Re: [PATCH v3] kernel/module: Reschedule while waiting for modules to finish loading

[Prarit Bhargava]

On 5/2/19 5:48 AM, Jessica Yu wrote:
> +++ Prarit Bhargava [01/05/19 17:26 -0400]:
>>
>>
>> On 4/30/19 6:22 PM, Prarit Bhargava wrote:
>>> On a s390 z14 LAR with 2 cpus about stalls about 3% of the time while
>>> loading the s390_trng.ko module.
>>>
>>> Add a reschedule point to the loop that waits for modules to complete
>>> loading.
>>>
>>> v3: cleanup Fixes line.
>>
>> Jessica, even with this additional patch there appears to be some other issues
>> in the module code that are causing significant delays in boot up on large
>> systems.
> 
> Is this limited to only s390? Or are you seeing this on other arches
> as well? And is it limited to specific modules (like s390_trng)?

Other arches.  We're seeing a hang on a new 192 CPU x86_64 box & the
acpi_cpufreq driver.  The system is MUCH faster than any other x86_64 box I've
seen and that's likely why I'm seeing a problem.

> 
>> FWIW, the logic in the original patch is correct.  It's just that there's, as
>> Heiko discovered, some poor scheduling, etc., that is impacting the module
>> loading code after these changes.
> 
> I am really curious to see what these performance regressions look
> like :/ Please update us when you find out more.
> 

I sent Heiko a private v4 RFC last night with this patch (sorry for the
cut-and-paste)

diff --git a/kernel/module.c b/kernel/module.c
index 1c429d8d2d74..a4ef8628f26f 100644
--- a/kernel/module.c
+++ b/kernel/module.c
@@ -3568,12 +3568,12 @@ static int add_unformed_module(struct module *mod)
	mutex_lock(&module_mutex);
	old = find_module_all(mod->name, strlen(mod->name), true);
	if (old != NULL) {
-		if (old->state == MODULE_STATE_COMING
-		    || old->state == MODULE_STATE_UNFORMED) {
+		if (old->state != MODULE_STATE_LIVE) {
			/* Wait in case it fails to load. */
			mutex_unlock(&module_mutex);
-			err = wait_event_interruptible(module_wq,
-					       finished_loading(mod->name));
+			err = wait_event_interruptible_timeout(module_wq,
+					       finished_loading(mod->name),
+					       HZ / 10000);
			if (err)
				goto out_unlocked;
			goto again;

The original module dependency race issue is fixed simply by changing the
conditional to checking !MODULE_STATE_LIVE.  This, unfortunately, exposed some
other problems within the code.

The module_wq is only run when a module fails to load.  It's possible that
the time between the module's failed init() call and running module_wq
(kernel/module.c:3455) takes a while.  Any thread entering the
add_unformed_module() code while the old module is unloading is put to sleep
waiting for the module_wq to execute.

On the 192 thread box I have noticed that the acpi_cpufreq module attempts
to load 392 times (that is not a typo and I am going to try to figure that
problem out after this one).  This means 191 cpus are put to sleep, and one
cpu is executing the acpi_cpufreq module unload which is executing
do_init_module() and is now at

fail_free_freeinit:
        kfree(freeinit);
fail:
        /* Try to protect us from buggy refcounters. */
        mod->state = MODULE_STATE_GOING;
        synchronize_rcu();
        module_put(mod);
        blocking_notifier_call_chain(&module_notify_list,
                                     MODULE_STATE_GOING, mod);
        klp_module_going(mod);
        ftrace_release_mod(mod);
        free_module(mod);
        wake_up_all(&module_wq);
        return ret;
}

The 191 threads cannot schedule and the system is effectively stuck.  It *does*
eventually free itself but in some cases it takes minutes to do so.

A simple fix for this is to, as I've done above, to add a timeout so that
the threads can be scheduled which allows other processes to run.  After
thinking about it a bit more, however, I wonder if a better approach is to
change the mod->state to MODULE_FAILED & running the module_wq immediately so
that the threads can return an error.

I'm experimenting with that now.

P.