net: ethtool: Set the req_info->dev on DUMP requests for each dev

[PATCH net] net: ethtool: Set the req_info->dev on DUMP requests for each dev

Posted by Maxime Chevallier 11 months, 2 weeks ago

There are a few netlink commands that rely on the req_info->dev field
being populated by ethnl in their ->prepare_data() and ->fill_reply().

For a regular GET request, this will be set by ethnl_default_parse(),
which calls ethnl_parse_header_dev_get().

In the case of a DUMP request, the ->prepare_data() and ->fill_reply()
callbacks will be called with the req_info->dev being NULL, which can
cause discrepancies in the behaviour between GET and DUMP results.

The main impact is that ethnl_req_get_phydev() will not find any
phy_device, impacting :
 - plca
 - pse-pd
 - stats

Some other commands rely on req_info->dev, namely :
 - coalesce in ->fill_reply to look for an irq_moder

Although cable_test and tunnels also rely on req_info->dev being set,
that's not a problem for these commands as :
 - cable_test doesn't support DUMP
 - tunnels rolls its own ->dumpit (and sets dev in the req_info).
 - phy also has its own ->dumpit

All other commands use reply_data->dev (probably the correct way of
doing things) and aren't facing this issue.

Simply set the dev in the req_info context when iterating to dump each
dev.

Fixes: c15e065b46dc ("net: ethtool: Allow passing a phy index for some commands")
Signed-off-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
---

Fixes tag targets the phy-index commit, as it introduced a change in
behaviour for PLCA. From what I can tell, coalesce never correctly
detected irq_moder in DUMP requests.

We could also consider fixing all individual commands that use
req_info->dev, however I'm not actually sure it's incorrect to do so,
feel free to correct me though.

Maxime

 net/ethtool/netlink.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/net/ethtool/netlink.c b/net/ethtool/netlink.c
index b4c45207fa32..de967961d8fe 100644
--- a/net/ethtool/netlink.c
+++ b/net/ethtool/netlink.c
@@ -582,6 +582,7 @@ static int ethnl_default_dumpit(struct sk_buff *skb,
 		dev_hold(dev);
 		rcu_read_unlock();
 
+		ctx->req_info->dev = dev;
 		ret = ethnl_default_dump_one(skb, dev, ctx, genl_info_dump(cb));
 
 		rcu_read_lock();
-- 
2.48.1

Re: [PATCH net] net: ethtool: Set the req_info->dev on DUMP requests for each dev

Posted by Kory Maincent 11 months, 1 week ago

On Sun,  2 Mar 2025 17:21:36 +0100
Maxime Chevallier <maxime.chevallier@bootlin.com> wrote:

> There are a few netlink commands that rely on the req_info->dev field
> being populated by ethnl in their ->prepare_data() and ->fill_reply().
> 
> For a regular GET request, this will be set by ethnl_default_parse(),
> which calls ethnl_parse_header_dev_get().
> 
> In the case of a DUMP request, the ->prepare_data() and ->fill_reply()
> callbacks will be called with the req_info->dev being NULL, which can
> cause discrepancies in the behaviour between GET and DUMP results.
> 
> The main impact is that ethnl_req_get_phydev() will not find any
> phy_device, impacting :
>  - plca
>  - pse-pd
>  - stats
> 
> Some other commands rely on req_info->dev, namely :
>  - coalesce in ->fill_reply to look for an irq_moder
> 
> Although cable_test and tunnels also rely on req_info->dev being set,
> that's not a problem for these commands as :
>  - cable_test doesn't support DUMP
>  - tunnels rolls its own ->dumpit (and sets dev in the req_info).
>  - phy also has its own ->dumpit
> 
> All other commands use reply_data->dev (probably the correct way of
> doing things) and aren't facing this issue.
> 
> Simply set the dev in the req_info context when iterating to dump each
> dev.
> 
> Fixes: c15e065b46dc ("net: ethtool: Allow passing a phy index for some
> commands") Signed-off-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
> ---
> 
> Fixes tag targets the phy-index commit, as it introduced a change in
> behaviour for PLCA. From what I can tell, coalesce never correctly
> detected irq_moder in DUMP requests.
> 
> We could also consider fixing all individual commands that use
> req_info->dev, however I'm not actually sure it's incorrect to do so,
> feel free to correct me though.
> 
> Maxime
> 
>  net/ethtool/netlink.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/net/ethtool/netlink.c b/net/ethtool/netlink.c
> index b4c45207fa32..de967961d8fe 100644
> --- a/net/ethtool/netlink.c
> +++ b/net/ethtool/netlink.c
> @@ -582,6 +582,7 @@ static int ethnl_default_dumpit(struct sk_buff *skb,
>  		dev_hold(dev);
>  		rcu_read_unlock();
>  
> +		ctx->req_info->dev = dev;

I would rather put it in ethnl_default_dump_one() before
ethnl_init_reply_data() call.

With this change:
Reviewed-by: Kory Maincent <kory.maincent@bootlin.com>
Tested-by: Kory Maincent <kory.maincent@bootlin.com>

Thank you!

>  		ret = ethnl_default_dump_one(skb, dev, ctx,
> genl_info_dump(cb)); 
>  		rcu_read_lock();



-- 
Köry Maincent, Bootlin
Embedded Linux and kernel engineering
https://bootlin.com

Re: [PATCH net] net: ethtool: Set the req_info->dev on DUMP requests for each dev

Posted by Maxime Chevallier 11 months, 1 week ago

Hi Köry,

On Tue, 4 Mar 2025 17:00:51 +0100
Kory Maincent <kory.maincent@bootlin.com> wrote:

> On Sun,  2 Mar 2025 17:21:36 +0100
> Maxime Chevallier <maxime.chevallier@bootlin.com> wrote:
> 
> > There are a few netlink commands that rely on the req_info->dev field
> > being populated by ethnl in their ->prepare_data() and ->fill_reply().
> > 
> > For a regular GET request, this will be set by ethnl_default_parse(),
> > which calls ethnl_parse_header_dev_get().
> > 
> > In the case of a DUMP request, the ->prepare_data() and ->fill_reply()
> > callbacks will be called with the req_info->dev being NULL, which can
> > cause discrepancies in the behaviour between GET and DUMP results.
> > 
> > The main impact is that ethnl_req_get_phydev() will not find any
> > phy_device, impacting :
> >  - plca
> >  - pse-pd
> >  - stats
> > 
> > Some other commands rely on req_info->dev, namely :
> >  - coalesce in ->fill_reply to look for an irq_moder
> > 
> > Although cable_test and tunnels also rely on req_info->dev being set,
> > that's not a problem for these commands as :
> >  - cable_test doesn't support DUMP
> >  - tunnels rolls its own ->dumpit (and sets dev in the req_info).
> >  - phy also has its own ->dumpit
> > 
> > All other commands use reply_data->dev (probably the correct way of
> > doing things) and aren't facing this issue.
> > 
> > Simply set the dev in the req_info context when iterating to dump each
> > dev.
> > 
> > Fixes: c15e065b46dc ("net: ethtool: Allow passing a phy index for some
> > commands") Signed-off-by: Maxime Chevallier <maxime.chevallier@bootlin.com>
> > ---
> > 
> > Fixes tag targets the phy-index commit, as it introduced a change in
> > behaviour for PLCA. From what I can tell, coalesce never correctly
> > detected irq_moder in DUMP requests.
> > 
> > We could also consider fixing all individual commands that use
> > req_info->dev, however I'm not actually sure it's incorrect to do so,
> > feel free to correct me though.
> > 
> > Maxime
> > 
> >  net/ethtool/netlink.c | 1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/net/ethtool/netlink.c b/net/ethtool/netlink.c
> > index b4c45207fa32..de967961d8fe 100644
> > --- a/net/ethtool/netlink.c
> > +++ b/net/ethtool/netlink.c
> > @@ -582,6 +582,7 @@ static int ethnl_default_dumpit(struct sk_buff *skb,
> >  		dev_hold(dev);
> >  		rcu_read_unlock();
> >  
> > +		ctx->req_info->dev = dev;  
> 
> I would rather put it in ethnl_default_dump_one() before
> ethnl_init_reply_data() call.

Ah yes true, that'll be more consistent with the other callpaths.

> With this change:
> Reviewed-by: Kory Maincent <kory.maincent@bootlin.com>
> Tested-by: Kory Maincent <kory.maincent@bootlin.com>

Thank you for testing, I'll respin.

Maxime

>