LKML Archive on lore.kernel.org
help / color / mirror / Atom feed
From: Barret Rhoden <brho@google.com>
To: Prarit Bhargava <prarit@redhat.com>, Jessica Yu <jeyu@kernel.org>
Cc: linux-kernel@vger.kernel.org,
	Heiko Carstens <heiko.carstens@de.ibm.com>,
	David Arcari <darcari@redhat.com>
Subject: Re: [PATCH v3] kernel/module: Reschedule while waiting for modules to finish loading
Date: Fri, 10 May 2019 14:40:30 -0400	[thread overview]
Message-ID: <be47ac01-a5ac-7be1-d387-5c841007b45f@google.com> (raw)
In-Reply-To: <4849a5af-b2bb-95ad-6b58-2f0c403c9ebb@redhat.com>

Hi -

On 5/2/19 1:46 PM, Prarit Bhargava wrote:
> On 5/2/19 8:41 AM, Prarit Bhargava wrote:
>> 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 taking a much closer look the above patch appears to be correct.  I am not
> seeing any boot failures associated with it anywhere.  I would like to hear from
> Heiko as to whether or not this works for him though.

I think I found the issue here.  The original patch changed the state 
check to "not LIVE", which made it include GOING.  However, the wake 
condition was not changed.  That could lead to a livelock, which I 
experienced.

I have a patch that fixes it, which I'll send out shortly.  With my 
change, I think you won't need any of the scheduler functions 
(cond_resched(), wait timeouts, etc).  Those were probably just papering 
over the issue.

Barret


  reply	other threads:[~2019-05-10 18:40 UTC|newest]

Thread overview: 13+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-04-30 22:22 Prarit Bhargava
2019-05-01  7:49 ` Jessica Yu
2019-05-01 21:26 ` Prarit Bhargava
2019-05-02  9:48   ` Jessica Yu
2019-05-02 12:41     ` Prarit Bhargava
2019-05-02 17:46       ` Prarit Bhargava
2019-05-10 18:40         ` Barret Rhoden [this message]
2019-05-10 18:42           ` [PATCH] modules: fix livelock in add_unformed_module() Barret Rhoden
2019-05-13 11:23             ` Prarit Bhargava
2019-05-13 14:37               ` Barret Rhoden
2019-05-22 17:08                 ` Prarit Bhargava
2019-05-28 14:30                   ` Prarit Bhargava
2019-05-28 14:47                     ` Jessica Yu

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=be47ac01-a5ac-7be1-d387-5c841007b45f@google.com \
    --to=brho@google.com \
    --cc=darcari@redhat.com \
    --cc=heiko.carstens@de.ibm.com \
    --cc=jeyu@kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=prarit@redhat.com \
    --subject='Re: [PATCH v3] kernel/module: Reschedule while waiting for modules to finish loading' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).