Re: [PATCH 18/21] fs: nfs: Make nfs boot time y2038 safe

[Trond Myklebust]

On 6/10/16, 09:12, "Anna Schumaker" <Anna.Schumaker@xxxxxxxxxx> wrote:

>On 06/09/2016 05:10 PM, Deepa Dinamani wrote:
>>>> boot_time is represented as a struct timespec.
>>>> struct timespec and CURRENT_TIME are not y2038 safe.
>>>> Overall, the plan is to use timespec64 for all internal
>>>> kernel representation of timestamps.
>>>> CURRENT_TIME will also be removed.
>>>> Use struct timespec64 to represent boot_time.
>>>> And, ktime_get_real_ts64() for the boot_time value.
>>>>
>>>> boot_time is used to construct the nfs client boot verifier.
>>>> This will now wrap in 2106 instead of 2038 on 32-bit systems.
>>>> The server only relies on the value being persistent until
>>>> reboot so the wrapping should be fine.
>>>
>>> We really do not give a damn about wraparound here, since the boot time is
>>> only ever compared for an exact match, and the odds of two reboots occurring
>>> exactly 2^32 * 10^9 nanoseconds apart are cosmically small...
>>> If struct timespec is going away, can we just convert this into a ktime_t?
>> 
>> timespec64 is the same as timespec already on 64 bit machines.
>> But, yes, we can use ktime_t here.
>> 
>> Did you mean the internal storage value or the wire boo_time used for verifier?
>> In case you don't want to change the wire value, then we will have a division
>> operation, every time the verifier needs to be sent.
>
>The verifier is mostly used during mounting, so we don't send too many of them.  I don't think we need to worry about adding an extra division operation here, they're pretty cheap compared to making RPC calls! :)
>

The only requirement for the verifier is that it be unique, so changing the format is not a problem either. 

Cheers
  Trond