Hi Kirill, and thanks for your time,

On 31 Jan 19 14:49, Kirill Tkhai ktkhai@virtuozzo.com wrote :

Hi, Alexandre,

On 31.01.2019 16:20, alexandre.besnard@softathome.com wrote:

quoted

From: Alexandre Besnard <redacted>

quoted

Device remaining references counter is get as a signed integer.

quoted

When unregistering network devices, the loop waiting for this counter
to decrement tests the 0 strict equality. Thus if an error occurs and
two references are given back by a protocol, we are stuck in the loop
forever, with a -1 value.

quoted

Robustness is added by checking a negative value: the device is then
considered free of references, and a warning is issued (it should not
happen, one should check that behavior)

quoted

Signed-off-by: Alexandre Besnard <redacted>
---
net/core/dev.c | 5 +++++
1 file changed, 5 insertions(+)

quoted

diff --git a/net/core/dev.c b/net/core/dev.c
index ddc551f..e4190ae 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c

@@ -8687,6 +8687,11 @@ static void netdev_wait_allrefs(struct net_device *dev)

refcnt = netdev_refcnt_read(dev);

quoted

while (refcnt != 0) {
+ if (refcnt < 0) {
+ pr_warn("Device %s refcnt negative: device considered free, but it should not
happen\n",
+ dev->name);
+ break;
+ }

1)I don't think this is a good approach. Negative value does not guarantee
there is just a double put of device reference. Negative value is an indicator
something goes wrong, and we definitely should not free device memory in
this case.

2)Not related to your patch -- it looks like we have problem in existing
code with this netdev_refcnt_read(). It does not imply a memory ordering
or some guarantees about reading percpu values. For example, in generic
code struct percpu_ref switches a counter into atomic mode before it checks
for the last reference. But there is nothing in netdev_refcnt_read().

I agree with you, as it is not a full fix for a bad behavior of the refcnt: many
wrong things could happen here, and that's why I added a warning (short of a
more critical flag I could think of).

However, I think this is a good approach as a global workaround for any critical
situation caused by a negative refcnt, acting as a failsafe. What I try to avoid
here is not the bug, but a situation such as a deadlock keeping a system from
powering off, or way worse in the system life.
On the other hand, I can't think of a critical situation caused by freeing
the device memory. Processes or even systems may crash in some cases, but it
should be an expected behavior in such a case IMHO.

Actually, I think that with the current implementation, most of the systems
locked in the problem are powered off.

Do you think of any issue beyond this behavior ?