Re: [PATCH/RFC] NFS: Minimize COMMIT calls during writeback.

[Trond Myklebust <trondmy@hammerspace.com>]

On Mon, 2020-03-16 at 14:39 +1100, NeilBrown wrote:
> On Mon, Mar 16 2020, Trond Myklebust wrote:
> 
> > On Mon, 2020-03-16 at 11:05 +1100, NeilBrown wrote:
> > > Since 4.13 (see Fixes: commit) NFS has sent a COMMIT request for
> > > every
> > > .writepages() call it received - when the writes submitted for
> > > that
> > > call
> > > have all completed.
> > > 
> > > This works well enough when COMMIT is fast, but if COMMIT is slow
> > > these
> > > calls can overlap each other, overload the server, and
> > > substantially
> > > reduce throughput.
> > > 
> > > I looked at this due to a report of regression when writing to a
> > > Ganesha
> > > NFS server with tracing showing multiple overlapping COMMITs, and
> > > individual commits typically taking 200ms to 300ms.  This
> > > resulted in
> > > 2
> > > orders of magnitude slow down when writing 1000x1M from /dev/zero
> > > with dd.
> > > This is easily reproduced by adding 'msleep(300)' to
> > > nfsd_commit() in
> > > the
> > > Linux NFS server.
> > 
> > When there is overlap of writeback to the same file, then it is
> > almost
> > always because we're hitting memory reclaim on the client. In that
> > case
> > we want to ensure that memory reclaim completes, which it won't do
> > if
> > we delay COMMIT.
> > IOW: this behaviour is very much intentional.
> 
> What do you mean by "overlap of writeback" ?
> The ->writepages calls are sequential - they don't overlap.
> The WRITE rpcs are asynchronous so of course they overlap.
> The COMMITs are expected to overlap with subsequent WRITEs.
> It is only when COMMITs overlap with COMMITs that I think we might
> overload the server, and I cannot see how that would ever be a useful
> thing to do.
> 

I mean that the NFS writeback code has exactly two cases where it
starts to write out data before an fsync() or close() forces it to do
so:

1) Read of a page that was written to, but not up to date
2) Memory pressure.

IOW: If you're seeing overlapping commits, it is due to a situation
where we're trying to free up the pages as soon as possible.

> In the cases I've experiment with there is plenty of memory, but the
> total write size exceeds dirty_threshold, so writeout starts to limit
> the amount of dirty memory.  Certainly we want the COMMIT promptly as
> a group of writes finish, but that group doesn't need to be tiny.
> 
> (Our initial explorations found that increasing dirty_threshold to
> larger than the write size restored normal performance)
> 

That's the definition of memory pressure here. It's not the job of the
filesystem to second guess what is going on in the higher layers. Its
job is to write out data as quickly as possible when asked to do so.

> > > This patch changes the details of when COMMIT is sent without
> > > changing
> > > the general approach.
> > > 
> > > Where previously the writes from a single .writepages() call were
> > > accounted together in a nfs_io_completion, now the writes from
> > > multiple
> > > writepages() calls are accounted together.  The set of writepages
> > > calls
> > > that are grouped together are all those from when one COMMIT call
> > > ends
> > > to when the next COMMIT call ends.  This automatically reduces
> > > the
> > > rate
> > > of COMMIT requests when COMMIT itself is slow.
> > > (If there are no COMMIT calls pending, the first .writepages will
> > > get
> > >  an nfs_io_completion to itself, then subsequenct writepages
> > > requests
> > >  until the first COMMIT completes will go in the next
> > > nfs_io_completion)
> > > 
> > > There are typically at most two nfs_io_completion structures
> > > allocated
> > > for writeback to a given inode.
> > > 
> > > - If there was been any writepages calls since the last time a
> > > COMMIT
> > >   completed, there will be an nfs_io_completion stored in the
> > > inode
> > > (in
> > >   nfs_mds_commit_info) which new writepages requests are
> > > accounted
> > > it.
> > > 
> > > - If there is no pending COMMIT request, but there are pending
> > > writeback
> > >   WRITES, there will be another nfs_io_completion which is not
> > > attached
> > >   and which is draining.  When it fully drains a COMMIT will be
> > > sent.
> > >   When that COMMIT completes, the attached nfs_io_completion will
> > > be
> > >   detached and allowed to drain.
> > > 
> > > The rpcs_out counter now counts the unattached nfs_io_completion
> > > as
> > > well
> > > as any pending COMMIT requests.  As an unattached
> > > nfs_io_completion
> > > will
> > > soon turn into a COMMIT request, this seems reasonable.  It
> > > allows us
> > > to
> > > clearly detect when there are no longer any COMMITs expected to
> > > complete, so we know to detach any nfs_io_completion from the
> > > inode
> > > and
> > > allow it to drain.
> > > 
> > > As we use atomic accesses (e.g.  xchg and kref_get_unless_zero())
> > > to
> > > access the attached nfs_io_completion, we now use kfree_rcu() to
> > > free
> > > it, to ensure it is not accessed after it is freed.
> > > 
> > > With 300ms commits, this improves throught of a 1G dd by 2 orders
> > > of
> > > magnitude.  Even without the 300ms delay, this noticeably
> > > improves
> > > throughput to a Linux NFS server is a simple VM.
> > > 
> > > Fixes: 919e3bd9a875 ("NFS: Ensure we commit after writeback is
> > > complete")
> > > Signed-off-by: NeilBrown <neilb@suse.de>
> > > ---
> > >  fs/nfs/inode.c          |  1 +
> > >  fs/nfs/write.c          | 50
> > > ++++++++++++++++++++++++++++++++++++---
> > > --
> > >  include/linux/nfs_xdr.h |  7 ++++++
> > >  3 files changed, 53 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> > > index 11bf15800ac9..c00b54491949 100644
> > > --- a/fs/nfs/inode.c
> > > +++ b/fs/nfs/inode.c
> > > @@ -2110,6 +2110,7 @@ static void init_once(void *foo)
> > >  	INIT_LIST_HEAD(&nfsi->commit_info.list);
> > >  	atomic_long_set(&nfsi->nrequests, 0);
> > >  	atomic_long_set(&nfsi->commit_info.ncommit, 0);
> > > +	nfsi->commit_info.ioc = NULL;
> > >  	atomic_set(&nfsi->commit_info.rpcs_out, 0);
> > >  	init_rwsem(&nfsi->rmdir_sem);
> > >  	mutex_init(&nfsi->commit_mutex);
> > > diff --git a/fs/nfs/write.c b/fs/nfs/write.c
> > > index c478b772cc49..57e209f964e4 100644
> > > --- a/fs/nfs/write.c
> > > +++ b/fs/nfs/write.c
> > > @@ -47,6 +47,7 @@ struct nfs_io_completion {
> > >  	void (*complete)(void *data);
> > >  	void *data;
> > >  	struct kref refcount;
> > > +	struct rcu_head rcu;
> > >  };
> > >  
> > >  /*
> > > @@ -134,7 +135,7 @@ static void nfs_io_completion_release(struct
> > > kref
> > > *kref)
> > >  	struct nfs_io_completion *ioc = container_of(kref,
> > >  			struct nfs_io_completion, refcount);
> > >  	ioc->complete(ioc->data);
> > > -	kfree(ioc);
> > > +	kfree_rcu(ioc, rcu);
> > >  }
> > >  
> > >  static void nfs_io_completion_get(struct nfs_io_completion *ioc)
> > > @@ -720,6 +721,8 @@ static void nfs_io_completion_commit(void
> > > *inode)
> > >  	nfs_commit_inode(inode, 0);
> > >  }
> > >  
> > > +static void nfs_commit_end(struct nfs_mds_commit_info *cinfo);
> > > +
> > >  int nfs_writepages(struct address_space *mapping, struct
> > > writeback_control *wbc)
> > >  {
> > >  	struct inode *inode = mapping->host;
> > > @@ -729,9 +732,37 @@ int nfs_writepages(struct address_space
> > > *mapping, struct writeback_control *wbc)
> > >  
> > >  	nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
> > >  
> > > -	ioc = nfs_io_completion_alloc(GFP_KERNEL);
> > > -	if (ioc)
> > > -		nfs_io_completion_init(ioc, nfs_io_completion_commit,
> > > inode);
> > > +	rcu_read_lock();
> > > +	ioc = rcu_dereference(NFS_I(inode)->commit_info.ioc);
> > > +	if (ioc && kref_get_unless_zero(&ioc->refcount)) {
> > > +		rcu_read_unlock();
> > > +		/* We've successfully piggybacked on the attached ioc
> > > */
> > > +	} else {
> > > +		rcu_read_unlock();
> > > +		ioc = nfs_io_completion_alloc(GFP_KERNEL);
> > > +		if (ioc) {
> > > +			struct nfs_io_completion *ioc_prev;
> > > +
> > > +			nfs_io_completion_init(ioc,
> > > nfs_io_completion_commit,
> > > +					       inode);
> > > +			/* Temporarily elevate rpcs_out to ensure a
> > > commit
> > > +			 * completion *will* happen after we attach
> > > this ioc,
> > > +			 * so it *will* get a chance to drain.
> > > +			 */
> > > +			atomic_inc(&NFS_I(inode)-
> > > > commit_info.rpcs_out);
> > > +			nfs_io_completion_get(ioc);
> > > +			ioc_prev = xchg(&NFS_I(inode)->commit_info.ioc,
> > > +				    ioc);
> > > +			/* ioc_prev is normally NULL, but racing
> > > writepages
> > > +			 * calls might result in it being non-NULL.
> > > +			 * In either case it is safe to put it - worst
> > > case
> > > +			 * we get an extra COMMIT.
> > > +			 */
> > > +			nfs_io_completion_put(ioc_prev);
> > > +			/* release temporary ref on rpcs_out */
> > > +			nfs_commit_end(&NFS_I(inode)->commit_info);
> > 
> > So won't this normally trigger the xchg(NULL) in nfs_commmit_end()?
> > For
> > most cases, I'd expect commit_info.rpcs_out to be zero until we
> > start
> > actually sending out the COMMITs.
> 
> On the first writepages call after a period of quiet, that is exactly
> correct.  The ioc that we have just allocated and attached will
> immediately be detached.  It will have a refcount of 1 at this point
> so
> nothing further happens.
> Then WRITEs get queued elevating the refcount.  This will eventually
> drain and a COMMIT will be sent.  So the behaviour for the firstpages
> call is exactly like the current code (I think I mentioned that in
> the
> commit message).
> If the WRITEs and COMMIT complete before the next .writepages, then
> that
> writepages will behave the same way.
> But if the WRITEs or COMMIT are still pending, then rpcs_out will
> still
> be elevated....
> Only - no.  Unattached iocs are meant to be counted in rpcs_out, but
> I
> haven't got that right.  So we can still end up with parallel
> COMMITs.
> I think that might explain some slightly odd numbers in my testing.
> 
> Here is the new version.
> 
> Thanks for the review.
> 
> NeilBrown
> 
> From: NeilBrown <neilb@suse.de>
> Subject: [PATCH] NFS: Minimize COMMIT calls during writeback.
> 
> Since 4.13 (see Fixes: commit) NFS has sent a COMMIT request for
> every
> .writepages() call it received - when the writes submitted for that
> call
> have all completed.
> 
> This works well enough when COMMIT is fast, but if COMMIT is slow
> these
> calls can overlap each other, overload the server, and substantially
> reduce throughput.
> 
> I looked at this due to a report of regression when writing to a
> Ganesha
> NFS server with tracing showing multiple overlapping COMMITs, and
> individual commits typically taking 200ms to 300ms.  This resulted in
> 2
> orders of magnitude slow down when writing 1000x1M from /dev/zero
> with dd.
> This is easily reproduced by adding 'msleep(300)' to nfsd_commit() in
> the
> Linux NFS server.
> 
> This patch changes the details of when COMMIT is sent without
> changing
> the general approach.
> 
> Where previously the writes from a single .writepages() call were
> accounted together in a nfs_io_completion, now the writes from
> multiple
> writepages() calls are accounted together.  The set of writepages
> calls
> that are grouped together are all those from when one COMMIT call
> ends
> to when the next COMMIT call ends.  This automatically reduces the
> rate
> of COMMIT requests when COMMIT itself is slow.
> (If there are no COMMIT calls pending, the first .writepages will get
>  an nfs_io_completion to itself, then subsequenct writepages requests
>  until the first COMMIT completes will go in the next
> nfs_io_completion)
> 
> There are typically at most two nfs_io_completion structures
> allocated
> for writeback to a given inode.
> 
> - If there was been any writepages calls since the last time a COMMIT
>   completed, there will be an nfs_io_completion stored in the inode
> (in
>   nfs_mds_commit_info) which new writepages requests are accounted
> it.
> 
> - If there is no pending COMMIT request, but there are pending
> writeback
>   WRITES, there will be another nfs_io_completion which is not
> attached
>   and which is draining.  When it fully drains a COMMIT will be sent.
>   When that COMMIT completes, the attached nfs_io_completion will be
>   detached and allowed to drain.
> 
> The rpcs_out counter now counts the unattached nfs_io_completion as
> well
> as any pending COMMIT requests.  As an unattached nfs_io_completion
> will
> soon turn into a COMMIT request, this seems reasonable.  It allows us
> to
> clearly detect when there are no longer any COMMITs expected to
> complete, so we know to detach any nfs_io_completion from the inode
> and
> allow it to drain.
> 
> As we use atomic accesses (e.g.  xchg and kref_get_unless_zero()) to
> access the attached nfs_io_completion, we now use kfree_rcu() to free
> it, to ensure it is not accessed after it is freed.
> 
> With 300ms commits, this improves throught of a 1G dd by 2 orders of
> magnitude.  Even without the 300ms delay, this noticeably improves
> throughput to a Linux NFS server is a simple VM.
> 
> Fixes: 919e3bd9a875 ("NFS: Ensure we commit after writeback is
> complete")
> Signed-off-by: NeilBrown <neilb@suse.de>
> ---
>  fs/nfs/inode.c          |  1 +
>  fs/nfs/write.c          | 69 +++++++++++++++++++++++++++++++++++++
> ----
>  include/linux/nfs_xdr.h |  7 +++++
>  3 files changed, 71 insertions(+), 6 deletions(-)
> 
> diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> index 11bf15800ac9..c00b54491949 100644
> --- a/fs/nfs/inode.c
> +++ b/fs/nfs/inode.c
> @@ -2110,6 +2110,7 @@ static void init_once(void *foo)
>  	INIT_LIST_HEAD(&nfsi->commit_info.list);
>  	atomic_long_set(&nfsi->nrequests, 0);
>  	atomic_long_set(&nfsi->commit_info.ncommit, 0);
> +	nfsi->commit_info.ioc = NULL;
>  	atomic_set(&nfsi->commit_info.rpcs_out, 0);
>  	init_rwsem(&nfsi->rmdir_sem);
>  	mutex_init(&nfsi->commit_mutex);
> diff --git a/fs/nfs/write.c b/fs/nfs/write.c
> index c478b772cc49..6dbb055f80bf 100644
> --- a/fs/nfs/write.c
> +++ b/fs/nfs/write.c
> @@ -47,6 +47,7 @@ struct nfs_io_completion {
>  	void (*complete)(void *data);
>  	void *data;
>  	struct kref refcount;
> +	struct rcu_head rcu;
>  };
>  
>  /*
> @@ -134,7 +135,7 @@ static void nfs_io_completion_release(struct kref
> *kref)
>  	struct nfs_io_completion *ioc = container_of(kref,
>  			struct nfs_io_completion, refcount);
>  	ioc->complete(ioc->data);
> -	kfree(ioc);
> +	kfree_rcu(ioc, rcu);
>  }
>  
>  static void nfs_io_completion_get(struct nfs_io_completion *ioc)
> @@ -715,9 +716,15 @@ static int nfs_writepages_callback(struct page
> *page, struct writeback_control *
>  	return ret;
>  }
>  
> +static void nfs_commit_end(struct nfs_mds_commit_info *cinfo);
> +
>  static void nfs_io_completion_commit(void *inode)
>  {
>  	nfs_commit_inode(inode, 0);
> +	/* this came from a detached nfs_io_completion, which is now
> +	 * no longer active, so must decrement rpcs_out.
> +	 */
> +	nfs_commit_end(&NFS_I(inode)->commit_info);
>  }
>  
>  int nfs_writepages(struct address_space *mapping, struct
> writeback_control *wbc)
> @@ -729,9 +736,46 @@ int nfs_writepages(struct address_space
> *mapping, struct writeback_control *wbc)
>  
>  	nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
>  
> -	ioc = nfs_io_completion_alloc(GFP_KERNEL);
> -	if (ioc)
> -		nfs_io_completion_init(ioc, nfs_io_completion_commit,
> inode);
> +	rcu_read_lock();
> +	ioc = rcu_dereference(NFS_I(inode)->commit_info.ioc);
> +	if (ioc && kref_get_unless_zero(&ioc->refcount)) {
> +		rcu_read_unlock();
> +		/* We've successfully piggybacked on the attached ioc
> */
> +	} else {
> +		rcu_read_unlock();
> +		ioc = nfs_io_completion_alloc(GFP_KERNEL);
> +		if (ioc) {
> +			struct nfs_io_completion *ioc_prev;
> +
> +			nfs_io_completion_init(ioc,
> nfs_io_completion_commit,
> +					       inode);
> +			/* This is now a detached ioc so we count it in
> +			 * rpcs_out.  *After* we successfully attach it
> +			 * below (likely, but not certain), we will
> call
> +			 * nfs_commit_end() which might detach it
> immedately
> +			 * if there are no outstanding commit.
> +			 */
> +			atomic_inc(&NFS_I(inode)-
> >commit_info.rpcs_out);

commit_info.rpcs_out is there in order to count the number of
outstanding COMMITs. I'm not seeing why we need to change that
definition, particularly given that there can be processes out there
that are trying to wait for the count to go to zero.

The I/O completion structure already has its own reference counter, and
I can see nothing stopping you from modifying
nfs_io_completion_commit() to do the same things you are trying to do
in nfs_commit_end() here.

However as I said, I have my doubts as to whether or not this is a good
idea. It sounds to me as if you're trying to fix a memory management
layer problem in the NFS layer.

> +			nfs_io_completion_get(ioc);
> +			ioc_prev = xchg(&NFS_I(inode)->commit_info.ioc,
> +				    ioc);
> +			/* ioc_prev is normally NULL, but racing
> writepages
> +			 * calls might result in it being non-NULL.
> +			 * In either case it is safe to put it - worst
> case
> +			 * we get an extra COMMIT.
> +			 */
> +			if (ioc_prev)
> +				/* This ioc_prev is now detached, so
> put it, and
> +				 * keep the ref on rpcs_out
> +				 */
> +				nfs_io_completion_put(ioc_prev);
> +			else
> +				/* Attachment successful, so drop ref
> +				 * on rpcs_out
> +				 */
> +				nfs_commit_end(&NFS_I(inode)-
> >commit_info);
> +		}
> +	}
>  
>  	nfs_pageio_init_write(&pgio, inode, wb_priority(wbc), false,
>  				&nfs_async_write_completion_ops);
> @@ -1677,8 +1721,21 @@ static void nfs_commit_begin(struct
> nfs_mds_commit_info *cinfo)
>  
>  static void nfs_commit_end(struct nfs_mds_commit_info *cinfo)
>  {
> -	if (atomic_dec_and_test(&cinfo->rpcs_out))
> -		wake_up_var(&cinfo->rpcs_out);
> +	if (atomic_dec_and_test(&cinfo->rpcs_out)) {
> +		/* When a commit finishes, we must release any attached
> +		 * nfs_io_completion so that it can drain and then
> request
> +		 * another commit.
> +		 */
> +		struct nfs_io_completion *ioc;
> +
> +		ioc = xchg(&cinfo->ioc, NULL);
> +		if (ioc) {
> +			/* Count the detached completion */
> +			atomic_inc(&cinfo->rpcs_out);
> +			nfs_io_completion_put(ioc);
> +		} else
> +			wake_up_var(&cinfo->rpcs_out);
> +	}
>  }
>  
>  void nfs_commitdata_release(struct nfs_commit_data *data)
> diff --git a/include/linux/nfs_xdr.h b/include/linux/nfs_xdr.h
> index 94c77ed55ce1..89db1e9d461d 100644
> --- a/include/linux/nfs_xdr.h
> +++ b/include/linux/nfs_xdr.h
> @@ -1557,9 +1557,16 @@ struct nfs_pgio_header {
>  };
>  
>  struct nfs_mds_commit_info {
> +	/* rpcs_out counts pending COMMIT rpcs plus pendng
> nfs_io_completions
> +	 * which are *not* attached at 'ioc' below.  Such
> nfs_io_compleions
> +	 * (normally at most one) will drain as writes complete and
> then trigger
> +	 * a COMMIT, so they can be considered as pending COMMITs which
> haven't
> +	 * been sent yet
> +	 */
>  	atomic_t rpcs_out;
>  	atomic_long_t		ncommit;
>  	struct list_head	list;
> +	struct nfs_io_completion *ioc;
>  };
>  
>  struct nfs_commit_info;
-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com