Based on discussions in the ofiwg, I'm most of the way down the path of
treating utility providers as 'shadow' providers. They're mostly invisible to
an app.
EXCEPT, I shouldn't have saved updating fi_fabric() for last... We need some
way to go from:
struct fi_fabric_attr {
struct fid_fabric *fabric;
char *name;
char *prov_name;
uint32_t prov_version;
uint32_t api_version;
};
to the correct provider. The proposed behavior for utility/shadow providers is:
1. Are not affected by any filtering logic.
They are always enabled with debugging turned on.
2. Have their getinfo calls invoked if prov_name is any core provider.
If prov_name is a util provider, the call routes directly
3. Set prov_name to: core_name:ofi-xxx
4. Will only layer over core providers.
An internal flag is used to prevent nesting
5. Report themselves directly for FI_PROV_ATTR_ONLY getinfo calls.
For debug purposes
From the viewpoint of an app, this means that the returned prov_name may differ
from the getinfo input. The man pages don't specify the behavior on this. If
this isn't acceptable, then we either require the app to update to 1.5 to use a
util provider or keep the existing behavior with util providers exposed as
separate providers. (I don't like the idea of encoding the provider in the
fabric name). So, which option sucks less?
- Sean
_______________________________________________
ofiwg mailing list
[email protected]
http://lists.openfabrics.org/mailman/listinfo/ofiwg
