March 6, 2026 at 20:22, "Nikolay Aleksandrov" <[email protected] mailto:[email protected]?to=%22Nikolay%20Aleksandrov%22%20%3Crazor%40blackwall.org%3E > wrote:
> > On Fri, Mar 06, 2026 at 02:42:05AM +0000, Jiayuan Chen wrote: > > > > > 2026/3/6 05:03, "Jay Vosburgh" <[email protected] > > mailto:[email protected]?to=%22Jay%20Vosburgh%22%20%3Cjv%40jvosburgh.net%3E > > > wrote: > > > > > > > > Nikolay Aleksandrov <[email protected]> wrote: > > > > > > > > On Wed, Mar 04, 2026 at 09:27:28AM -0800, Jay Vosburgh wrote: > > > > > > > > > > > Nikolay Aleksandrov <[email protected]> wrote: > > > > > > > > >On Wed, Mar 04, 2026 at 03:42:57PM +0800, Jiayuan Chen wrote: > > > > >> From: Jiayuan Chen <[email protected]> > > > > >> > > > > >> bond_rr_gen_slave_id() dereferences bond->rr_tx_counter without a > > NULL > > > > >> check. rr_tx_counter is a per-CPU counter only allocated in > > bond_open() > > > > >> when the bond mode is round-robin. If the bond device was never > > brought > > > > >> up, rr_tx_counter remains NULL, causing a null-ptr-deref. > > > > >> > > > > >> The XDP redirect path can reach this code even when the bond is not > > up: > > > > >> bpf_master_redirect_enabled_key is a global static key, so when any > > bond > > > > >> device has native XDP attached, the XDP_TX -> xdp_master_redirect() > > > > >> interception is enabled for all bond slaves system-wide. This > > allows the > > > > >> path xdp_master_redirect() -> bond_xdp_get_xmit_slave() -> > > > > >> bond_xdp_xmit_roundrobin_slave_get() -> bond_rr_gen_slave_id() to be > > > > >> reached on a bond that was never opened. > > > > >> > > > > >> Fix this by allocating rr_tx_counter unconditionally in bond_init() > > > > >> (ndo_init), which is called by register_netdevice() and covers both > > > > >> device creation paths (bond_create() and bond_newlink()). This also > > > > >> handles the case where bond mode is changed to round-robin after > > device > > > > >> creation. The conditional allocation in bond_open() is removed. > > Since > > > > >> bond_destructor() already unconditionally calls > > > > >> free_percpu(bond->rr_tx_counter), the lifecycle is clean: allocate > > at > > > > >> ndo_init, free at destructor. > > > > >> > > > > >> Note: rr_tx_counter is only used by round-robin mode, so this > > > > >> deliberately allocates a per-cpu u32 that goes unused for other > > modes. > > > > >> Conditional allocation (e.g., in bond_option_mode_set) was > > considered > > > > >> but rejected: the XDP path can race with mode changes on a downed > > bond, > > > > >> and adding memory barriers to the XDP hot path is not justified for > > > > >> saving 4 bytes per CPU. > > > > >> > > > > >> Fixes: 879af96ffd72 ("net, core: Add support for XDP redirection to > > slave device") > > > > >> Reported-by: [email protected] > > > > >> Closes: > > https://lore.kernel.org/all/[email protected]/T/ > > > > >> Signed-off-by: Jiayuan Chen <[email protected]> > > > > >> --- > > > > >> drivers/net/bonding/bond_main.c | 19 +++++++++++++------ > > > > >> 1 file changed, 13 insertions(+), 6 deletions(-) > > > > >> > > > > > > > > > >IMO it's not worth it to waste memory in all modes, for an unpopular > > mode. > > > > >I think it'd be better to add a null check in bond_rr_gen_slave_id(), > > > > >READ/WRITE_ONCE() should be enough since it is allocated only once, > > and > > > > >freed when the xmit code cannot be reachable anymore (otherwise we'd > > have > > > > >more bugs now). The branch will be successfully predicted practically > > always, > > > > >and you can also mark the ptr being null as unlikely. That way only > > RR takes > > > > >a very minimal hit, if any. > > > > > > > > Is what you're suggesting different from Jiayuan's proposal[0], > > > > in the sense of needing barriers in the XDP hot path to insure > > ordering? > > > > > > > > If I understand correctly, your suggestion is something like > > > > (totally untested): > > > > > > > Basically yes, that is what I'm proposing + an unlikely() around that > > > null check since it is really unlikely and will be always predicted > > > correctly, this way it's only for RR mode. > > > > > Jiayuan, > > > > Do you agree that the patch below (including Nikolay's > > suggestion to add "unlikely") resolves the original issue without memory > > waste, and without introducing performance issues (barriers) into the > > XDP path? > > > > > > Sure, it's basically similar to what my v1 did, but the patch below can be > > more generic. > > > > > > https://lore.kernel.org/netdev/[email protected]/T/#m08e3e53a8aa8d837ddc9242f4b14f2651a2b00aa > > > IMO that is worse, you'll add 2 new tests and potentially 1 more cache line > for everyone in a hot path that is used a lot and must be kept as fast as > possible. Hi Nikolay, Apologies for the confusion. When I said "the patch below," I was not referring to the v1 link, but to the diff snippet appended at the bottom of my email. That specific snippet (using a local temp pointer) is the optimal approach. It ensures the NULL check for bond->rr_tx_counter is handled efficiently without introducing the performance overhead or memory waste you were concerned about. Best regards, Jiayuan > > > > -J > > > > > > > > > > > > > diff --git a/drivers/net/bonding/bond_main.c > > b/drivers/net/bonding/bond_main.c > > > > index eb27cacc26d7..ac2a4fc0aad0 100644 > > > > --- a/drivers/net/bonding/bond_main.c > > > > +++ b/drivers/net/bonding/bond_main.c > > > > @@ -4273,13 +4273,17 @@ void bond_work_cancel_all(struct bonding *bond) > > > > static int bond_open(struct net_device *bond_dev) > > > > { > > > > struct bonding *bond = netdev_priv(bond_dev); > > > > + u32 __percpu *rr_tx_tmp; > > > > struct list_head *iter; > > > > struct slave *slave; > > > > > > > > - if (BOND_MODE(bond) == BOND_MODE_ROUNDROBIN && !bond->rr_tx_counter) > > { > > > > - bond->rr_tx_counter = alloc_percpu(u32); > > > > - if (!bond->rr_tx_counter) > > > > + if (BOND_MODE(bond) == BOND_MODE_ROUNDROBIN && > > > > + !READ_ONCE(bond->rr_tx_counter)) { > > > > + rr_tx_tmp = alloc_percpu(u32); > > > > + if (!rr_tx_tmp) > > > > return -ENOMEM; > > > > + WRITE_ONCE(bond->rr_tx_counter, rr_tx_tmp); > > > > + > > > > } > > > > > > > > /* reset slave->backup and slave->inactive */ > > > > @@ -4866,6 +4870,9 @@ static u32 bond_rr_gen_slave_id(struct bonding > > *bond) > > > > struct reciprocal_value reciprocal_packets_per_slave; > > > > int packets_per_slave = bond->params.packets_per_slave; > > > > > > > > + if (!READ_ONCE(bond->rr_tx_counter)) > > > > + packets_per_slave = 0; > > > > + > > > > switch (packets_per_slave) { > > > > case 0: > > > > slave_id = get_random_u32(); > > > > > > > > -J > > > > > > > > > > > > [0] > > https://lore.kernel.org/netdev/[email protected]/ > > > > > > > > > --- > > -Jay Vosburgh, [email protected] > > >
