Re: [PATCH] EVMS core (3/9) discover.c

["Mark Peloquin" <peloquin@us.ibm.com>]

On 10/10/2002 at 03:48 PM, Andi Kleen wrote:

> > + list_for_each_entry(plugin, &plugin_head, headers) {
> > +       if (GetPluginType(plugin->id) == EVMS_DEVICE_MANAGER) {
> > +             spin_unlock(&plugin_lock);
> > +             DISCOVER(plugin, disk_list);
> > +             spin_lock(&plugin_lock);
> > +       }

> How do you know "plugin" and its successors are still valid when retaking
> the spinlock? Looks like you need a reference count on the object here.

The spinlock itself should protect the integrity of the
list. If a prev or next element in the list should be
removed, while in a discover function, then the prev
or next field in the current plugin will get updated,
but I don't believe that should cause the list_for_each_entry
macro problems traversing the remainding elements in the list.

The first instruction in every plugin's discover function
is a MOD_INC_USE_COUNT and the last before the return is
MOD_DEC_USE_COUNT. So there exists a small window by
which the current plugin might be unloaded between the
spinlock release and MOD_INC_USE_COUNT, and the
MOD_DEC_USE_COUNT and the spinlock reacquire.

We plan to register a "__this_module.can_unload()" that
should prevent plugin modules from unloading during
discovery.

Mark