[3/5] ethdev: allow iterating with only class filter

  If no rte_device is given in the iterator,
eth_dev_match() is looking at all ports without any restriction,
except the ethdev kvargs filter.

It allows to iterate with a devargs filter referencing only
some ethdev parameters. The format (from the new devargs syntax) is:
	class=eth,paramY=Y

Fixes: e815a7f69371 ("ethdev: register as a class")

Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
---
 lib/librte_ethdev/rte_class_eth.c |  2 +-
 lib/librte_ethdev/rte_ethdev.c    | 13 +++++++++++--
 2 files changed, 12 insertions(+), 3 deletions(-)
  

On 10/8/18 1:25 AM, Thomas Monjalon wrote:
> If no rte_device is given in the iterator,
> eth_dev_match() is looking at all ports without any restriction,
> except the ethdev kvargs filter.
>
> It allows to iterate with a devargs filter referencing only
> some ethdev parameters. The format (from the new devargs syntax) is:
> 	class=eth,paramY=Y
>
> Fixes: e815a7f69371 ("ethdev: register as a class")
>
> Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
> ---
>   lib/librte_ethdev/rte_class_eth.c |  2 +-
>   lib/librte_ethdev/rte_ethdev.c    | 13 +++++++++++--
>   2 files changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/lib/librte_ethdev/rte_class_eth.c b/lib/librte_ethdev/rte_class_eth.c
> index 84b646291..f0af51c36 100644
> --- a/lib/librte_ethdev/rte_class_eth.c
> +++ b/lib/librte_ethdev/rte_class_eth.c
> @@ -42,7 +42,7 @@ eth_dev_match(const struct rte_eth_dev *edev,
>   
>   	if (edev->state == RTE_ETH_DEV_UNUSED)
>   		return -1;
> -	if (edev->device != arg->device)
> +	if (arg->device != NULL && arg->device != edev->device)
>   		return -1;
>   	if (kvlist == NULL)
>   		/* Empty string matches everything. */

It looks like it is the only hunk which
Fixes: e815a7f69371 ("ethdev: register as a class")

everything else adjusts the previous patch.
I think this fix should go before and the rest should be squashed
in the previous patch. It was really questionable why it is safe
to dereference iter->bus without checking that it is not NULL.

> diff --git a/lib/librte_ethdev/rte_ethdev.c b/lib/librte_ethdev/rte_ethdev.c
> index 83ab28c23..a43e0ab3a 100644
> --- a/lib/librte_ethdev/rte_ethdev.c
> +++ b/lib/librte_ethdev/rte_ethdev.c
> @@ -199,10 +199,18 @@ rte_eth_iterator_init(struct rte_dev_iterator *iter, const char *devargs_str)
>   	 * The devargs string may use various syntaxes:
>   	 *   - 0000:08:00.0,representor=[1-3]
>   	 *   - pci:0000:06:00.0,representor=[0,5]
> +	 *   - class=eth,mac=00:11:22:33:44:55
>   	 * A new syntax is in development (not yet supported):
>   	 *   - bus=X,paramX=x/class=Y,paramY=y/driver=Z,paramZ=z
>   	 */
>   
> +	/* Handle a case from future syntax, without any bus-level argument. */
> +	if (strncmp(devargs_str, iter_anybus_str,
> +			strlen(iter_anybus_str)) == 0) {
> +		iter->cls_str = devargs_str + strlen(iter_anybus_str);
> +		goto end;
> +	}
> +

It looks like a hack, but I guess we need it since rte_devargs_parse()
cannot handle the case. May be it is acceptable if we have no time
to solve it, but it would be good to highlight it better in the comments.


>   	/* Split bus, device and parameters. */
>   	ret = rte_devargs_parse(&devargs, devargs_str);
>   	if (ret != 0)
> @@ -239,6 +247,7 @@ rte_eth_iterator_init(struct rte_dev_iterator *iter, const char *devargs_str)
>   	}
>   	iter->bus_str = bus_str;
>   
> +end:
>   	iter->cls = rte_class_find_by_name("eth");
>   	return 0;
>   }
> @@ -250,7 +259,7 @@ rte_eth_iterator_next(struct rte_dev_iterator *iter)
>   		return RTE_MAX_ETHPORTS;
>   
>   	do { /* loop for matching rte_device */
> -		if (iter->class_device == NULL) {
> +		if (iter->bus != NULL && iter->class_device == NULL) {
>   			iter->device = iter->bus->dev_iterate(
>   					iter->device, iter->bus_str, iter);
>   			if (iter->device == NULL)
> @@ -260,7 +269,7 @@ rte_eth_iterator_next(struct rte_dev_iterator *iter)
>   				iter->class_device, iter->cls_str, iter);
>   		if (iter->class_device != NULL)
>   			return eth_dev_to_id(iter->class_device);
> -	} while (iter->class_device == NULL);
> +	} while (iter->bus != NULL && iter->class_device == NULL);
>   
>   	/* No more ethdev port to iterate. */
>   	free(RTE_CAST_FIELD(iter, bus_str, char *)); /* workaround const */
  

08/10/2018 09:20, Andrew Rybchenko:
> On 10/8/18 1:25 AM, Thomas Monjalon wrote:
> > If no rte_device is given in the iterator,
> > eth_dev_match() is looking at all ports without any restriction,
> > except the ethdev kvargs filter.
> >
> > It allows to iterate with a devargs filter referencing only
> > some ethdev parameters. The format (from the new devargs syntax) is:
> > 	class=eth,paramY=Y
> >
> > Fixes: e815a7f69371 ("ethdev: register as a class")
> >
> > Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
> > ---
> > --- a/lib/librte_ethdev/rte_class_eth.c
> > +++ b/lib/librte_ethdev/rte_class_eth.c
> > @@ -42,7 +42,7 @@ eth_dev_match(const struct rte_eth_dev *edev,
> >   
> >   	if (edev->state == RTE_ETH_DEV_UNUSED)
> >   		return -1;
> > -	if (edev->device != arg->device)
> > +	if (arg->device != NULL && arg->device != edev->device)
> >   		return -1;
> >   	if (kvlist == NULL)
> >   		/* Empty string matches everything. */
> 
> It looks like it is the only hunk which
> Fixes: e815a7f69371 ("ethdev: register as a class")

Yes this hunk is fixing above commit.

> everything else adjusts the previous patch.

Yes but the whole goal of this patch is to allow ethdev pure filter.
All is related in this patch.

> I think this fix should go before and the rest should be squashed
> in the previous patch. It was really questionable why it is safe
> to dereference iter->bus without checking that it is not NULL.

No iter->bus was safe because iter->cls was checked before and
implied that iter->bus was successfully set.

I still think it is better for understanding to split different
kind of filters in 2 patches.
I may remove the Fixes line however. Opinion?

[...]
> > +	/* Handle a case from future syntax, without any bus-level argument. */
> > +	if (strncmp(devargs_str, iter_anybus_str,
> > +			strlen(iter_anybus_str)) == 0) {
> > +		iter->cls_str = devargs_str + strlen(iter_anybus_str);
> > +		goto end;
> > +	}
> > +
> 
> It looks like a hack, but I guess we need it since rte_devargs_parse()
> cannot handle the case. May be it is acceptable if we have no time
> to solve it, but it would be good to highlight it better in the comments.

This function is a mix of old and new syntax.
When only new syntax will be supported, it will be better.
It is a work in progress.
I do not know how to better explain it.
  

On 10/8/18 11:07 AM, Thomas Monjalon wrote:
> 08/10/2018 09:20, Andrew Rybchenko:
>> On 10/8/18 1:25 AM, Thomas Monjalon wrote:
>>> If no rte_device is given in the iterator,
>>> eth_dev_match() is looking at all ports without any restriction,
>>> except the ethdev kvargs filter.
>>>
>>> It allows to iterate with a devargs filter referencing only
>>> some ethdev parameters. The format (from the new devargs syntax) is:
>>> 	class=eth,paramY=Y
>>>
>>> Fixes: e815a7f69371 ("ethdev: register as a class")
>>>
>>> Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
>>> ---
>>> --- a/lib/librte_ethdev/rte_class_eth.c
>>> +++ b/lib/librte_ethdev/rte_class_eth.c
>>> @@ -42,7 +42,7 @@ eth_dev_match(const struct rte_eth_dev *edev,
>>>    
>>>    	if (edev->state == RTE_ETH_DEV_UNUSED)
>>>    		return -1;
>>> -	if (edev->device != arg->device)
>>> +	if (arg->device != NULL && arg->device != edev->device)
>>>    		return -1;
>>>    	if (kvlist == NULL)
>>>    		/* Empty string matches everything. */
>> It looks like it is the only hunk which
>> Fixes: e815a7f69371 ("ethdev: register as a class")
> Yes this hunk is fixing above commit.
>
>> everything else adjusts the previous patch.
> Yes but the whole goal of this patch is to allow ethdev pure filter.
> All is related in this patch.
>
>> I think this fix should go before and the rest should be squashed
>> in the previous patch. It was really questionable why it is safe
>> to dereference iter->bus without checking that it is not NULL.
> No iter->bus was safe because iter->cls was checked before and
> implied that iter->bus was successfully set.

OK, I see now.

> I still think it is better for understanding to split different
> kind of filters in 2 patches.
> I may remove the Fixes line however. Opinion?

Yes, it makes sense. If so, I think it is better to remove Fixes line.
With Fixes I would expect to be able to apply it before previous patch.

> [...]
>>> +	/* Handle a case from future syntax, without any bus-level argument. */
>>> +	if (strncmp(devargs_str, iter_anybus_str,
>>> +			strlen(iter_anybus_str)) == 0) {
>>> +		iter->cls_str = devargs_str + strlen(iter_anybus_str);
>>> +		goto end;
>>> +	}
>>> +
>> It looks like a hack, but I guess we need it since rte_devargs_parse()
>> cannot handle the case. May be it is acceptable if we have no time
>> to solve it, but it would be good to highlight it better in the comments.
> This function is a mix of old and new syntax.
> When only new syntax will be supported, it will be better.
> It is a work in progress.
> I do not know how to better explain it.

I'd suggest to highligh rte_devargs_parse() limitations here to
explain why special handling is required here.