Re: [RFC PATCH] ftrace: speed up module enable by binary searching module symbols

[Yaxiong Tian <tianyaxiong@kylinos.cn>]

On 2026/10/19 23:25, Masami Hiramatsu (Google) wrote:
> Hi,
>
> On Wed, 16 Sep 2026 16:27:15 +0800, Yaxiong Tian <tianyaxiong@kylinos.cn> wrote:
>
>> diff --git a/kernel/trace/ftrace.c b/kernel/trace/ftrace.c
>> --- a/kernel/trace/ftrace.c
>> +++ b/kernel/trace/ftrace.c
>> @@ -8010,11 +8011,137 @@ void ftrace_release_mod(struct module *mod)
>> +
>> +	/* mod->kallsyms is not replaced until do_init_module() returns. */
>> +	rcu_read_lock();
>> +	kallsyms = rcu_dereference(mod->kallsyms);
> Consider using rcu_access_pointer(mod->kallsyms) or
> rcu_dereference_raw(mod->kallsyms) rather than a transient
> rcu_read_lock()/rcu_read_unlock() pair.
Agreed, I will use rcu_dereference_raw() and drop the lock pair.

>> +	rcu_read_unlock();
>> +
>> +	if (!kallsyms || kallsyms->num_symtab < 2)
>> [...]
>> +	 * ftrace never came up, where the loop below would not run anyway; that
>> +	 * also makes reading ftrace_disabled without the lock safe.
>> +	 */
>> +	syms = ftrace_disabled ? NULL : ftrace_mod_syms_alloc(mod);
> Because ftrace_disabled is one-way flag, if you find ftrace_disabled here,
> you can return from this function soon.
>
> Thanks,
Right. I will check it once before the allocation and jump to a common
out_cached_mods label, so ftrace_lock is not taken at all in that case and
process_cached_mods() still runs exactly once. I keep the second check
under the lock because ftrace_kill() can set the flag after the unlocked
read.

    if (ftrace_disabled)
        goto out_cached_mods;

    syms = ftrace_mod_syms_alloc(mod);

    mutex_lock(&ftrace_lock);

    if (ftrace_disabled)
        goto out_unlock;
    ...
 out_unlock:
    mutex_unlock(&ftrace_lock);
    kvfree(syms);
 out_cached_mods:
    process_cached_mods(mod->name);

>> +
>>  	mutex_lock(&ftrace_lock);
>>