Re: [PATCH net-next 02/13] ipv6: annotate data-races around cnf.disable_ipv6

[Jiri Pirko <jiri@resnulli.us>]

Mon, Feb 26, 2024 at 05:14:36PM CET, edumazet@google.com wrote:
>On Mon, Feb 26, 2024 at 5:09 PM Jiri Pirko <jiri@resnulli.us> wrote:
>>
>> Mon, Feb 26, 2024 at 04:50:44PM CET, edumazet@google.com wrote:
>> >disable_ipv6 is read locklessly, add appropriate READ_ONCE()
>> >and WRITE_ONCE() annotations.
>> >
>> >Signed-off-by: Eric Dumazet <edumazet@google.com>
>> >---
>> > net/ipv6/addrconf.c   | 12 ++++++------
>> > net/ipv6/ip6_input.c  |  4 ++--
>> > net/ipv6/ip6_output.c |  2 +-
>> > 3 files changed, 9 insertions(+), 9 deletions(-)
>> >
>> >diff --git a/net/ipv6/addrconf.c b/net/ipv6/addrconf.c
>> >index a280614b37652deee0d1f3c70ba1b41b01cc7d91..0d7746b113cc65303b5c2ec223b3331c3598ded6 100644
>> >--- a/net/ipv6/addrconf.c
>> >+++ b/net/ipv6/addrconf.c
>> >@@ -4214,7 +4214,7 @@ static void addrconf_dad_work(struct work_struct *w)
>> >                       if (!ipv6_generate_eui64(addr.s6_addr + 8, idev->dev) &&
>> >                           ipv6_addr_equal(&ifp->addr, &addr)) {
>> >                               /* DAD failed for link-local based on MAC */
>> >-                              idev->cnf.disable_ipv6 = 1;
>> >+                              WRITE_ONCE(idev->cnf.disable_ipv6, 1);
>> >
>> >                               pr_info("%s: IPv6 being disabled!\n",
>> >                                       ifp->idev->dev->name);
>> >@@ -6388,7 +6388,8 @@ static void addrconf_disable_change(struct net *net, __s32 newf)
>> >               idev = __in6_dev_get(dev);
>> >               if (idev) {
>> >                       int changed = (!idev->cnf.disable_ipv6) ^ (!newf);
>> >-                      idev->cnf.disable_ipv6 = newf;
>> >+
>> >+                      WRITE_ONCE(idev->cnf.disable_ipv6, newf);
>> >                       if (changed)
>> >                               dev_disable_change(idev);
>> >               }
>> >@@ -6397,15 +6398,14 @@ static void addrconf_disable_change(struct net *net, __s32 newf)
>> >
>> > static int addrconf_disable_ipv6(struct ctl_table *table, int *p, int newf)
>> > {
>> >-      struct net *net;
>> >+      struct net *net = (struct net *)table->extra2;
>>
>> How is this related to the rest of the patch and why is it okay to
>> access table->extra2 without holding rtnl mutex?
>
>table->extra2 is immutable, it can be fetched before grabbing RTNL.
>Everything that can be done before acquiring RTNL is a win under RTNL pressure.
>
>I had a followup minor patch, but the patch series was already too big.

I see, so this hunk should be part of that patch, not this one, I
believe.


>
>We do not need to grab rtnl when changing net->ipv6.devconf_dflt->disable_ipv6
>
>diff --git a/net/ipv6/addrconf.c b/net/ipv6/addrconf.c
>index 08b4728733e3ed16d139d2bd4b50328552b3c27f..befe2709acdffdce8c6a3304df8dec598246a682
>100644
>--- a/net/ipv6/addrconf.c
>+++ b/net/ipv6/addrconf.c
>@@ -6398,17 +6398,16 @@ static int addrconf_disable_ipv6(struct
>ctl_table *table, int *p, int newf)
>        struct net *net = (struct net *)table->extra2;
>        int old;
>
>+       if (p == &net->ipv6.devconf_dflt->disable_ipv6) {
>+               WRITE_ONCE(*p, newf);
>+               return 0;
>+       }
>        if (!rtnl_trylock())
>                return restart_syscall();
>
>        old = *p;
>        WRITE_ONCE(*p, newf);
>
>-       if (p == &net->ipv6.devconf_dflt->disable_ipv6) {
>-               rtnl_unlock();
>-               return 0;
>-       }
>-
>        if (p == &net->ipv6.devconf_all->disable_ipv6) {
>                WRITE_ONCE(net->ipv6.devconf_dflt->disable_ipv6, newf);
>                addrconf_disable_change(net, newf);