Switch vfs.nfsd.issue_delegations to TUNABLE ?

[Rick Macklem]

I think the attached patch should fix your performance problem.
It simply checks for nfsrv_delegatecnt != 0 before doing all the
locking stuff in nfsrv_checkgetattr().

If this fixes your performance problem (with delegations disabled),
I'll probably add a separate counter for write delegations and
use atomics to increment/decrement it so that it is SMP safe without
acquiring any lock.

If you can test this, please let me know how it goes? rick

________________________________________
From: Rick Macklem <rmacklem at uoguelph.ca>
Sent: Tuesday, November 28, 2017 2:09:51 PM
To: Emmanuel Vadot
Cc: Konstantin Belousov; FreeBSD Current; freebsd-fs at freebsd.org; Rick Macklem
Subject: Re: Switch vfs.nfsd.issue_delegations to TUNABLE ?

Emmanuel Vadot wrote:
>I wrote:
>> Since it defaults to "disabled", I don't see why a tunable would be necessary?
>> (Just do nothing and delegations don't happen. If you want the server
>>  to issue delegations, then use the sysctl to turn them on. If you want to turn
>>  them off again at the server, reboot the server without setting the sysctl.)
>
>If you need to reboot to make things working again without delegation
>this shouldn't be a sysctl.
Turning them off without rebooting doesn't break anything.
I just wrote it that way since you seemed to want to use a tunable, which
implied rebooting was your preference.
> > > >  The reason behind it is recent problem at work on some on our filer
> > > > related to NFS.
> > > >  We use NFSv4 without delegation as we never been able to have good
> > > > performance with FreeBSD server and Linux client (we need to do test
> > > > again but that's for later).
> Delegations are almost never useful, especially with Linux clients.
Emmanuel Vadot wrote:
>Can you elaborate ? Reading what delegation are I see that this is
>exactly what I'm looking for to have better performance with NFS for my
>workload (I only have one client per mount point).
Delegations allow the client to do Opens locally without contacting the
server. Unless, the delegation is a write delegation, this only applied
to read-only delegations. Also, since most implementors couldn`t agree
on how to check permissions via the delegation, most client implementations
still do an Access check at open, which is almost as much overhead as the
Open RPC. (For example, Solaris servers always specified an empty ACE in the
delegation, forcing the client to do an Access. I have no idea what the
current Linux serveréclient does. If you capture packets when a Linux
client is mounted with delegations enabled, you could look for RPCs like Access when
are Opened multiple times. If you see them, then delegations aren`t saving RPCs.
Also, they are `per file`, so are only useful if the client is Opening the
same file multiple times.
Further, if another client Opens the same file and the first client got a Write
delegation, then the write delegation must be recalled, which is a lot of
overhead and one of the few cases where the FreeBSD server must exclusively
lock the state lists, forcing all other RPCs related to Open, Close to wait.

They sounded good in theory and might have worked well if the implementors
had agreed to do them, but that didnèt happen. (Companies pay for servers, but the
clients donèt get funded so delegation support in the clients are lacking. I tried
to make them useful in the FreeBSD client, but Ièm not sure I succeeded.)

> [stuff snipped]
If I understood your original post, you have a performance problem caused
by lock contention, where the server grabs the state lock to check for delegations
for every Getattr from a client.

As below, I think the fix is to add code to check for no delegations issued that
does not require acquisition of the state lock.

Btw, large numbers of Getattrs will not perform well with delegations.
(Again, the client should be able to do Getattr operations locally in the
 client when it has a delegation for the file, but if the client is not doing that...)

I wrote:
>
> Having a per-mount version of this would be overkill, I think. It would have to
> disable callbacks on the mount point, since there is no way for a client to say
> "don't give me delegations" except disabling callbacks, which the server
> requires for delegation issue.
> [stuff snipped]
> The case where there has never been delegations issued will result in an
> empty delegation queue. Maybe this case can be handled without acquiring
> the "global NFSv4 state lock", which is what sounds like is causing the
> performance issue. Maybe a global counter of how many delegations are
> issued that is handled by atomic ops.
> --> If it is 0, nfsrv_checkgetattr() can just return without acquiring the lock.
>
> I'll take a look at this, rick
>
rick