| Message ID | f92677bd76573f543b6486c6b1e8d5c4a6c3fa45.1537372746.git.gaetan.rivet@6wind.com (mailing list archive) |
|---|---|
| State | Superseded, archived |
| Delegated to: | Thomas Monjalon |
| Headers | show |
| Series | Implement new devargs framework | expand |
| Context | Check | Description |
|---|---|---|
| ci/checkpatch | success | coding style OK |
| ci/Intel-compilation | success | Compilation OK |
19/09/2018 18:03, Gaetan Rivet: > The eth device class can now parse a field name, > matching the eth_dev name with one passed as > > "class=eth,name=xxxxxx" I am not sure what is the purpose of the "name" property. I think we should not need it to choose a port by its ethdev name. If you are thinking about a vdev, we can use the rte_device name (at bus level).
On Thu, Sep 20, 2018 at 06:17:13PM +0200, Thomas Monjalon wrote: > 19/09/2018 18:03, Gaetan Rivet: > > The eth device class can now parse a field name, > > matching the eth_dev name with one passed as > > > > "class=eth,name=xxxxxx" > > I am not sure what is the purpose of the "name" property. > I think we should not need it to choose a port by its ethdev name. rte_eth_dev_get_port_by_name seems pretty close. Which fields do you think should be proposed first as a getter on the eth class? Or do you have a list of fields the eth class should be restricted to? > If you are thinking about a vdev, we can use the rte_device name (at bus level). This patch is only about eth class, it has no impact on buses.
21/09/2018 14:16, Gaëtan Rivet: > On Thu, Sep 20, 2018 at 06:17:13PM +0200, Thomas Monjalon wrote: > > 19/09/2018 18:03, Gaetan Rivet: > > > The eth device class can now parse a field name, > > > matching the eth_dev name with one passed as > > > > > > "class=eth,name=xxxxxx" > > > > I am not sure what is the purpose of the "name" property. > > I think we should not need it to choose a port by its ethdev name. > > rte_eth_dev_get_port_by_name seems pretty close. Exact, this function already exists to get a port id by name. I think we should discourage the use of ethdev "internal" name. The point of the devargs is to match device with explicit known properties. > Which fields do you think should be proposed first as a getter on the eth class? > Or do you have a list of fields the eth class should be restricted to? In other words, which ethdev properties can help to match a port? I think about "mac=" (which is already implemented in OVS devargs), and "representor=" (which requires a new field in ethdev). About representors, we could also match all representor ports of a switch (see rte_eth_switch_info). We could also have a property for the kernel netdev we are (or were) bound. Does it make sense? > > If you are thinking about a vdev, we can use the rte_device name (at bus level). > > This patch is only about eth class, it has no impact on buses. So we agree :)
diff --git a/lib/librte_ethdev/rte_class_eth.c b/lib/librte_ethdev/rte_class_eth.c index b6557db97..66fd48dc2 100644 --- a/lib/librte_ethdev/rte_class_eth.c +++ b/lib/librte_ethdev/rte_class_eth.c @@ -15,10 +15,12 @@ #include "ethdev_private.h" enum eth_params { + RTE_ETH_PARAMS_NAME, RTE_ETH_PARAMS_MAX, }; static const char * const eth_params_keys[] = { + [RTE_ETH_PARAMS_NAME] = "name", [RTE_ETH_PARAMS_MAX] = NULL, }; @@ -39,6 +41,7 @@ eth_dev_match(const struct rte_eth_dev *edev, { const struct eth_dev_match_arg *arg = _arg; const struct rte_kvargs *kvlist = arg->kvlist; + struct rte_eth_dev_data *data; if (edev->state == RTE_ETH_DEV_UNUSED) return -1; @@ -47,6 +50,10 @@ eth_dev_match(const struct rte_eth_dev *edev, if (kvlist == NULL) /* Empty string matches everything. */ return 0; + data = edev->data; + if (rte_kvargs_process(kvlist, "name", + rte_kvargs_strcmp, data->name)) + return -1; return 0; }