[Gluster-devel] Eager-lock and nfs graph generation

[Anand Avati]

On Tue, Feb 19, 2013 at 3:59 AM, Pranith Kumar K <pkarampu at redhat.com>wrote:

>  On 02/19/2013 11:26 AM, Anand Avati wrote:
>
> Thinking over this, looks like there is a problem!
>
> Write-behind guarantees: That a second write request arriving after the
> acknowledgement of a first overlapping request (whether written-behind or
> otherwise) will be guaranteed to be fulfilled in the backend in the same
> order (i.e, the second overlapping request will be "serialized" behind the
> first one in the fulfillment process)
>
> Eager-lock requirement: That write-behind will send no two write requests
> on an overlapping region at the same time.
>
> The requirement-set and guarantee-set have a big overlap, but the
> requirement-set is not a subset.
>
> This is because of O_SYNC writes. write-behind performs
> write-serialization at fulfillment only for written behind requests (which
> get covered under the conflict detection code during liability
> fulfillment). However, if two threads (or apps) issue overlapping O_SYNC
> writes to the same region at approx same time, then write-behind will let
> both of them go by without any kind of serialization, into eager lock,
> violating the assumptions!
>
> I'm wondering if it is a safer idea to implement overlap checks within
> eager-lock code itself rather than depend on write-behind :|
>
> Avati
>
> On Mon, Feb 11, 2013 at 10:07 PM, Anand Avati <anand.avati at gmail.com>wrote:
>
>>
>>
>>  On Mon, Feb 11, 2013 at 9:32 PM, Pranith Kumar K <pkarampu at redhat.com>wrote:
>>
>>>  hi,
>>> Please note that this is a case in theory and I did not run into such
>>> situation, but I feel it is important to address this.
>>> Configuration with 'Eager-lock on" and "write-behind off" should not be
>>> allowed as it leads to lock synchronization problems which lead to data
>>> in-consistency among replicas in nfs.
>>> lets say bricks b1, b2 are in replication.
>>> Gluster Nfs server uses 1 anonymous fd to perform all write-fops. If
>>> eager-lock is enabled in afr, the lock-owner is used as fd's address which
>>> will be same for all write-fops, so there will never be any inodelk
>>> contention. If write-behind is disabled, there can be writes that overlap.
>>> (Does nfs makes sure that the ranges don't overlap?)
>>>
>>> Now imagine the following scenario:
>>> lets say w1, w2 are 2 write fops on same offset and length. w1 with all
>>> '0's and w2 with all '1's. If these 2 write fops are executed in 2
>>> different threads, the order of arrival of write fops on b1 can be w1, w2
>>> where as on b2 it is w2, w1 leading to data inconsistency between the two
>>> replicas. The lock contention will not happen as both lk-owner, transport
>>> are same for these 2 fops.
>>>
>>
>>  Write-behind has to functions - a) performing operations in the
>> background and b) serializing overlapping operations.
>>
>>  While the problem does exist, the specifics are different from what you
>> describe. since all writes coming in from NFS will always use the same
>> anonymous FD, two near-in-time/overlapping writes will never contend with
>> inodelk() but instead the second write will inherit the lock and changelog
>> from the first. In either case, it is a problem.
>>
>>
>>>  We can add a check in glusterd for volume set to disallow such
>>> configuration, BUT by default write-behind is off in nfs graph and by
>>> default eager-lock is on. So we should either turn on write-behind for nfs
>>> or turn off eager-lock by default.
>>>
>>> Could you please suggest how to proceed with this if you agree that I
>>> did not miss any important detail that makes this theory invalid.
>>>
>>
>>  It seems loading write-behind xlator in NFS graph  looks like a simpler
>> solution. eager-locking is crucial for replicated NFS write performance.
>>
>>  Avati
>>
>
>  Shall we disable eager-lock for files opened with O_SYNC, for now?
>

Bad news: the problem is slightly worse than just this. Even with
non-O_SYNC writes, there is a possibility in write-behind where, if a
second overlapping write request comes so close to the first request that,
if wb_enqueue() of the second one happens after wb_enqueue() of the first
write, but before any unwind() after the first wb_enqueue() (i.e
wb_inode->gen is not bumped), then the two write requests can be wound down
together to eager lock.

Avati
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://supercolony.gluster.org/pipermail/gluster-devel/attachments/20130219/e518e4a5/attachment-0001.html>