Re: [PATCH v2 2/3] NFSv4 introduce max_connect mount options

[Trond Myklebust <trondmy@hammerspace.com>]

On Thu, 2021-06-10 at 12:14 -0400, Olga Kornievskaia wrote:
> On Thu, Jun 10, 2021 at 10:56 AM Trond Myklebust
> <trondmy@hammerspace.com> wrote:
> > 
> > On Thu, 2021-06-10 at 10:31 -0400, Olga Kornievskaia wrote:
> > > On Thu, Jun 10, 2021 at 10:13 AM Trond Myklebust
> > > <trondmy@hammerspace.com> wrote:
> > > > 
> > > > On Thu, 2021-06-10 at 13:56 +0000, Chuck Lever III wrote:
> > > > > 
> > > > > 
> > > > > > On Jun 10, 2021, at 9:34 AM, Trond Myklebust <
> > > > > > trondmy@hammerspace.com> wrote:
> > > > > > 
> > > > > > On Thu, 2021-06-10 at 13:30 +0000, Chuck Lever III wrote:
> > > > > > > 
> > > > > > > 
> > > > > > > > On Jun 9, 2021, at 5:53 PM, Olga Kornievskaia <
> > > > > > > > olga.kornievskaia@gmail.com> wrote:
> > > > > > > > 
> > > > > > > > From: Olga Kornievskaia <kolga@netapp.com>
> > > > > > > > 
> > > > > > > > This option will control up to how many xprts can the
> > > > > > > > client
> > > > > > > > establish to the server. This patch parses the value
> > > > > > > > and
> > > > > > > > sets
> > > > > > > > up structures that keep track of max_connect.
> > > > > > > > 
> > > > > > > > Signed-off-by: Olga Kornievskaia <kolga@netapp.com>
> > > > > > > > ---
> > > > > > > > fs/nfs/client.c           |  1 +
> > > > > > > > fs/nfs/fs_context.c       |  8 ++++++++
> > > > > > > > fs/nfs/internal.h         |  2 ++
> > > > > > > > fs/nfs/nfs4client.c       | 12 ++++++++++--
> > > > > > > > fs/nfs/super.c            |  2 ++
> > > > > > > > include/linux/nfs_fs_sb.h |  1 +
> > > > > > > > 6 files changed, 24 insertions(+), 2 deletions(-)
> > > > > > > > 
> > > > > > > > diff --git a/fs/nfs/client.c b/fs/nfs/client.c
> > > > > > > > index 330f65727c45..486dec59972b 100644
> > > > > > > > --- a/fs/nfs/client.c
> > > > > > > > +++ b/fs/nfs/client.c
> > > > > > > > @@ -179,6 +179,7 @@ struct nfs_client
> > > > > > > > *nfs_alloc_client(const
> > > > > > > > struct nfs_client_initdata *cl_init)
> > > > > > > > 
> > > > > > > >         clp->cl_proto = cl_init->proto;
> > > > > > > >         clp->cl_nconnect = cl_init->nconnect;
> > > > > > > > +       clp->cl_max_connect = cl_init->max_connect ?
> > > > > > > > cl_init-
> > > > > > > > > max_connect : 1;
> > > > > > > 
> > > > > > > So, 1 is the default setting, meaning the "add another
> > > > > > > transport"
> > > > > > > facility is disabled by default. Would it be less
> > > > > > > surprising
> > > > > > > for
> > > > > > > an admin to allow some extra connections by default?
> > > > > > > 
> > > > > > > 
> > > > > > > >         clp->cl_net = get_net(cl_init->net);
> > > > > > > > 
> > > > > > > >         clp->cl_principal = "*";
> > > > > > > > diff --git a/fs/nfs/fs_context.c b/fs/nfs/fs_context.c
> > > > > > > > index d95c9a39bc70..cfbff7098f8e 100644
> > > > > > > > --- a/fs/nfs/fs_context.c
> > > > > > > > +++ b/fs/nfs/fs_context.c
> > > > > > > > @@ -29,6 +29,7 @@
> > > > > > > > #endif
> > > > > > > > 
> > > > > > > > #define NFS_MAX_CONNECTIONS 16
> > > > > > > > +#define NFS_MAX_TRANSPORTS 128
> > > > > > > 
> > > > > > > This maximum seems excessive... again, there are
> > > > > > > diminishing
> > > > > > > returns to adding more connections to the same server.
> > > > > > > what's
> > > > > > > wrong with re-using NFS_MAX_CONNECTIONS for the maximum?
> > > > > > > 
> > > > > > > As always, I'm a little queasy about adding yet another
> > > > > > > mount
> > > > > > > option. Are there real use cases where a whole-client
> > > > > > > setting
> > > > > > > (like a sysfs attribute) would be inadequate? Is there a
> > > > > > > way
> > > > > > > the client could figure out a reasonable maximum without
> > > > > > > a
> > > > > > > human intervention, say, by counting the number of NICs
> > > > > > > on
> > > > > > > the system?
> > > > > > 
> > > > > > Oh, hell no! We're not tying anything to the number of
> > > > > > NICs...
> > > > > 
> > > > > That's a bit of an over-reaction. :-) A little more
> > > > > explanation
> > > > > would be welcome. I mean, don't you expect someone to ask
> > > > > "How
> > > > > do I pick a good value?" and someone might reasonably answer
> > > > > "Well, start with the number of NICs on your client times 3"
> > > > > or
> > > > > something like that.
> > > > > 
> > > > > IMO we're about to add another admin setting without
> > > > > understanding
> > > > > how it will be used, how to select a good maximum value, or
> > > > > even
> > > > > whether this maximum needs to be adjustable. In a previous e-
> > > > > mail
> > > > > Olga has already demonstrated that it will be difficult to
> > > > > explain
> > > > > how to use this setting with nconnect=.
> > > > > 
> > > > > Thus I would favor a (moderate) soldered-in maximum to start
> > > > > with,
> > > > > and then as real world use cases arise, consider adding a
> > > > > tuning
> > > > > mechanism based on actual requirements.
> > > > 
> > > > It's not an overreaction. It's insane to think that counting
> > > > NICs
> > > > gives
> > > > you any notion whatsoever about the network topology and
> > > > connectivity
> > > > between the client and server. It doesn't even tell you how
> > > > many of
> > > > those NICs might potentially be available to your application.
> > > > 
> > > > We're not doing any automation based on that kind of layering
> > > > violation.
> > > 
> > > I'm not suggesting to programmatically determine the number of
> > > NIC to
> > > determine the value of max_connect.
> > > > 
> > 
> > No, but that's what Chuck appeared to be suggesting in order to
> > avoid
> > the need for the mount option.
> > 
> > To me, the main reason for the mount option is to allow the user to
> > limit the number of new IP addresses being added so that if the DNS
> > server is configured to hand out lots of different addresses for
> > the
> > same servername, the user can basically say 'no, I just want to use
> > the
> > one IP address that I'm already connected to' (i.e. max_connect=1).
> > I
> > can imagine that some clustered setups might need that ability in
> > order
> > to work efficiently.
> > 
> > I'm fine with the idea of nconnect setting the number of
> > connections
> > per IP address, but that would need some plumbing in
> > rpc_clnt_test_and_add_xprt() to allow us to add up to 'nconnect'
> > copies
> > of a given transport.
> > Presumably rpc_xprt_switch_has_addr() would need to return a count
> > of
> > the number of copies of the transport that are already present so
> > that
> > we can decide whether or not we should add a new one.
> 
> I think the last paragraph is what I'm asking for. But I would like
> to
> again confirm if you still mean "max_connect" to be the total number
> of connections since you say we could/will allow for nconnect number
> of connections per IP address. Would max_connect need to be a
> multiple
> of nconnect (max_connect = X *nconnect)?

No. Your suggestion to make the two independent is growing on me,
however in that case we do want to ensure that if nconnect=X, then we
always add X transports when we add a new IP address.

> 
> Actually when I said supporting (or rather allowing for) nconnect *
> max_connect transport, is that correct? Given how the code works now
> this is going to be nconnect + max_connect (only if 1st mount had
> nconnect option). We can't "add" nconnect connections to the new
> mounts (but with my patch we can add a single trunk connection). By
> that I mean: say the first was "mount IP1:/vol1 /mnt1" (1 connection
> to IP2). Now the client is doing "mount IP2:/vol2 /mnt2". IP1 and IP2
> are trunkable addresses of the same server so we add a trunk. We
> currently don't allow for doing "mount -o nconnec=2 IP2:vol2 /mnt2"
> and then also add "nconnect" connections to IP2 along with a trunk.
> In
> the 2nd example, we'd have 1 connections to IP1, then 2 connections
> to
> IP2. Can we allow for that (with needed code change)?  If not, then
> we
> really need to commit to only support nconnect (16) connections +
> some
> number of trunkable connections.


I think we want to have nconnect be server-global. i.e. nconnect
entries of each IP address.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com