveu:

quoted

On Mon, Feb 10, 2020 at 04:51:08PM +0100, Björn Töpel wrote:

quoted

On Sat, 8 Feb 2020 at 16:42, Jiri Olsa [off-list ref] wrote:

quoted

this patchset adds trampoline and dispatcher objects
to be visible in /proc/kallsyms. The last patch also
adds sorting for all bpf objects in /proc/kallsyms.

quoted

quoted

Thanks for working on this!

quoted

quoted

I'm probably missing something with my perf setup; I've applied your
patches, and everything seem to work fine from an kallsyms
perspective:

quoted

quoted

# grep bpf_dispatcher_xdp /proc/kallsyms
...
ffffffffc0511000 t bpf_dispatcher_xdp   [bpf]

However, when I run
# perf top

I still see the undecorated one:
0.90%  [unknown]                  [k] 0xffffffffc0511037

Any ideas?

quoted

yea strange.. it should be picked up from /proc/kallsyms as
fallback if there's no other source, I'll check on that
(might be the problem with perf depending on address going
only higher in /proc/kallsyms, while bpf symbols are at the
end and start over from the lowest bpf address)

anyway, in perf we enumerate bpf_progs via the perf events
PERF_BPF_EVENT_PROG_LOAD,PERF_BPF_EVENT_PROG_UNLOAD interface
together with PERF_RECORD_KSYMBOL_TYPE_BPF events

we might need to add something like:
  PERF_RECORD_KSYMBOL_TYPE_BPF_TRAMPOLINE
  PERF_RECORD_KSYMBOL_TYPE_BPF_DISPATCHER

to notify about the area, I'll check on that

however the /proc/kallsyms fallback should work in any
case.. thanks for report ;-)

We should by now move kallsyms to be the preferred source of symbols,
not vmlinux, right?

Perhaps what is happening is:

[root@quaco ~]# strace -f -e open,openat -o /tmp/bla perf top
[root@quaco ~]# grep vmlinux /tmp/bla
11013 openat(AT_FDCWD, "vmlinux", O_RDONLY) = -1 ENOENT (No such file or directory)
11013 openat(AT_FDCWD, "/boot/vmlinux", O_RDONLY) = -1 ENOENT (No such file or directory)
11013 openat(AT_FDCWD, "/boot/vmlinux-5.5.0+", O_RDONLY) = -1 ENOENT (No such file or directory)
11013 openat(AT_FDCWD, "/usr/lib/debug/boot/vmlinux-5.5.0+", O_RDONLY) = -1 ENOENT (No such file or directory)
11013 openat(AT_FDCWD, "/lib/modules/5.5.0+/build/vmlinux", O_RDONLY) = 152
[root@quaco ~]#

I.e. it is using vmlinux for resolving symbols and he should try with:

[root@quaco ~]# strace -f -e open,openat -o /tmp/bla perf top --ignore-vmlinux
[root@quaco ~]# perf top -h vmlinux

 Usage: perf top [<options>]

    -k, --vmlinux <file>  vmlinux pathname
        --ignore-vmlinux  don't load vmlinux even if found

[root@quaco ~]# grep vmlinux /tmp/bla
[root@quaco ~]#

Historically vmlinux was preferred because it contains function sizes,
but with all these out of the blue symbols, we need to prefer starting
with /proc/kallsyms and, as we do now, continue getting updates via
PERF_RECORD_KSYMBOL.

Humm, but then trampolines don't generate that, right? Or does it? If it
doesn't, then we will know about just the trampolines in place when the
record/top session starts, reparsing /proc/kallsyms periodically seems
excessive?

I plan to extend the KSYMBOL interface to contain trampolines/dispatcher
data, plus we could do some inteligent fallback to /proc/kallsyms in case
vmlinux won't have anything

jirka