RFC: adding 'proxy' nodes to provider ports (with patch)

[Pawel Jakub Dawidek]

On Thu, Mar 19, 2009 at 09:19:36AM +0100, Luigi Rizzo wrote:
> Hi,
> 
> Fabio Checconi and I have been thinking on how to implement "proxy"
> geom nodes, i.e. nodes that have exactly 1 provider and 1 consumer
> port, do not do any data transformation, and can be transparently
> inserted or removed on top of a provider port while the tree is
> actively moving data.
> 
> Our immediate need was to add/remove a scheduler while a disk is
> mounted, but there are possibly other uses e.g. if one wants to
> "sniff" the traffic through a disk, or do other ops that are
> transparent for the data stream.
> 
> We would like opinion on the following solution, which seems
> extremely simple in terms of implementation.
> 
> The idea is to intercept requests coming on a provider port, pp, and
> redirect them to a geom node acting as a proxy if the port
> is configured in this way:
> 
>      +=====...===...======+
>      H                    H
>      H                    H
>      H                    H
>      +=====...====== cp ==+
>                      |          +---------------+
>                      V          |               V
>      +=====.....==== pp ==+     |    +======= proxy_pp ==+
>      H           'ad0s1'  H     |    H                   H
>      H                ------->--+    H                   H
>      H        gp      -------<--+    H    proxy_node     H
>      H                    H     |    H                   H
>      +=======....===...===+     |    +======= proxy_cp ==+
>                                 |               V
>                                 +---------------+
> 
> Normally the proxy does not exist, and the geom tree works as it does now.
> 
> When we create a 'proxy' node, with something like
> 
> 	geom my_proxy_class proxy ad0s1
> 
> we do something very similar to a 'create', but:
> 
> - the proxy node is marked specially in gp->flags, so the core will
>   not generate a g_new_provider_event when the provider port is created
>   (this means there is no taste() call and nobody should be able
>   to attach to the port).
> 
> - the provider port we attach to is linked, with two pointers,
>   to the provider and consumer ports of the proxy_node.
> 
> In this situation, g_io_request() finds that port pp has a proxy attached
> to it, and immediately redirects the requests to the proxy, which
> does everything a geom node does (cloning requests, etc).
>    When the proxy wants to pass the request down, it sends it again to pp,
> but now there is no redirection because the source can be identified
> as the proxy.  The pointers in the bio insure a correct flow of the
> requests on the reverse path.

The one advantage I see for this over using regular GEOM rules is that
new consumers go through proxy automatically. When I was working on
similar functionality I more wanted to do something like this:

consumer1   consumer2
    \          /
     \        /
      provider

Insert the proxy in the middle of any provider-consumer pair:

consumer1      consumer2
    |             |
proxy_provider    |
    |            /
proxy_consumer  /
      \        /
       provider

This can be done (almost I think) atomically:

/* First attach to the destination provider. */
g_attach(proxy_consumer, provider);
/* Then switch original consumer to use proxy_provider (should be almost atomic). */
consumer1->provider = proxy_provider;
/* handle access counts */

In-flight I/O requests know how to go back, because they have source and
destination stored in bio_from and bio_to fields, so no races here.

> Disconnecting a proxy is almost trivial: apart from handling possible
> races on the data path, we just need to clear pp->proxy_pp and pp->proxy_cp.
> After that, we can just send the regular destroy events to the proxy
> node, who will have to take care of flushing any pending bio's (e.g.
> see our geom_sched node that already does this).
> 
> Overall the change is very small (see attached patch):
> a couple of lines in g_io_request, two extra fields in the g_provider,
> and the addition of a flag to gp->flags to control the generation
> of g_new_provider_event.
> There is basically no overhead on regular operation, and only
> a couple of extra pointers in struct g_provider (we use a spare
> bit in gp->flags to mark G_GEOM_PROXY nodes).
> 
> The only things missing in the patch should be:
> 
> - a check to avoid races on creation&destruction of a proxy.
>   I am not so sure on how to achieve this, but creation and destruction
>   are rare and can normally wait, so we could just piggyback the
>   small critical section (manipulating pp->proxy_cp and pp->proxy_cp)
>   into some other piece of code that is guaranteed to be race-free.
> 
> - a check to prevent attaching to a provider port of a proxy
>   (not a problem, i believe);
> 
> - a check to prevent attaching a proxy to a provider port that already
>   has one. Of course you can attach a proxy to another proxy, and
>   if you want to change the order it is as simple as removing the
>   existing proxy and reattaching it after the new one.

Could you provide link for the patch, as it was removed from your e-mail?

-- 
Pawel Jakub Dawidek                       http://www.wheel.pl
pjd at FreeBSD.org                           http://www.FreeBSD.org
FreeBSD committer                         Am I Evil? Yes, I Am!
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 187 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-arch/attachments/20090319/8e89dfc7/attachment.pgp