Re: [PATCH net] net: sunhme: Return an error when we are out of slots

From: Sean Anderson
Date: Wed Feb 22 2023 - 16:07:54 EST


On 2/22/23 15:59, Simon Horman wrote:
On Wed, Feb 22, 2023 at 12:09:35PM -0500, Sean Anderson wrote:
We only allocate enough space for four devices when the parent is a QFE. If
we couldn't find a spot (because five devices were created for whatever
reason), we would not return an error from probe(). Return ENODEV, which
was what we did before.

Fixes: 96c6e9faecf1 ("sunhme: forward the error code from pci_enable_device()")

I think the hash for that commit is acb3f35f920b.

Ah, sorry that's my local copy. The upstream commit is as you noted.


However, I also think this problem was introduced by the first hunk of
5b3dc6dda6b1 ("sunhme: Regularize probe errors").

Which is:

--- a/drivers/net/ethernet/sun/sunhme.c
+++ b/drivers/net/ethernet/sun/sunhme.c
@@ -2945,7 +2945,6 @@ static int happy_meal_pci_probe(struct pci_dev *pdev,
if (err)
goto err_out;
pci_set_master(pdev);
- err = -ENODEV;
if (!strcmp(prom_name, "SUNW,qfe") || !strcmp(prom_name, "qfe")) {
qp = quattro_pci_find(pdev);


Yes. That's the one I should have blamed.

Which leads me to wonder if simpler fixes would be either:

1) Reverting the hunk above
2) Or, more in keeping with the rest of that patch,
explicitly setting err before branching to err_out,
as you your patch does, but without other logic changes.

Something like this (*compile tested only!*:

diff --git a/drivers/net/ethernet/sun/sunhme.c b/drivers/net/ethernet/sun/sunhme.c
index 1c16548415cd..2409e7d6c29e 100644
--- a/drivers/net/ethernet/sun/sunhme.c
+++ b/drivers/net/ethernet/sun/sunhme.c
@@ -2863,8 +2863,10 @@ static int happy_meal_pci_probe(struct pci_dev *pdev,
if (!qp->happy_meals[qfe_slot])
break;
- if (qfe_slot == 4)
+ if (qfe_slot == 4) {
+ err = -ENOMEM;
goto err_out;
+ }
}
dev = devm_alloc_etherdev(&pdev->dev, sizeof(struct happy_meal));

That's of course simpler, but this also does some cleanup to make it more
obvious what's going on.

Also, I am curious why happy_meal_pci_probe() doesn't just return instaed
of branching to err_out. As err_out only returns err. I guess there is a
reason for it. But simply returning would probably simplify error handling.
(I'm not suggesting that approach for this fix.)

I think it's because there used to be cleanup in err_out. But you're right,
we can just return directly and avoid a goto.

--Sean

Reported-by: kernel test robot <lkp@xxxxxxxxx>
Reported-by: Dan Carpenter <error27@xxxxxxxxx>
Signed-off-by: Sean Anderson <seanga2@xxxxxxxxx>
---

drivers/net/ethernet/sun/sunhme.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/sun/sunhme.c b/drivers/net/ethernet/sun/sunhme.c
index 1c16548415cd..523e26653ec8 100644
--- a/drivers/net/ethernet/sun/sunhme.c
+++ b/drivers/net/ethernet/sun/sunhme.c
@@ -2861,12 +2861,13 @@ static int happy_meal_pci_probe(struct pci_dev *pdev,
for (qfe_slot = 0; qfe_slot < 4; qfe_slot++)
if (!qp->happy_meals[qfe_slot])
- break;
+ goto found_slot;
- if (qfe_slot == 4)
- goto err_out;
+ err = -ENODEV;
+ goto err_out;
}
+found_slot:
dev = devm_alloc_etherdev(&pdev->dev, sizeof(struct happy_meal));
if (!dev) {
err = -ENOMEM;
--
2.37.1