[lustre-devel] [PATCH 00/24] lustre - more cleanups including module reduction.

[NeilBrown <neilb@suse.com>]

On Wed, Jun 27 2018, Andreas Dilger wrote:

> On Jun 27, 2018, at 05:08, NeilBrown <neilb@suse.com> wrote:
>> 
>> On Tue, Jun 26 2018, Patrick Farrell wrote:
>> 
>>> Ah, sorry, lost your mail in the shuffle, Neil.
>>> 
>>> The name, mostly.  ptlrpc is one subdirectory and one subsystem, so I don?t want to use it for a module that explicitly includes all of several.  I?m not sure of a better name immediately - is this module intended to be shared between client and server?
>>> 
>> 
>> Thanks for clarifying.
>> Note that I'm still feeling my way around here so "intended" can at most
>> be a soft intention, open to change.  But yes, I expect this module
>> would ultimately be shared with the server.
>> 
>> I think things should only be in different modules if it might sometimes
>> make sense to use one without the other.
>> 
>> So everything that is always used for a lustre client mount, and is only
>> used for a lustre client mount, should be in a module called "lustre".
>> And everything always and only used for a lustre server should be in a
>> module called (something like) "lustre-server".
>> 
>> It appears to me that ptlrpc, ldlm, and obdclass are always and only
>> used together, but can be used for client or server.  So I think they
>> should be together in a module.
>> 
>> I'm tempted to add lnet to that list.  There has been mention that Cray
>> has some other functionality that uses lnet - if that doesn't use
>> ptlrpc, then maybe that is a case for keeping lnet separate.  However I
>> would rather see it as a potential case to separate lnet from the others
>> at a later date if that functionality from cray ever goes into mainline.
>> 
>> The klnds modules can presumably be used independently(?) so they can
>> sensibly remain as separate modules, though I'm beginning to wonder if
>> lnet can actually function without socklnd, so I wonder if that should
>> be permanently part of the lnet module (with NFS, the xprt_sock code is
>> a permanent part of "sunrpc", while the xprt_rdma (aka rpcrdma) code
>> can be a separate module).
>
> Definitely LNet can run without ksocklnd, and it does so on clients that
> only have IB connections.  Lots of HPC systems have client nodes that do
> not have Ethernet interfaces, since it is a cabling/networking nightmare
> to add a few thousand additional Ethernet cables/ports in addition to the
> IB cables/ports, and it is always possible to run IPoIB for TCP traffic.

Thanks for the clarification.
This came up because I'd been looking at lnet/lib-socket.c and wondering
why that code wasn't in socklnd.  If found that it was also used by
lnet/acceptor which seems to always run a thread that listens on a TCP
port (I think - I skim lots of code while trying to figure out the
big-picture structure).
If there is always a socket acceptor, I thought you might always need a
socket back-end (lnd).

Should the 'acceptor' thing really be part of socklnd?

>
>> In the interests of a concrete strawman, what objections would I get if
>> I suggested that ptlrpc, ldlm, obdclass, lnet, and socklnd were all included in
>> the one module named "lnet" ??
>
> This would break all of the module parameters for ptlrpc.  There are only
> a few obdclass module parameters that are rarely used, so I'm less worried
> about those ones.
>
> Given that lnet is a separate thing, I'd prefer to keep "ptlrpc" and "lnet"
> as separate modules, then "lustre" for the remainder of the client
> code.

This does seem to be the idea that will be most widely accepted.
The module parameters seem to be the main practical issue with merging
modules.  I wouldn't want that to force us to keep things separate that
should be together, but nor do I want to introduce unnecessary
breakage.

>
> The "ptlrpc" module should also include the "fld", "fid", "quota", "mgc"
> modules, since these are shared with the server as well.
>
> Server modules:
> [adilger at mookie ~]$ lsmod | grep ^obdclass
> obdclass             1731964  83 mdd,lod,mdt,osp,ofd,lfsck,ost,mgs,mgc,osd_ldiskfs,fid,fld,lquota,ptlrpc
>
> Client modules:
> [adilger at twoshoes ~]$ lsmod | grep ^obdclass
> obdclass             1568040  86 osc,mgc,lustre,lov,mdc,fid,lmv,fld,ptlrpc

Thanks for this - a useful perspective.
I don't see quota (or lquota) on the client side though ??

>
>
> James, if we are going to move in the direction of having a separate
> "lustre-server" module, it would make sense to land a patch to master
> now that adds both a module alias, as well as a filesystem type alias
> for lustre-server, so that we can start moving systems over to using
> "lustre-server" as the filesystem type in /etc/fstab or HA mounting
> scripts, and it will auto-load the module at the same time.  This might
> also allow the Lustre mount code to be simplified, since I think it has
> some messy code to have the same code, but distinguish between client
> and server mounts.

This sounds like an excellent idea!
I have a patch in my lustre-testing tree (probably will post next week)
which moves the client-side mounting from obdclass to llite.  It is a
nice simplification.

Thanks,
NeilBrown

>
> Cheers, Andreas
> ---
> Andreas Dilger
> Principal Lustre Architect
> Whamcloud
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 832 bytes
Desc: not available
URL: <http://lists.lustre.org/pipermail/lustre-devel-lustre.org/attachments/20180628/5d3ab435/attachment.sig>