A small concern on driver loading and unloading

Keith Busch keith.busch at intel.com
Mon May 19 15:38:03 PDT 2014


I think I agree writing 0 to the CC.SHN when clearing CC.EN should
have no effect. I don't have hardware that behaves any differently in
any case. I'm all for using the procedure that aligns best with the
specification.

On Sun, 18 May 2014, Xuehua Chen wrote:
> Felt this is related to the understanding of the spec. We could think of
> writing 00b means no notification and no effect for the current write
> operation.  Or we could think of writing 00b means notification status
> changed for the current write operation and controller needs to do something.
>
> Didn't see spec mentioned anywhere specifically that bits not changing means
> no further notification. If the controller took the former approach, the
> driver could cause undesirable behavior. Do you know any evidence in spec
> supporting the latter? If yes, could you share that?
>
> Thanks
>
> On Sun, May 18, 2014 at 1:59 PM, Xuehua Chen <xuehua at gmail.com> wrote:
>>> We write whatever the previous CC values were but without CC.EN, so it
>>> should not trigger another shutdown on the device.
>>
>>> I'm not sure we want to do either of those. The comment in the driver
>>> above "nvme_disable_ctrl" indicates that changing the shutdown bits
>>> might cause undesired behavior when the controller is "enabled" and in
>>> a ready state.
>>
>> It seems that controller behaved incorrectly and should be fixed if
>> writing 00b to the
>> field caused undesirable behavior.
>>
>> From spec, 01b means "Normal shutdown notification". The spec does not
>> say if the
>> bits not changed, 01b means no notification. Instead, spec says 00b means
>> "no notification, no effect".
>>
>> So using 00b seems to be safer.
>>
>> Thanks,
>>
>>
>>
>> On Sun, May 18, 2014 at 11:08 AM, Keith Busch <keith.busch at intel.com> wrote:
>>> On Fri, 16 May 2014, Xuehua Chen wrote:
>>>>
>>>> Noticed something strange about the driver unloading and loading and
>>>> would like to ask a question here.
>>>>
>>>> when driver is unloaded, nvme_shutdown_ctrl is called by nvme_remove.
>>>>
>>>> cc = (readl(&dev->bar->cc) & ~NVME_CC_SHN_MASK) | NVME_CC_SHN_NORMAL;
>>>>       writel(cc, &dev->bar->cc);
>>>>
>>>> This set NVME_CC_SHN_NORMAL in CC.
>>>>
>>>> Then during driver loading,  nvme_disable_ctrl is called, which does this.
>>>>
>>>> u32 cc = readl(&dev->bar->cc);
>>>>
>>>> if (cc & NVME_CC_ENABLE)
>>>>        writel(cc & ~NVME_CC_ENABLE, &dev->bar->cc);
>>>>
>>>>
>>>> Checked the NVME spec 1.a  and didn't find it mention anywhere that
>>>> controller will reset Shutdown Notification field after shutdown
>>>> complete(please correct if this is incorrect). So the bits stay there.
>>>
>>>
>>> That should be okay: we are about to clear the shutdown bits right after
>>> controller reset completes and the CSTS.RDY bit is cleared.
>>>
>>>
>>>> When loading the driver, the NVME_CC_SHN_NORMAL is written when
>>>> disabling controller. This can be confusing. Why disabling the
>>>> controller and do a shutdown at the same time.
>>>
>>>
>>> We write whatever the previous CC values were but without CC.EN, so it
>>> should not trigger another shutdown on the device.
>>>
>>>
>>>> Maybe changing
>>>> writel(cc & ~NVME_CC_ENABLE, &dev->bar->cc);
>>>> to
>>>> writel(cc & ~NVME_CC_ENABLE & ~NVME_CC_SHN_MASK , &dev->bar->cc);
>>>> is safer.
>>>>
>>>> Also it seems that
>>>> writel(0 , &dev->bar->cc);
>>>
>>>
>>> I'm not sure we want to do either of those. The comment in the driver
>>> above "nvme_disable_ctrl" indicates that changing the shutdown bits
>>> might cause undesired behavior when the controller is "enabled" and in
>>> a ready state.
>>>
>>>
>>>> can be simpler and the effects are the same. Disabling controller
>>>> reset controller registers including CC to default values. All fields
>>>> of the CC will be gone any way.
>



More information about the Linux-nvme mailing list