On Thu 13-06-24 16:04:12, Jeff Layton wrote:
> On Thu, 2024-06-13 at 10:28 +0200, Jan Kara wrote:
> > Commit 6df25e58532b ("nfs: remove reliance on bdi congestion")
> > introduced NFS-private solution for limiting number of writes
> > outstanding against a particular server. Unlike previous bdi congestion
> > this algorithm actually works and limits number of outstanding writeback
> > pages to nfs_congestion_kb which scales with amount of client's memory
> > and is capped at 256 MB. As a result some workloads such as random
> > buffered writes over NFS got slower (from ~170 MB/s to ~126 MB/s). The
> > fio command to reproduce is:
> >
>
> That 256M cap was set back in 2007. I wonder if we ought to reconsider
> that. It might be worth experimenting these days with a higher cap on
> larger machines. Might that improve throughput even more?
In theory it may help but in my setup I didn't observe it. So I've decided
to leave the limit alone for now. But I certainly would not object if
someone decided to increase it as 256MB seems a bit low buffer space for
writeback given contemporary IO speeds, I just didn't have numbers to back
that decision.
> > fio --direct=0 --ioengine=sync --thread --invalidate=1 --group_reporting=1
> > --runtime=300 --fallocate=posix --ramp_time=10 --new_group --rw=randwrite
> > --size=64256m --numjobs=4 --bs=4k --fsync_on_close=1 --end_fsync=1
> >
> > This happens because the client sends ~256 MB worth of dirty pages to
> > the server and any further background writeback request is ignored until
> > the number of writeback pages gets below the threshold of 192 MB. By the
> > time this happens and clients decides to trigger another round of
> > writeback, the server often has no pages to write and the disk is idle.
> >
> > To fix this problem and make the client react faster to eased congestion
> > of the server by blocking waiting for congestion to resolve instead of
> > aborting writeback. This improves the random 4k buffered write
> > throughput to 184 MB/s.
> >
> > Signed-off-by: Jan Kara <jack@xxxxxxx>
> > ---
> > fs/nfs/client.c | 1 +
> > fs/nfs/write.c | 12 ++++++++----
> > include/linux/nfs_fs_sb.h | 1 +
> > 3 files changed, 10 insertions(+), 4 deletions(-)
> >
> > diff --git a/fs/nfs/client.c b/fs/nfs/client.c
> > index 3b252dceebf5..8286edd6062d 100644
> > --- a/fs/nfs/client.c
> > +++ b/fs/nfs/client.c
> > @@ -994,6 +994,7 @@ struct nfs_server *nfs_alloc_server(void)
> >
> > server->change_attr_type = NFS4_CHANGE_TYPE_IS_UNDEFINED;
> >
> > + init_waitqueue_head(&server->write_congestion_wait);
> > atomic_long_set(&server->writeback, 0);
> >
> > ida_init(&server->openowner_id);
> > diff --git a/fs/nfs/write.c b/fs/nfs/write.c
> > index c6255d7edd3c..21a5f1e90859 100644
> > --- a/fs/nfs/write.c
> > +++ b/fs/nfs/write.c
> > @@ -423,8 +423,10 @@ static void nfs_folio_end_writeback(struct folio *folio)
> >
> > folio_end_writeback(folio);
> > if (atomic_long_dec_return(&nfss->writeback) <
> > - NFS_CONGESTION_OFF_THRESH)
> > + NFS_CONGESTION_OFF_THRESH) {
> > nfss->write_congested = 0;
> > + wake_up_all(&nfss->write_congestion_wait);
> > + }
> > }
> >
> > static void nfs_page_end_writeback(struct nfs_page *req)
> > @@ -698,12 +700,14 @@ int nfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
> > struct nfs_pageio_descriptor pgio;
> > struct nfs_io_completion *ioc = NULL;
> > unsigned int mntflags = NFS_SERVER(inode)->flags;
> > + struct nfs_server *nfss = NFS_SERVER(inode);
> > int priority = 0;
> > int err;
> >
> > - if (wbc->sync_mode == WB_SYNC_NONE &&
> > - NFS_SERVER(inode)->write_congested)
> > - return 0;
> > + /* Wait with writeback until write congestion eases */
> > + if (wbc->sync_mode == WB_SYNC_NONE && nfss->write_congested)
> > + wait_event(nfss->write_congestion_wait,
> > + nfss->write_congested == 0);
> >
>
> It seems odd to block here, but if that helps throughput then so be it.
Well, I've kept the place where writeback throttling is decided. Another
logical possibility would be to handle the blocking in
nfs_writepages_callback() more like it happens for standard block device
filesystems which block in submit_bio(). Just there it would potentially
require more work as AFAIU we can have some unfinished RPC requests which
we may need to flush before blocking or something similar (similarly to how
we unplug IO on schedule). So I wasn't really comfortable in moving the
blocking there.
> > nfs_inc_stats(inode, NFSIOS_VFSWRITEPAGES);
> >
> > diff --git a/include/linux/nfs_fs_sb.h b/include/linux/nfs_fs_sb.h
> > index 92de074e63b9..38a128796a76 100644
> > --- a/include/linux/nfs_fs_sb.h
> > +++ b/include/linux/nfs_fs_sb.h
> > @@ -140,6 +140,7 @@ struct nfs_server {
> > struct rpc_clnt * client_acl; /* ACL RPC client handle */
> > struct nlm_host *nlm_host; /* NLM client handle */
> > struct nfs_iostats __percpu *io_stats; /* I/O statistics */
> > + wait_queue_head_t write_congestion_wait; /* wait until write congestion eases */
> > atomic_long_t writeback; /* number of writeback pages */
> > unsigned int write_congested;/* flag set when writeback gets too high */
> > unsigned int flags; /* various flags */
>
> Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>
Thanks!
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
