Re: [PATCH bpf-next v2] libbpf: also check /sys/kernel/tracing for tracefs files

[Andrii Nakryiko]

On Mon, May 23, 2022 at 2:28 PM Connor O'Brien <connoro@xxxxxxxxxx> wrote:
>
> libbpf looks for tracefs files only under debugfs, but tracefs may be
> mounted even if debugfs is not. When /sys/kernel/debug/tracing is
> absent, try looking under /sys/kernel/tracing instead.
>
> Signed-off-by: Connor O'Brien <connoro@xxxxxxxxxx>
> ---
> v1->v2: cache result of debugfs check.
>
>  src/libbpf.c | 32 +++++++++++++++++++++++++-------
>  1 file changed, 25 insertions(+), 7 deletions(-)
>
> diff --git a/src/libbpf.c b/src/libbpf.c
> index 2262bcd..cc47c52 100644
> --- a/src/libbpf.c
> +++ b/src/libbpf.c
> @@ -9945,10 +9945,22 @@ static void gen_kprobe_legacy_event_name(char *buf, size_t buf_sz,
>                  __sync_fetch_and_add(&index, 1));
>  }
>
> +static bool debugfs_available(void)
> +{
> +       static bool initialized = false, available;
> +
> +       if (!initialized) {
> +               available = !access("/sys/kernel/debug/tracing", F_OK);
> +               initialized = true;
> +       }
> +       return available;
> +}

so thinking about this caching a bit, I'm not so sure we want to cache
this decision. Mounting and unmounting of tracefs can happen after BPF
application starts, so this debugfs_available flag can actually change
while program is running. On the other hand, we don't do this check
all that frequently, only during attach/detach, so it might be ok not
to cache this result at all? WDYT?

> +
>  static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
>                                    const char *kfunc_name, size_t offset)
>  {
> -       const char *file = "/sys/kernel/debug/tracing/kprobe_events";
> +       const char *file = debugfs_available() ? "/sys/kernel/debug/tracing/kprobe_events" :
> +               "/sys/kernel/tracing/kprobe_events";

reading through this patch, it's now quite hard to see what differs,
to be honest. While I do like full file path spelled out, now that we
have two different prefixes it seems better to have prefixes separate.
How about we do

#define TRACEFS_PFX "/sys/kernel/tracing"
#define DEBUGFS_PFX "/sys/kernel/debug/tracing"

and then use that to construct strings. Like in the above example


const char *file = has_debugfs() ? DEBUGFS_PFX "/kprobe_events" :
TRACEFS_PFX "/kprobe_events";

and similarly below. That way at least we see clearly the part that's
not dependent on debugfs/tracefs differences.

>
>         return append_to_file(file, "%c:%s/%s %s+0x%zx",
>                               retprobe ? 'r' : 'p',
> @@ -9958,7 +9970,8 @@ static int add_kprobe_event_legacy(const char *probe_name, bool retprobe,
>
>  static int remove_kprobe_event_legacy(const char *probe_name, bool retprobe)
>  {
> -       const char *file = "/sys/kernel/debug/tracing/kprobe_events";
> +       const char *file = debugfs_available() ? "/sys/kernel/debug/tracing/kprobe_events" :
> +               "/sys/kernel/tracing/kprobe_events";
>
>         return append_to_file(file, "-:%s/%s", retprobe ? "kretprobes" : "kprobes", probe_name);
>  }
> @@ -9968,7 +9981,8 @@ static int determine_kprobe_perf_type_legacy(const char *probe_name, bool retpro
>         char file[256];
>
>         snprintf(file, sizeof(file),
> -                "/sys/kernel/debug/tracing/events/%s/%s/id",
> +                debugfs_available() ? "/sys/kernel/debug/tracing/events/%s/%s/id" :
> +                "/sys/kernel/tracing/events/%s/%s/id",
>                  retprobe ? "kretprobes" : "kprobes", probe_name);
>
>         return parse_uint_from_file(file, "%d\n");
> @@ -10144,7 +10158,8 @@ static void gen_uprobe_legacy_event_name(char *buf, size_t buf_sz,
>  static inline int add_uprobe_event_legacy(const char *probe_name, bool retprobe,
>                                           const char *binary_path, size_t offset)
>  {
> -       const char *file = "/sys/kernel/debug/tracing/uprobe_events";
> +       const char *file = debugfs_available() ? "/sys/kernel/debug/tracing/uprobe_events" :
> +               "/sys/kernel/tracing/uprobe_events";
>
>         return append_to_file(file, "%c:%s/%s %s:0x%zx",
>                               retprobe ? 'r' : 'p',
> @@ -10154,7 +10169,8 @@ static inline int add_uprobe_event_legacy(const char *probe_name, bool retprobe,
>
>  static inline int remove_uprobe_event_legacy(const char *probe_name, bool retprobe)
>  {
> -       const char *file = "/sys/kernel/debug/tracing/uprobe_events";
> +       const char *file = debugfs_available() ? "/sys/kernel/debug/tracing/uprobe_events" :
> +               "/sys/kernel/tracing/uprobe_events";
>
>         return append_to_file(file, "-:%s/%s", retprobe ? "uretprobes" : "uprobes", probe_name);
>  }
> @@ -10164,7 +10180,8 @@ static int determine_uprobe_perf_type_legacy(const char *probe_name, bool retpro
>         char file[512];
>
>         snprintf(file, sizeof(file),
> -                "/sys/kernel/debug/tracing/events/%s/%s/id",
> +                debugfs_available() ? "/sys/kernel/debug/tracing/events/%s/%s/id" :
> +                "/sys/kernel/tracing/events/%s/%s/id",

like here, "events/%s/%s/id" is important to separate, so

snprintf(file, sizeof(file), "%s/events/%s/%s/id",
         has_debugfs() ? DEBUGFS_PFX : TRACEFS_PFX,
         retprobe ? "uretprobes" : "uprobes", probe_name);

seems easier to follow?


>                  retprobe ? "uretprobes" : "uprobes", probe_name);
>
>         return parse_uint_from_file(file, "%d\n");
> @@ -10295,7 +10312,8 @@ static int determine_tracepoint_id(const char *tp_category,
>         int ret;
>
>         ret = snprintf(file, sizeof(file),
> -                      "/sys/kernel/debug/tracing/events/%s/%s/id",
> +                      debugfs_available() ? "/sys/kernel/debug/tracing/events/%s/%s/id" :
> +                      "/sys/kernel/tracing/events/%s/%s/id",
>                        tp_category, tp_name);
>         if (ret < 0)
>                 return -errno;
> --
> 2.36.1.124.g0e6072fb45-goog
>