Re: flock() and NFS [Was: Re: [PATCH] locks: rename file-private locks to file-description locks]

["Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com>]

On 04/29/2014 01:34 PM, Jeff Layton wrote:
> On Tue, 29 Apr 2014 11:53:40 +0200
> "Michael Kerrisk (man-pages)" <mtk.manpages@gmail.com> wrote:
> 
>> On 04/29/2014 11:24 AM, NeilBrown wrote:
>>> On Tue, 29 Apr 2014 11:07:16 +0200 "Michael Kerrisk (man-pages)"
>>> <mtk.manpages@gmail.com> wrote:
>>>
>>>> On 04/27/2014 11:28 PM, NeilBrown wrote:
>>>>> On Sun, 27 Apr 2014 13:11:33 +0200 "Michael Kerrisk (man-pages)"
>>>>> <mtk.manpages@gmail.com> wrote:
>>>>>
>>>>>> On Sun, Apr 27, 2014 at 12:04 PM, NeilBrown <neilb@suse.de> wrote:
>>>>>>> On Sun, 27 Apr 2014 11:16:02 +0200 "Michael Kerrisk (man-pages)"
>>>>>>> <mtk.manpages@gmail.com> wrote:
>>>>>>>
>>>>>>>> [Trimming some folk from CC, and adding various NFS people]
>>>>>>>>
>>>>>>>> On 04/27/2014 06:51 AM, NeilBrown wrote:
>>>>>>>>
>>>>>>>> [...]
>>>>>>>>
>>>>>>>>> Note to Michael: The text
>>>>>>>>>    flock() does not lock files over NFS.
>>>>>>>>> in flock(2) is no longer accurate.  The reality is ... complex.
>>>>>>>>> See nfs(5), and search for "local_lock".
>>>>>>>>
>>>>>>>> Ahhh -- I see:
>>>>>>>> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=5eebde23223aeb0ad2d9e3be6590ff8bbfab0fc2
>>>>>>>>
>>>>>>>> Thanks for the heads up.
>>>>>>>>
>>>>>>>> Just in general, it would be great if the flock(2) and fcntl(2) man pages
>>>>>>>> contained correct details for NFS, of course. So, for example, if there
>>>>>>>> are any current gotchas for NFS and fcntl() byte-range locking, I'd like
>>>>>>>> to add those to the fcntl(2) man page.
>>>>>>>
>>>>>>> The only peculiarities I can think of are:
>>>>>>>  - With NFS, locking or unlocking a region forces a flush of any cached data
>>>>>>>    for that file (or maybe for the region of the file).  I'm not sure if this
>>>>>>>    is worth mentioning.
>>>>>>
>>>>>> I agree that it's probably not necessary to mention.
>>>>>>
>>>>>>>  - With NFSv4 the client can lose a lock if it is out of contact with the
>>>>>>>    server for a period of time.  When this happens, any IO to the file by a
>>>>>>>    process which "thinks" it holds a lock will fail until that process closes
>>>>>>>    and re-opens the file.
>>>>>>>    This behaviour is since 3.12.  Prior to that the client might lose and
>>>>>>>    regain the lock without ever knowing thus potentially risking corruption
>>>>>>>    (but only if client and server lost contact for an extended period).
>>>>>>
>>>>>> Do you have a pointer for that commit to 3.12?
>>>>>>
>>>>>
>>>>> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=ef1820f9be27b6ad158f433ab38002ab8131db4d
>>>>>
>>>>> did most of the work while  the subsequent commit
>>>>>
>>>>> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=f6de7a39c181dfb8a2c534661a53c73afb3081cd
>>>>>
>>>>> changed some details, added some documentation, and inverted the default
>>>>> behaviour.
>>>>
>>>> Thanks for that detail. What do you think of the following text for the 
>>>> fcntl(2) man page:
>>>>
>>>>        Before  Linux 3.12, if an NFS client is out of contact with the
>>>>        server for a period of time, it might lose and  regain  a  lock
>>>>        without  ever  being  aware  of the fact.  This scenario poten‐
>>>>        tially risks  data  corruption,  since  another  process  might
>>>>        acquire  a lock in the intervening period and perform file I/O.
>>>>        Since Linux 3.12, if the client loses contact with the  server,
>>>>        any I/O to the file by a process which "thinks" it holds a lock
>>>>        will fail until that process closes and reopens  the  file.   A
>>>>        kernel  parameter,  nfs.recover_lost_locks,  can be set to 1 to
>>>>        obtain the pre-3.12 behavior, whereby the client  will  attempt
>>>>        to  recover  lost  locks when contact is reestablished with the
>>>>        server.  Because of the attendant risk of data corruption, this
>>>>        parameter defaults to 0 (disabled).
>>>>
>>>
>>> Mostly good.
>>>
>>> I'm just a little concerned about "if the client loses contact with the
>>> server" in the middle there.  It is no longer qualified and it isn't clear
>>> that the "for a period of time" qualification still applied.  And we should
>>> probably quantify the period of time - which defaults to 90 seconds.
>>> I don't remember just now the difference between
>>>    /proc/fs/nfsd/nfsv4{lease,grace}time
>>> but this 90 seconds is one of those.
>>>
>>> Also this is NFSv4 specific.  With NFSv3 the failure mode is the reverse.  If
>>> the server loses contact with a client then any lock stays in place
>>> indefinitely ("why can't I read my mail"... I remember it well).
>>>
>>>   Before Linux 3.12, if an NFSv4 client loses contact with the server
>>>   (defined as more than 90 seconds with no communication), it might lose
>>>   and regain ....
>>
>> Thanks, Neil. Changed as you suggest. I'd quite like to mention
>> which of /proc/fs/nfsd/nfsv4{lease,grace}time is relevant here. I had a 
>> quick scan, but could not determine it with complete confidence. My suspicion, 
>> looking at fs/lockd/svcproc.c and fs/lockd/grace.c::locks_in_grace()
>> is that it is /proc/fs/nfsd/nfsv4gracetime that is relevant here. Can anyone
>> confirm?
>>
> 
> The difference here is subtle. The gracetime is how long after a reboot
> should knfsd allow clients to reclaim state (and deny the creation of
> new locks and opens). The leasetime is how long the NFSv4 lease period
> is. There is a relationship between the two that's illustrated in the
> comments above write_gracetime:
> 
> /**
>  * write_gracetime - Set or report current NFSv4 grace period time
>  *
>  * As above, but sets the time of the NFSv4 grace period.
>  *
>  * Note this should never be set to less than the *previous*
>  * lease-period time, but we don't try to enforce this.  (In the common
>  * case (a new boot), we don't know what the previous lease time was
>  * anyway.)
>  */
> 
> The value you're interested in here is the nfsv4leasetime. If the
> client doesn't renew its lease within that period, then it's subject to
> the server giving up on it and dropping any state that it holds on that
> clients' behalf.
> 
> Note that this is not a firm timeout. The server runs a job
> periodically to clean out expired stateful objects, and it's likely
> that there is some time (maybe even up to another whole lease period)
> between when the timeout expires and the job actually runs. If the
> client gets a RENEW in there within that window, its lease will be
> renewed and its state preserved.
> 
> Also note that all of the above just applies to the Linux knfsd. There
> are many other servers in the field and they have different rules for
> dropping state held by clients that have gone AWOL.

Thanks for the detailed explanation, Jeff. I've updated the draft text to
mention nfsv4gracetime. I won't add the subtleties you mention above
(but they'll go into the commit message).

The text is now:

   Record locking and NFS
       Before  Linux  3.12,  if an NFSv4 client loses contact with the
       server for a period of time (defined as more  than  90  seconds
       with no communication), it might lose and regain a lock without
       ever being aware of the fact.  (The period of time after  which
       contact is assumed lost is defined by /proc/fs/nfsd/nfsv4lease‐
       time, which expresses the period in seconds.  The default value
       for  this  file  is  90.)  This scenario potentially risks data
       corruption, since another process might acquire a lock  in  the
       intervening period and perform file I/O.

       Since  Linux  3.12,  if  an NFSv4 client loses contact with the
       server, any I/O to the file by  a  process  which  "thinks"  it
       holds  a  lock  will fail until that process closes and reopens
       the file.  A kernel parameter, nfs.recover_lost_locks,  can  be
       set  to  1  to obtain the pre-3.12 behavior, whereby the client
       will attempt to recover lost locks  when  contact  is  reestab‐
       lished  with the server.  Because of the attendant risk of data
       corruption, this parameter defaults to 0 (disabled).

Cheers,

Michael



-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/