Re: [PATCH 2/2] NFS: avoid deadlocks with loop-back mounted NFS filesystems.

[Trond Myklebust <trond.myklebust@primarydata.com>]

On Wed, Aug 20, 2014 at 9:11 PM, NeilBrown <neilb@suse.de> wrote:
> On Wed, 20 Aug 2014 20:45:16 -0400 Trond Myklebust
> <trond.myklebust@primarydata.com> wrote:
>
>> On Mon, 2014-08-18 at 16:22 +1000, NeilBrown wrote:
>> > Support for loop-back mounted NFS filesystems is useful when NFS is
>> > used to access shared storage in a high-availability cluster.
>> >
>> > If the node running the NFS server fails, some other node can mount the
>> > filesystem and start providing NFS service.  If that node already had
>> > the filesystem NFS mounted, it will now have it loop-back mounted.
>> >
>> > nfsd can suffer a deadlock when allocating memory and entering direct
>> > reclaim.
>> > While direct reclaim does not write to the NFS filesystem it can send
>> > and wait for a COMMIT through nfs_release_page().
>> >
>> > This patch modifies nfs_release_page() and the functions it calls so
>> > that if the COMMIT is sent to the local host (i.e. is routed over the
>> > loop-back interface) then nfs_release_page() will not wait forever for
>> > that COMMIT to complete.  This is achieved using a new flag
>> > FLUSH_COND_CONNECTED.  When this is set the flush is conditional and
>> > will only wait if the client is connected to a non-local server.
>> >
>> > Aborting early should be safe as kswapd uses the same code but never
>> > waits for the COMMIT.
>> >
>> > Always aborting early could upset the balance of memory management so
>> > when the commit was sent to the local host we still wait but with a
>> > 100ms timeout.  This is enough to significantly slow down any process
>> > that is allocating lots of memory and often long enough to let the
>> > commit complete.
>> >
>> > In those rare cases where it is nfsd, or something that nfsd is
>> > waiting for, that is calling nfs_release_page(), 100ms is not so long
>> > that throughput will be seriously affected.
>> >
>> > When fail-over happens a remote (foreign) client will first become
>> > disconnected and then turn into a local client.
>> > To prevent deadlocks from happening at this point, we still have a
>> > timeout when the COMMIT has been sent to a remote client. In this case
>> > the timeout is longer (1 second).
>> >
>> > So when a node that has mounted a remote filesystem loses the
>> > connection, nfs_release_page() will stop blocking and start failing.
>> > Memory allocators will then be able to make progress without blocking
>> > in NFS.  Any process which writes to a file will still be blocked in
>> > balance_dirty_pages().
>> >
>> > This patch makes use of the new 'private' field in
>> > "struct wait_bit_key" to store the start time of a commit, so the
>> > 'action' function called by __wait_on_bit() knows how much longer
>> > it is appropriate to wait.
>> >
>>
>> This puts way too much RPC connection logic in the NFS layer: we really
>> should not have to care. Is there any reason why we could not just use
>> RPC_TASK_SOFTCONN to have the commit fail if the connection is broken?
>
> I tried to keep as much in the RPC layer as I could....
>
> There really is a big difference between talking to an 'nfsd' on the same
> machine and talking to one on a different machine.  In the first case you
> need to be cautious of deadlocks, in the second you don't.  That is a
> difference that matters to NFS, not to RPC.
>
> I guess we could always have a timeout, even when connected to a remote
> server.  We would just end up blocking somewhere else when the server was not
> responding.
>
> I don't think SOFTCONN is really appropriate.  We don't want the COMMIT to
> stop being retried.  We just don't want the memory reclaim code to block
> waiting for the COMMIT.
>
>>
>> Note that all this has to come with a huge WARNING: all your data may be
>> lost even if you think you just fsync()ed it. Is there any difference
>> between doing this and using the 'async' export option on the knfsd
>> side?
>>
>
> I don't think this patch risks losing data at all - that certainly isn't the
> intent.
> This patch only affects the behaviour of ->releasepage, and only allows it to
> fail instead of block.  It is perfectly acceptable for releasepage to fail
> and it doesn't result in data loss.  It just means that the page is busy.
>

So why not just change the behaviour of ->releasepage() to always
initiate a non-blocking commit and then wait up to 1 second for the
PG_private bit to be released?

-- 
Trond Myklebust

Linux NFS client maintainer, PrimaryData

trond.myklebust@primarydata.com