On Wed, 2021-06-30 at 09:26 +0800, Xiubo Li wrote:
> On 6/29/21 9:25 PM, Jeff Layton wrote:
> > On Tue, 2021-06-29 at 12:42 +0800, xiubli@xxxxxxxxxx wrote:
> > > From: Xiubo Li <xiubli@xxxxxxxxxx>
> > >
> > > For the client requests who will have unsafe and safe replies from
> > > MDS daemons, in the MDS side the MDS daemons won't flush the mdlog
> > > (journal log) immediatelly, because they think it's unnecessary.
> > > That's true for most cases but not all, likes the fsync request.
> > > The fsync will wait until all the unsafe replied requests to be
> > > safely replied.
> > >
> > > Normally if there have multiple threads or clients are running, the
> > > whole mdlog in MDS daemons could be flushed in time if any request
> > > will trigger the mdlog submit thread. So usually we won't experience
> > > the normal operations will stuck for a long time. But in case there
> > > has only one client with only thread is running, the stuck phenomenon
> > > maybe obvious and the worst case it must wait at most 5 seconds to
> > > wait the mdlog to be flushed by the MDS's tick thread periodically.
> > >
> > > This patch will trigger to flush the mdlog in all the MDSes manually
> > > just before waiting the unsafe requests to finish.
> > >
> > > Signed-off-by: Xiubo Li <xiubli@xxxxxxxxxx>
> > > ---
> > > fs/ceph/caps.c | 9 +++++++++
> > > 1 file changed, 9 insertions(+)
> > >
> > > diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
> > > index c6a3352a4d52..6e80e4649c7a 100644
> > > --- a/fs/ceph/caps.c
> > > +++ b/fs/ceph/caps.c
> > > @@ -2286,6 +2286,7 @@ static int caps_are_flushed(struct inode *inode, u64 flush_tid)
> > > */
> > > static int unsafe_request_wait(struct inode *inode)
> > > {
> > > + struct ceph_mds_client *mdsc = ceph_sb_to_client(inode->i_sb)->mdsc;
> > > struct ceph_inode_info *ci = ceph_inode(inode);
> > > struct ceph_mds_request *req1 = NULL, *req2 = NULL;
> > > int ret, err = 0;
> > > @@ -2305,6 +2306,14 @@ static int unsafe_request_wait(struct inode *inode)
> > > }
> > > spin_unlock(&ci->i_unsafe_lock);
> > >
> > > + /*
> > > + * Trigger to flush the journal logs in all the MDSes manually,
> > > + * or in the worst case we must wait at most 5 seconds to wait
> > > + * the journal logs to be flushed by the MDSes periodically.
> > > + */
> > > + if (req1 || req2)
> > > + flush_mdlog(mdsc);
> > > +
> > So this is called on fsync(). Do we really need to flush all of the mds
> > logs on every fsync? That sounds like it might have some performance
> > impact. Would it be possible to just flush the mdslog on the MDS that's
> > authoritative for this inode?
>
> I hit one case before, the mds.0 is the auth mds, but the client just
> sent the request to mds.2, then when the mds.2 tried to gather the
> rdlocks then it was stuck for waiting for the mds.0 to flush the mdlog.
> I think it also will happen that if the mds.0 could also be stuck just
> like this even its the auth MDS.
>
It sounds like mds.0 should flush its own mdlog in this situation once
mds.2 started requesting locks that mds.0 was holding. Shouldn't it?
> Normally the mdlog submit thread will be triggered per MDS's tick,
> that's 5 seconds. But this is not always true mostly because any other
> client request could trigger the mdlog submit thread to run at any time.
> Since the fsync is not running all the time, so IMO the performance
> impact should be okay.
>
>
I'm not sure I'm convinced.
Consider a situation where we have a large(ish) ceph cluster with
several MDSs. One client is writing to a file that is on mds.0 and there
is little other activity there. Several other clients are doing heavy
I/O on other inodes (of which mds.1 is auth).
The first client then calls fsync, and now the other clients stall for a
bit while mds.1 unnecessarily flushes its mdlog. I think we need to take
care to only flush the mdlog for mds's that we care about here.
> >
> > > dout("unsafe_request_wait %p wait on tid %llu %llu\n",
> > > inode, req1 ? req1->r_tid : 0ULL, req2 ? req2->r_tid : 0ULL);
> > > if (req1) {
>
--
Jeff Layton <jlayton@xxxxxxxxxx>
