Message-Id: <20250814121504.2784e740a4e6fd4e0dd563d6@kernel.org>
Date: Thu, 14 Aug 2025 12:15:04 +0900
From: Masami Hiramatsu (Google) <mhiramat@...nel.org>
To: Ryan Chung <seokwoo.chung130@...il.com>
Cc: rostedt@...dmis.org, mathieu.desnoyer@...icios.com,
 linux-kernel@...r.kernel.org, linux-trace-kernel@...r.kernel.org,
 linux-kernel-mentees@...ts.linux.dev
Subject: Re: [PATCH] trace/trace_fprobe.c: TODO: handle filter, nofilter or
 symbol list

Hi Ryan,

On Wed, 13 Aug 2025 01:21:01 +0900
Ryan Chung <seokwoo.chung130@...il.com> wrote:

> Resolve TODO in `__register_trace_fprobe()`: 
> parse `tf->symbol` robustly (support `sym!filter` and comma-separated lists), trim tokens, ignore empties, deduplicate symbols, use bulk registration for lists, return `-EEXIST` if already registered, and preserve lockdown/tracepoint deferral semantics.

Thanks for the improvement!
And could you add the new syntax in the document too ?

> 
> Please note that this was my personal interpretation of what TODO
> required here. Welcoming any feedback. 
> 
> Signed-off-by: Ryan Chung <seokwoo.chung130@...il.com>
> ---
>  kernel/trace/trace_fprobe.c | 102 +++++++++++++++++++++++++++++++++++-
>  1 file changed, 100 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/trace/trace_fprobe.c b/kernel/trace/trace_fprobe.c
> index b40fa59159ac..37d4260b9012 100644
> --- a/kernel/trace/trace_fprobe.c
> +++ b/kernel/trace/trace_fprobe.c
> @@ -12,6 +12,8 @@
>  #include <linux/security.h>
>  #include <linux/tracepoint.h>
>  #include <linux/uaccess.h>
> +#include <linux/string.h>
> +#include <linux/slab.h>

Headers should be sorted alphabetically.

>  
>  #include "trace_dynevent.h"
>  #include "trace_probe.h"
> @@ -762,8 +764,104 @@ static int __register_trace_fprobe(struct trace_fprobe *tf)
>  		return __regsiter_tracepoint_fprobe(tf);
>  	}
>  
> -	/* TODO: handle filter, nofilter or symbol list */
> -	return register_fprobe(&tf->fp, tf->symbol, NULL);
> +    /* Parse tf->symbol */

Please make this parse and check as a sub-function instead of new
scope. Also, it should be done in parse_symbol_and_return(), so that
we can handle wrong syntax when parsing it.

> +    {
> +        char *spec, *bang, *p;
> +        int n = 0, w = 0, j, rc;
> +        char **syms = NULL;
> +
> +        spec = kstrdup(tf->symbol, GFP_KERNEL);
> +        if (!spec)
> +            return -ENOMEM;
> +
> +        /* If a '!' exists, treat it as single symbol + filter */
> +        bang = strchr(spec, '!');
> +        if (bang) {
> +            char *sym, *flt;
> +
> +            *bang = '\0';
> +            sym = strim(spec);
> +            flt = strim(bang + 1);

You don't need to do strim, since if there is a space, it
should be parsed already. New syntax must be ',' separated.
My basic syntax for this probe event is;

WORD WORD WORD[:OPTWORD] SUBWORD[,SUBWORD]

OPTWORD is qualifying the previous WORD, SUBWORDs are not
quarifying, but the same-level words. (Currently using "%return"
for the return of the function, that is a special case.)

> +
> +            if (!*sym || !*flt) {
> +                kfree(spec);

Please use __free(kfree) instead of repeating kfree().

> +                return -EINVAL; /* reject empty symbol/filter */

Also, before returning an error, use trace_probe_log_err() to
notice the reason and the place of the error to user.

> +            }
> +
> +            rc = register_fprobe(&tf->fp, sym, flt);
> +            kfree(spec);
> +            return rc;
> +        }
> +
> +        /* Comma list (or single symbol without '!') */
> +        /* First pass: count non-empty tokens */
> +        p = spec;
> +        while (p) {
> +            char *tok = strsep(&p, ",");
> +            if (tok && *strim(tok))
> +                n++;
> +        }
> +
> +        if (n == 0){
> +            kfree(spec);
> +            return -EINVAL;
> +        }
> +
> +        /* Allocate array for pointers into spec (callee copies/consumes) */
> +        syms = kcalloc(n, sizeof(*syms), GFP_KERNEL);
> +        if (!syms) {
> +            kfree(spec);
> +            return -ENOMEM;
> +        }
> +
> +        /* Second pass: fill, skipping empties */

Again, symbol should not have a space.

> +        p = spec;
> +        while (p) {
> +            char *tok = strsep(&p, ",");
> +            char *s;
> +
> +            if (!tok)
> +                break;
> +            s = strim(tok);
> +            if (!*s)
> +                continue;
> +            syms[w++] = s; 
> +        }
> +        
> +        /* Dedup in-place */
> +        for (i = 0; i < w; i++){
> +            if (!syms[i])
> +                continue;
> +            for (j = i + 1; j < w; j++) {
> +                if (syms[j] && !strcmp(syms[i], syms[j]))
> +                    syms[j] = NULL;
> +            }

I think dedup will be done in ftrace, so we don't need to do this
costly operation.

> +        }
> +
> +        /* Compact */
> +        for (i = 0, j = 0; i < w; i++) {
> +            if (syms[i])
> +                syms[j++] = syms[i];
> +        }
> +        w = j;
> +
> +        /* After dedup, ensure we still have at least one symbol */
> +        if (w == 0){
> +            kfree(syms);
> +            kfree(spec);
> +            return -EINVAL;
> +        }
> +
> +        /* Register list or single symbol, using the existing bulk API */
> +        if (w == 1)
> +            rc = register_fprobe(&tf->fp, syms[0], NULL);

Hmm, you might misunderstand this. What you need to do is to classify
the list of symbols with '!' as nofilter, and others as "filter",
and pass those as "register_fprobe(&tf->fp, filter, nofilter)".

Thank you,

> +        else
> +            rc = register_fprobe_syms(&tf->fp, (const char **)syms, w);
> +
> +        kfree(syms);
> +        kfree(spec);
> +        return rc;
> +    }
>  }
>  
>  /* Internal unregister function - just handle fprobe and flags */
> -- 
> 2.43.0
> 


-- 
Masami Hiramatsu (Google) <mhiramat@...nel.org>

Powered by blists - more mailing lists