All of lore.kernel.org
 help / color / mirror / Atom feed
From: "Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>
To: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
Cc: Toshi Kani <toshi.kani@hp.com>,
	linux-acpi@vger.kernel.org, lenb@kernel.org,
	linux-kernel@vger.kernel.org
Subject: Re: [PATCH 1/2] acpi : cpu hot-remove returns error number when cpu_down() fails
Date: Tue, 10 Jul 2012 13:27:30 +0530	[thread overview]
Message-ID: <4FFBE06A.4090800@linux.vnet.ibm.com> (raw)
In-Reply-To: <4FFBB62E.9010707@jp.fujitsu.com>

On 07/10/2012 10:27 AM, Yasuaki Ishimatsu wrote:
> Hi Toshi,
> 
> 2012/07/10 6:15, Toshi Kani wrote:
>> On Mon, 2012-07-09 at 16:55 +0530, Srivatsa S. Bhat wrote:
>>> On 07/09/2012 08:01 AM, Yasuaki Ishimatsu wrote:
>>>> Hi Srivatsa,
>>>>
>>>> Thank you for your reviewing.
>>>>
>>>> 2012/07/06 18:51, Srivatsa S. Bhat wrote:
>>>>> On 07/06/2012 08:46 AM, Yasuaki Ishimatsu wrote:
>>>>>> Even if cpu_down() fails, acpi_processor_remove() continues to
>>>>>> remove the cpu.
>>>>>
>>>>> Ouch!
>>>>>
>>>>>> But in this case, it should return error number since some process
>>>>>> may run on
>>>>>> the cpu. If the cpu has a running process and the cpu is turned
>>>>>> the power off,
>>>>>> the system cannot work well.
>>>>>>
>>>>>> Signed-off-by: Yasuaki Ishimatsu <isimatu.yasuaki@jp.fujitsu.com>
>>>>>>
>>>>>> ---
>>>>>>    drivers/acpi/processor_driver.c |   18 ++++++++++++------
>>>>>>    1 file changed, 12 insertions(+), 6 deletions(-)
>>>>>>
>>>>>> Index: linux-3.5-rc4/drivers/acpi/processor_driver.c
>>>>>> ===================================================================
>>>>>> --- linux-3.5-rc4.orig/drivers/acpi/processor_driver.c   
>>>>>> 2012-06-25 04:53:04.000000000 +0900
>>>>>> +++ linux-3.5-rc4/drivers/acpi/processor_driver.c    2012-07-05
>>>>>> 21:02:58.711285382 +0900
>>>>>> @@ -610,7 +610,7 @@ err_free_pr:
>>>>>>    static int acpi_processor_remove(struct acpi_device *device,
>>>>>> int type)
>>>>>>    {
>>>>>>        struct acpi_processor *pr = NULL;
>>>>>> -
>>>>>> +    int ret;
>>>>>>
>>>>>>        if (!device || !acpi_driver_data(device))
>>>>>>            return -EINVAL;
>>>>>> @@ -621,8 +621,9 @@ static int acpi_processor_remove(struct
>>>>>>            goto free;
>>>>>>
>>>>>>        if (type == ACPI_BUS_REMOVAL_EJECT) {
>>>>>> -        if (acpi_processor_handle_eject(pr))
>>>>>> -            return -EINVAL;
>>>>>> +        ret = acpi_processor_handle_eject(pr);
>>>>>> +        if (ret)
>>>>>> +            return ret;
>>>>>>        }
>>>>>>
>>>>>>        acpi_processor_power_exit(pr, device);
>>>>>> @@ -841,12 +842,17 @@ static acpi_status acpi_processor_hotadd
>>>>>>
>>>>>>    static int acpi_processor_handle_eject(struct acpi_processor *pr)
>>>>>>    {
>>>>>> -    if (cpu_online(pr->id))
>>>>>> -        cpu_down(pr->id);
>>>>>> +    int ret;
>>>>>> +
>>>>>> +    if (cpu_online(pr->id)) {
>>>>>> +        ret = cpu_down(pr->id);
>>>>>> +        if (ret)
>>>>>> +            return ret;
>>>>>> +    }
>>>>>>
>>>>>
>>>>> Strictly speaking, this is not thorough enough. What prevents someone
>>>>> from onlining that same cpu again, at this point?
>>>>> So, IMHO, you need to wrap the contents of this function inside a
>>>>> get_online_cpus()/put_online_cpus() block, to prevent anyone else
>>>>> from messing with CPU hotplug at the same time.
>>>>
>>>> If I understand your comment by mistake, please let me know.
>>>> If the contents is wrapped a inside
>>>> get_online_cpus()/put_online_cpus() block
>>>> as below, cpu_down() will stop since cpu_down() calls
>>>> cpu_hotplug_begin() and
>>>> cpu_hotplug_begin() waits for cpu_hotplug.refcount to become 0.
>>>>
>>>
>>> You are right. Sorry, I overlooked that.
>>>
>>>> +    get_online_cpus()
>>>> +    if (cpu_online(pr->id)) {
>>>> +        ret = cpu_down(pr->id);
>>>> +        if (ret)
>>>> +            return ret;
>>>> +    }
>>>> +    put_online_cpus()
>>>>
>>>> I think following patch can prevent it correctly. How about the patch?
>>>>
>>>> ---
>>>>   drivers/acpi/processor_driver.c |   12 ++++++++++++
>>>>   kernel/cpu.c                    |    8 +++++---
>>>>   2 files changed, 17 insertions(+), 3 deletions(-)
>>>>
>>>> Index: linux-3.5-rc4/drivers/acpi/processor_driver.c
>>>> ===================================================================
>>>> --- linux-3.5-rc4.orig/drivers/acpi/processor_driver.c    2012-07-09
>>>> 09:59:01.280211202 +0900
>>>> +++ linux-3.5-rc4/drivers/acpi/processor_driver.c    2012-07-09
>>>> 11:05:34.559859236 +0900
>>>> @@ -844,14 +844,26 @@ static int acpi_processor_handle_eject(s
>>>>   {
>>>>       int ret;
>>>>
>>>> +retry:
>>>>       if (cpu_online(pr->id)) {
>>>>           ret = cpu_down(pr->id);
>>>>           if (ret)
>>>>               return ret;
>>>>       }
>>>>
>>>> +    get_online_cpus();
>>>> +    /*
>>>> +     * Someone might online the cpu again at this point. So we
>>>> check that
>>>> +     * cpu has been onlined or not. If cpu is online, we try to
>>>> offline
>>>> +     * the cpu again.
>>>> +     */
>>>> +    if (cpu_online(pr->id)) {
>>>
>>> How about this:
>>>     if (unlikely(cpu_online(pr->id)) {
>>> since the probability of this happening is quite small...
>>>
>>>> +        put_online_cpus();
>>>> +        goto retry;
>>>> +    }
>>>>       arch_unregister_cpu(pr->id);
>>>>       acpi_unmap_lsapic(pr->id);
>>>> +    put_online_cpus();
>>>>       return ret;
>>>>   }
>>>
>>> This retry logic doesn't look elegant, but I don't see any better
>>> method :-(
>>
>> Another possible option is to fail the request instead of retrying it.
> 
> Good idea!! I'll update it.
> 
>>
>> It would be quite challenging to allow on-lining and off-lining
>> operations to run concurrently.  In fact, even if we close this window,
>> there is yet another window right after the new put_online_cpus() call.
> 
> I think if we close the window, another window does not open.
> acpi_unmap_lsapic() sets cpu_present mask to false before new
> put_online_cpus()
> is called. So even if _cpu_up() is called, the function returns -EINAVL by
> following added code.
> 
> @@ -343,11 +343,13 @@ static int __cpuinit _cpu_up(unsigned in
>      unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
>      struct task_struct *idle;
> 
> -    if (cpu_online(cpu) || !cpu_present(cpu))
> -        return -EINVAL;
> -
>      cpu_hotplug_begin();
> 
> +    if (cpu_online(cpu) || !cpu_present(cpu)) {
> +        ret = -EINVAL;
> +        goto out;
> +    }
> +
> 

Right. Yasuaki's patch will ensure that there are no more race conditions
because it does the cpu_present() check after taking the cpu_hotplug.lock.
So I think it is safe and still doable from the kernel's perspective.

But the question is, "should we do it?" I think Toshi's suggestion of failing
the hot-remove request (if we find that the cpu has been onlined again by some
other task) sounds like a good idea for another reason: cpu hotplug is not
initiated by the OS by itself; its requested by the user; so if something is
onlining the cpu back again, the user better take a second look and decide
what exactly he wants to do with that cpu - whether keep it online or
hot-remove it out.

Trying to online as well as hot-remove the same cpu simultaneously sounds like
a crazy thing to do, and returning -EBUSY or -EAGAIN in the hot-remove case
(ie., failing that request) would give a warning to the user and a chance to
reflect upon what exactly he wants to do with the cpu.

So, IMHO, we should protect against the race condition (between cpu_up and
hot-remove) but choose to fail the hot-remove request, and add a comment saying
why we chose to fail the request, even though we could have gone ahead and
completed it.

Regards,
Srivatsa S. Bhat


  reply	other threads:[~2012-07-10  7:59 UTC|newest]

Thread overview: 21+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2012-07-06  3:16 [PATCH 1/2] acpi : cpu hot-remove returns error number when cpu_down() fails Yasuaki Ishimatsu
2012-07-06  3:16 ` Yasuaki Ishimatsu
2012-07-06  3:19 ` [PATCH 2/2] acpi_bus_trim() stops removing devices when failing to remove the device Yasuaki Ishimatsu
2012-07-06  3:19   ` Yasuaki Ishimatsu
2012-07-06  9:51 ` [PATCH 1/2] acpi : cpu hot-remove returns error number when cpu_down() fails Srivatsa S. Bhat
2012-07-09  2:31   ` Yasuaki Ishimatsu
2012-07-09  2:31     ` Yasuaki Ishimatsu
2012-07-09 11:25     ` Srivatsa S. Bhat
2012-07-09 21:15       ` Toshi Kani
2012-07-10  4:57         ` Yasuaki Ishimatsu
2012-07-10  4:57           ` Yasuaki Ishimatsu
2012-07-10  7:57           ` Srivatsa S. Bhat [this message]
2012-07-10  8:23             ` Yasuaki Ishimatsu
2012-07-10  8:23               ` Yasuaki Ishimatsu
2012-07-10 16:32             ` Toshi Kani
2012-07-10  0:13       ` Yasuaki Ishimatsu
2012-07-10  0:13         ` Yasuaki Ishimatsu
2012-07-10  5:14         ` Yasuaki Ishimatsu
2012-07-10  5:14           ` Yasuaki Ishimatsu
2012-07-10  6:52           ` Srivatsa S. Bhat
2012-07-10  6:51         ` Srivatsa S. Bhat

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=4FFBE06A.4090800@linux.vnet.ibm.com \
    --to=srivatsa.bhat@linux.vnet.ibm.com \
    --cc=isimatu.yasuaki@jp.fujitsu.com \
    --cc=lenb@kernel.org \
    --cc=linux-acpi@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=toshi.kani@hp.com \
    /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
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.