[net] Question about ipvs->sysctl_sync_threshold and READ_ONCE

[net] Question about ipvs->sysctl_sync_threshold and READ_ONCE

[Zijie Zhao]

Dear IPVS maintainers,

We encountered an unusual usage of sysctl parameter while analyzing 
kernel source code.


In include/net/ip_vs.h, line 1062 - 1070:

```
static inline int sysctl_sync_threshold(struct netns_ipvs *ipvs)
{
	return ipvs->sysctl_sync_threshold[0];
}

static inline int sysctl_sync_period(struct netns_ipvs *ipvs)
{
	return READ_ONCE(ipvs->sysctl_sync_threshold[1]);
}
```

Here, sysctl_sync_threshold[1] is accessed behind `READ_ONCE`, but 
sysctl_sync_threshold[0] is not. Should sysctl_sync_threshold[0] also be 
guarded by `READ_ONCE`?

Please kindly let us know if we missed any key information and this is 
actually intended. We appreciate your information and time! Thanks!


Links to the code:
https://elixir.bootlin.com/linux/v6.8.1/source/include/net/ip_vs.h#L1064
https://elixir.bootlin.com/linux/v6.8.1/source/include/net/ip_vs.h#L1069

Best,
Zijie

Re: [net] Question about ipvs->sysctl_sync_threshold and READ_ONCE

[Jason Xing]

On Wed, Mar 20, 2024 at 6:49 AM Zijie Zhao <zzjas98@gmail.com> wrote:
>
> Dear IPVS maintainers,
>
> We encountered an unusual usage of sysctl parameter while analyzing
> kernel source code.
>
>
> In include/net/ip_vs.h, line 1062 - 1070:
>
> ```
> static inline int sysctl_sync_threshold(struct netns_ipvs *ipvs)
> {
>         return ipvs->sysctl_sync_threshold[0];
> }
>
> static inline int sysctl_sync_period(struct netns_ipvs *ipvs)
> {
>         return READ_ONCE(ipvs->sysctl_sync_threshold[1]);
> }
> ```
>
> Here, sysctl_sync_threshold[1] is accessed behind `READ_ONCE`, but
> sysctl_sync_threshold[0] is not. Should sysctl_sync_threshold[0] also be
> guarded by `READ_ONCE`?

I'm not so sure and clear about the detailed history.

AFAIK, readers accessing this sysctl knob (sysctl_sync_threshold)
should be protected by READ_ONCE() because it can be changed
concurrently. Probably the commit 749c42b620a95 missed this point many
years ago and then commit 6aa7de059173a followed that.

Thanks,
Jason

>
> Please kindly let us know if we missed any key information and this is
> actually intended. We appreciate your information and time! Thanks!
>
>
> Links to the code:
> https://elixir.bootlin.com/linux/v6.8.1/source/include/net/ip_vs.h#L1064
> https://elixir.bootlin.com/linux/v6.8.1/source/include/net/ip_vs.h#L1069
>
> Best,
> Zijie
>

Re: [net] Question about ipvs->sysctl_sync_threshold and READ_ONCE

[Julian Anastasov]

	Hello,

On Tue, 19 Mar 2024, Zijie Zhao wrote:

> Dear IPVS maintainers,
> 
> We encountered an unusual usage of sysctl parameter while analyzing kernel
> source code.
> 
> 
> In include/net/ip_vs.h, line 1062 - 1070:
> 
> ```
> static inline int sysctl_sync_threshold(struct netns_ipvs *ipvs)
> {
> 	return ipvs->sysctl_sync_threshold[0];
> }
> 
> static inline int sysctl_sync_period(struct netns_ipvs *ipvs)
> {
> 	return READ_ONCE(ipvs->sysctl_sync_threshold[1]);
> }
> ```
> 
> Here, sysctl_sync_threshold[1] is accessed behind `READ_ONCE`, but
> sysctl_sync_threshold[0] is not. Should sysctl_sync_threshold[0] also be
> guarded by `READ_ONCE`?
> 
> Please kindly let us know if we missed any key information and this is
> actually intended. We appreciate your information and time! Thanks!

	Difference comes from commit 749c42b620a9 where we protect us
from division by zero by using ACCESS_ONCE at that time. The idea was
to read the value only once. Nowadays, READ_ONCE also has the role to
avoid load tearing, so, yes, all sysctl vars should be accessed with
READ_ONCE but this is a low priority goal for now.

> Links to the code:
> https://elixir.bootlin.com/linux/v6.8.1/source/include/net/ip_vs.h#L1064
> https://elixir.bootlin.com/linux/v6.8.1/source/include/net/ip_vs.h#L1069
> 
> Best,
> Zijie

Regards

--
Julian Anastasov <ja@ssi.bg>