Re: [PATCH 04/31] perf record, bpf: Create probe points for BPF programs

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Tue, Oct 20, 2015 at 04:12:14PM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Wed, Oct 14, 2015 at 12:41:15PM +0000, Wang Nan escreveu:
> > This patch introduces bpf__{un,}probe() functions to enable callers to
> > create kprobe points based on section names a BPF program. It parses
> > the section names in the program and creates corresponding 'struct
> > perf_probe_event' structures. The parse_perf_probe_command() function is
> > used to do the main parsing work. The resuling 'struct perf_probe_event'
> > is stored into program private data for further using.
> > 
> > By utilizing the new probing API, this patch creates probe points during
> > event parsing.
> > 
> > To ensure probe points be removed correctly, register an atexit hook
> > so even perf quit through exit() bpf__clear() is still called, so probing
> > points are cleared. Note that bpf_clear() should be registered before
> > bpf__probe() is called, so failure of bpf__probe() can still trigger
> > bpf__clear() to remove probe points which are already probed.
> > 
> > strerror style error reporting scaffold is created by this patch.
> > bpf__strerror_probe() is the first error reporting function in bpf-loader.c.
> 
> So, this one, for a non-root user gives me:
> 
> [acme@felicio linux]$ perf record --event /tmp/foo.o sleep 1
> event syntax error: '/tmp/foo.o'
>                      \___ Invalid argument
> 
> (add -v to see detail)
> Run 'perf list' for a list of valid events
> 
>  Usage: perf record [<options>] [<command>]
>     or: perf record [<options>] -- <command> [<options>]
> 
>     -e, --event <event>   event selector. use 'perf list' to list available events
> [acme@felicio linux]$
> 
> --------------------
> 
> I.e. no libbpf error (good!) but then, just an -EINVAL as the "event syntax
> error", which clearly isn't a syntax error, we need to tell the user that he or she
> needs special perfmissions for using sys_bpf() :-)
> 
> As root:
> 
> [root@felicio ~]# perf record --event /tmp/foo.o sleep
> event syntax error: '/tmp/foo.o'
>                      \___ Invalid argument
> 
> (add -v to see detail)
> Run 'perf list' for a list of valid events
> 
>  Usage: perf record [<options>] [<command>]
>     or: perf record [<options>] -- <command> [<options>]
> 
>     -e, --event <event>   event selector. use 'perf list' to list available events
> [root@felicio ~]# ls -la /tmp/foo.o
> -rw-rw-r--. 1 acme acme 824 Oct 20 12:35 /tmp/foo.o
> [root@felicio ~]# file /tmp/foo.o
> /tmp/foo.o: ELF 64-bit LSB relocatable, no machine, version 1 (SYSV), not stripped
> 
> 
> Humm, its something else, this is an ancient kernel, 4.2.0, probably without
> eBPF support? Nope, its there:
> 
> [root@felicio ~]# grep -i sys_bpf /proc/kallsyms
> ffffffff811829d0 T SyS_bpf
> ffffffff811829d0 T sys_bpf
> [root@felicio ~]#
> 
> Its something else, we need to improve this error reporting:
> 
> [root@felicio ~]# perf record -v --event /tmp/foo.o sleep 1
> libbpf: loading /tmp/foo.o
> libbpf: section .strtab, size 60, link 0, flags 0, type=3
> libbpf: section .text, size 0, link 0, flags 6, type=1
> libbpf: section .data, size 0, link 0, flags 3, type=1
> libbpf: section .bss, size 0, link 0, flags 3, type=8
> libbpf: section do_fork, size 16, link 0, flags 6, type=1
> libbpf: found program do_fork
> libbpf: section license, size 4, link 0, flags 3, type=1
> libbpf: license of /tmp/foo.o is GPL
> libbpf: section version, size 4, link 0, flags 3, type=1
> libbpf: kernel version of /tmp/foo.o is 40100
> libbpf: section .symtab, size 96, link 1, flags 0, type=2
> bpf: config program 'do_fork'
> symbol:do_fork file:(null) line:0 offset:0 return:0 lazy:(null)
> bpf: 'do_fork': event name is missing
> event syntax error: '/tmp/foo.o'
>                      \___ Invalid argument
> 
> (add -v to see detail)
> Run 'perf list' for a list of valid events
> 
>  Usage: perf record [<options>] [<command>]
>     or: perf record [<options>] -- <command> [<options>]
> 
>     -e, --event <event>   event selector. use 'perf list' to list available events
> [root@felicio ~]#
> 
> [root@felicio ~]# grep do_fork /proc/kallsyms 
> ffffffff81099ab0 T _do_fork
> ffffffff81ccc800 d do_fork_test
> [root@felicio ~]#
> 
> $ echo '__attribute__((section("_do_fork"), used)) int fork(void *ctx) {return 0;} char _license[] __attribute__((section("license"), used)) = "GPL";int _version __attribute__((section("version"), used)) = 0x40100;' | clang -D__KERNEL__ $CLANG_OPTIONS $KERNEL_INC_OPTIONS -Wno-unused-value -Wno-pointer-sign -working-directory $WORKING_DIR -c - -target bpf -O2 -o /tmp/foo.o
> 
> [root@felicio ~]# perf record  -v --event /tmp/foo.o sleep 1
> libbpf: loading /tmp/foo.o
> libbpf: section .strtab, size 61, link 0, flags 0, type=3
> libbpf: section .text, size 0, link 0, flags 6, type=1
> libbpf: section .data, size 0, link 0, flags 3, type=1
> libbpf: section .bss, size 0, link 0, flags 3, type=8
> libbpf: section _do_fork, size 16, link 0, flags 6, type=1
> libbpf: found program _do_fork
> libbpf: section license, size 4, link 0, flags 3, type=1
> libbpf: license of /tmp/foo.o is GPL
> libbpf: section version, size 4, link 0, flags 3, type=1
> libbpf: kernel version of /tmp/foo.o is 40100
> libbpf: section .symtab, size 96, link 1, flags 0, type=2
> bpf: config program '_do_fork'
> symbol:_do_fork file:(null) line:0 offset:0 return:0 lazy:(null)
> bpf: '_do_fork': event name is missing
> event syntax error: '/tmp/foo.o'
>                      \___ Invalid argument
> 
> (add -v to see detail)
> Run 'perf list' for a list of valid events
> 
>  Usage: perf record [<options>] [<command>]
>     or: perf record [<options>] -- <command> [<options>]
> 
>     -e, --event <event>   event selector. use 'perf list' to list available events
> [root@felicio ~]#
> 
> So it still doesn't work, doesn't look like it is trying to find a vmlinux,
> will look at another patch IIRC is in this patchkit allowing us to tell
> 'perf record' where to find it... But it can as well use kallsyms...


That doesn't seem to be the problem, the last check after
parse_perf_probe_command() is failing:


+       pr_debug("bpf: config program '%s'\n", config_str);
+       err = parse_perf_probe_command(config_str, pev);
+       if (err < 0) {
+               pr_debug("bpf: '%s' is not a valid config string\n",
+                        config_str);
+               err = -EINVAL;
+               goto errout;
+       }
+
+       if (pev->group && strcmp(pev->group, PERF_BPF_PROBE_GROUP)) {
+               pr_debug("bpf: '%s': group for event is set and not '%s'.\n",
+                        config_str, PERF_BPF_PROBE_GROUP);
+               err = -EINVAL;
+               goto errout;
+       } else if (!pev->group)
+               pev->group = strdup(PERF_BPF_PROBE_GROUP);
+
+       if (!pev->group) {
+               pr_debug("bpf: strdup failed\n");
+               err = -ENOMEM;
+               goto errout;
+       }
+
+       if (!pev->event) {
+               pr_debug("bpf: '%s': event name is missing\n",
+                        config_str);
+               err = -EINVAL;
+               goto errout;
+       }
+       pr_debug("bpf: config '%s' is ok\n", config_str);


I.e. pev->event is NULL, is this expected at this point? I'll continue
investigating later.

What I have is at the perf/ebpf branch in my tree.

- Arnaldo