[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Cluster-devel] gfs uevent and sysfs changes



Hi,

On Thu, 2008-12-04 at 16:38 -0600, David Teigland wrote:
> On Thu, Dec 04, 2008 at 04:59:23PM -0500, david m. richter wrote:
> > ah, so just to make sure i'm with you here: (1) gfs_controld is
> > generating this "id"-which-is-the-mountgroup-id, and (2) gfs_kernel
> > will no longer receive this in the hostdata string, so (3) i can just
> > rip out my in-kernel hostdata-parsing gunk and instead send in the
> > mountgroup id on my own (i have my own up/downcall channel)?  if i've
> > got it right, then everything's a cinch and i'll shut up :)
> 
> Yep.  Generally, the best way to uniquely identify and refer to a gfs
> filesystem is using the fsname string (specified during mkfs with -t and
> saved in the superblock).  But, sometimes it's just a lot easier have a
> numerical identifier instead.  I expect this is why you're using the id,
> and it's why we were using it for communicating about plocks.
> 
> In cluster1 and cluster2 the cluster infrastructure dynamically selected a
> unique id when needed, and it never worked great.  In cluster3 the id is
> just a crc of the fsname string.
> 
> Now that I think about this a bit more, there may be a reason to keep the
> id in the string.  There was some interest on linux-kernel about better
> using the statfs fsid field, and this id is what gfs should be putting
> there.
> 
In that case gfs2 should be able to generate the id itself from the
fsname and it still doesn't need it passed in, even if it continues to
expose the id in sysfs.

Perhaps better still, it should be possible for David to generate the id
directly if he really needs it from the fsname.

Since we also have a UUID now, for recently created filesystems, it
might be worth exposing that via sysfs and/or uevents too.

> > say, one tangential question (i won't be offended if you skip it -
> > heh): is there a particular reason that you folks went with the uevent
> > mechanism for doing upcalls?  i'm just curious, given the
> > seeming-complexity and possible overhead of using the whole layered
> > netlink apparatus vs. something like Trond Myklebust's rpc_pipefs
> > (don't let the "rpc" fool you; it's a barebones, dead-simple pipe).
> > -- and no, i'm not selling anything :)  my boss was asking for a list
> > of differences between rpc_pipefs and uevents and the best i could
> > come up with is the former's bidirectional.  Trond mentioned the
> > netlink overhead and i wondered if that was actually a significant
> > factor or just lost in the noise in most cases.
> 
> The uevents looked pretty simple when I was initially designing how the
> kernel/user interactions would work, and they fit well with sysfs files
> which I was using too.  I don't think the overhead of using uevents is too
> bad.  Sysfs files and uevents definately don't work great if you need any
> kind of sophisticated bi-directional interface.
> 
> Dave
> 
I think uevents are a reasonable choice as they are easy enough to parse
that it could be done by scripts, etc and easy to extend as well. We do
intend to use netlink in the future (bz #337691) for quota messages, but
in that case we would be using an existing standard for sending those
messages.

Netlink can be extended fairly easily, but you do have to be careful
when designing the message format. I've not come across rpc_pipefs
before, so I can't comment on that yet. I don't think we need to worry
about overhead on sending the messages (if you have so much recovery
message traffic that its a problem, you probably have bigger things to
worry about!), and I don't see that netlink should have any more
overhead than any other method of sending messages.

Steve.



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]