| Message ID | 20240821090458.10813-4-jianbol@nvidia.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | Fixes for IPsec over bonding | expand |
Jianbo Liu <jianbol@nvidia.com> wrote: >In the cited commit, bond->ipsec_lock is added to protect ipsec_list, >hence xdo_dev_state_add and xdo_dev_state_delete are called inside >this lock. As ipsec_lock is a spin lock and such xfrmdev ops may sleep, >"scheduling while atomic" will be triggered when changing bond's >active slave. > >[ 101.055189] BUG: scheduling while atomic: bash/902/0x00000200 >[ 101.055726] Modules linked in: >[ 101.058211] CPU: 3 PID: 902 Comm: bash Not tainted 6.9.0-rc4+ #1 >[ 101.058760] Hardware name: >[ 101.059434] Call Trace: >[ 101.059436] <TASK> >[ 101.060873] dump_stack_lvl+0x51/0x60 >[ 101.061275] __schedule_bug+0x4e/0x60 >[ 101.061682] __schedule+0x612/0x7c0 >[ 101.062078] ? __mod_timer+0x25c/0x370 >[ 101.062486] schedule+0x25/0xd0 >[ 101.062845] schedule_timeout+0x77/0xf0 >[ 101.063265] ? asm_common_interrupt+0x22/0x40 >[ 101.063724] ? __bpf_trace_itimer_state+0x10/0x10 >[ 101.064215] __wait_for_common+0x87/0x190 >[ 101.064648] ? usleep_range_state+0x90/0x90 >[ 101.065091] cmd_exec+0x437/0xb20 [mlx5_core] >[ 101.065569] mlx5_cmd_do+0x1e/0x40 [mlx5_core] >[ 101.066051] mlx5_cmd_exec+0x18/0x30 [mlx5_core] >[ 101.066552] mlx5_crypto_create_dek_key+0xea/0x120 [mlx5_core] >[ 101.067163] ? bonding_sysfs_store_option+0x4d/0x80 [bonding] >[ 101.067738] ? kmalloc_trace+0x4d/0x350 >[ 101.068156] mlx5_ipsec_create_sa_ctx+0x33/0x100 [mlx5_core] >[ 101.068747] mlx5e_xfrm_add_state+0x47b/0xaa0 [mlx5_core] >[ 101.069312] bond_change_active_slave+0x392/0x900 [bonding] >[ 101.069868] bond_option_active_slave_set+0x1c2/0x240 [bonding] >[ 101.070454] __bond_opt_set+0xa6/0x430 [bonding] >[ 101.070935] __bond_opt_set_notify+0x2f/0x90 [bonding] >[ 101.071453] bond_opt_tryset_rtnl+0x72/0xb0 [bonding] >[ 101.071965] bonding_sysfs_store_option+0x4d/0x80 [bonding] >[ 101.072567] kernfs_fop_write_iter+0x10c/0x1a0 >[ 101.073033] vfs_write+0x2d8/0x400 >[ 101.073416] ? alloc_fd+0x48/0x180 >[ 101.073798] ksys_write+0x5f/0xe0 >[ 101.074175] do_syscall_64+0x52/0x110 >[ 101.074576] entry_SYSCALL_64_after_hwframe+0x4b/0x53 > >As bond_ipsec_add_sa_all and bond_ipsec_del_sa_all are only called >from bond_change_active_slave, which requires holding the RTNL lock. >And bond_ipsec_add_sa and bond_ipsec_del_sa are xfrm state >xdo_dev_state_add and xdo_dev_state_delete APIs, which are in user >context. So ipsec_lock doesn't have to be spin lock, change it to >mutex, and thus the above issue can be resolved. > >Fixes: 9a5605505d9c ("bonding: Add struct bond_ipesc to manage SA") >Signed-off-by: Jianbo Liu <jianbol@nvidia.com> >Signed-off-by: Tariq Toukan <tariqt@nvidia.com> >Reviewed-by: Hangbin Liu <liuhangbin@gmail.com> >--- > drivers/net/bonding/bond_main.c | 67 +++++++++++++++------------------ > include/net/bonding.h | 2 +- > 2 files changed, 32 insertions(+), 37 deletions(-) > >diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c >index 0d1129eaf47b..f20f6d83ad54 100644 >--- a/drivers/net/bonding/bond_main.c >+++ b/drivers/net/bonding/bond_main.c >@@ -439,38 +439,33 @@ static int bond_ipsec_add_sa(struct xfrm_state *xs, > rcu_read_lock(); > bond = netdev_priv(bond_dev); > slave = rcu_dereference(bond->curr_active_slave); >- if (!slave) { >- rcu_read_unlock(); >+ real_dev = slave ? slave->dev : NULL; >+ rcu_read_unlock(); >+ if (!real_dev) > return -ENODEV; In reading these, I was confused as to why some changes use rcu_read_lock(), rcu_dereference() and others use rtnl_dereference(); I think it's because bond_ipsec_{add,del}_sa_all() are guaranteed to be called under RTNL, while the bond_ipsec_{add,del}_sa() functions are do not have that guarantee. Am I understanding correctly? >- } > >- real_dev = slave->dev; > if (!real_dev->xfrmdev_ops || > !real_dev->xfrmdev_ops->xdo_dev_state_add || > netif_is_bond_master(real_dev)) { > NL_SET_ERR_MSG_MOD(extack, "Slave does not support ipsec offload"); >- rcu_read_unlock(); > return -EINVAL; > } > >- ipsec = kmalloc(sizeof(*ipsec), GFP_ATOMIC); >- if (!ipsec) { >- rcu_read_unlock(); >+ ipsec = kmalloc(sizeof(*ipsec), GFP_KERNEL); >+ if (!ipsec) > return -ENOMEM; Presumably the switch from ATOMIC to KERNEL is safe because this is only called under RTNL (and therefore always has a process context), i.e., this change is independent of any other changes in the patch. Correct? >- } > > xs->xso.real_dev = real_dev; > err = real_dev->xfrmdev_ops->xdo_dev_state_add(xs, extack); > if (!err) { > ipsec->xs = xs; > INIT_LIST_HEAD(&ipsec->list); >- spin_lock_bh(&bond->ipsec_lock); >+ mutex_lock(&bond->ipsec_lock); > list_add(&ipsec->list, &bond->ipsec_list); >- spin_unlock_bh(&bond->ipsec_lock); >+ mutex_unlock(&bond->ipsec_lock); > } else { > kfree(ipsec); > } >- rcu_read_unlock(); > return err; > } > >@@ -481,35 +476,35 @@ static void bond_ipsec_add_sa_all(struct bonding *bond) > struct bond_ipsec *ipsec; > struct slave *slave; > >- rcu_read_lock(); >- slave = rcu_dereference(bond->curr_active_slave); >- if (!slave) >- goto out; >+ slave = rtnl_dereference(bond->curr_active_slave); >+ real_dev = slave ? slave->dev : NULL; >+ if (!real_dev) >+ return; > >- real_dev = slave->dev; >+ mutex_lock(&bond->ipsec_lock); > if (!real_dev->xfrmdev_ops || > !real_dev->xfrmdev_ops->xdo_dev_state_add || > netif_is_bond_master(real_dev)) { >- spin_lock_bh(&bond->ipsec_lock); > if (!list_empty(&bond->ipsec_list)) > slave_warn(bond_dev, real_dev, > "%s: no slave xdo_dev_state_add\n", > __func__); >- spin_unlock_bh(&bond->ipsec_lock); > goto out; > } > >- spin_lock_bh(&bond->ipsec_lock); > list_for_each_entry(ipsec, &bond->ipsec_list, list) { >+ /* If new state is added before ipsec_lock acquired */ >+ if (ipsec->xs->xso.real_dev == real_dev) >+ continue; >+ > ipsec->xs->xso.real_dev = real_dev; > if (real_dev->xfrmdev_ops->xdo_dev_state_add(ipsec->xs, NULL)) { > slave_warn(bond_dev, real_dev, "%s: failed to add SA\n", __func__); > ipsec->xs->xso.real_dev = NULL; > } > } >- spin_unlock_bh(&bond->ipsec_lock); > out: >- rcu_read_unlock(); >+ mutex_unlock(&bond->ipsec_lock); > } > > /** >@@ -530,6 +525,8 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) > rcu_read_lock(); > bond = netdev_priv(bond_dev); > slave = rcu_dereference(bond->curr_active_slave); >+ real_dev = slave ? slave->dev : NULL; >+ rcu_read_unlock(); Is it really safe to access real_dev once we've left the rcu critical section? What prevents the device referenced by real_dev from being deleted as soon as rcu_read_unlock() completes? -J > > if (!slave) > goto out; >@@ -537,7 +534,6 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) > if (!xs->xso.real_dev) > goto out; > >- real_dev = slave->dev; > WARN_ON(xs->xso.real_dev != real_dev); > > if (!real_dev->xfrmdev_ops || >@@ -549,7 +545,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) > > real_dev->xfrmdev_ops->xdo_dev_state_delete(xs); > out: >- spin_lock_bh(&bond->ipsec_lock); >+ mutex_lock(&bond->ipsec_lock); > list_for_each_entry(ipsec, &bond->ipsec_list, list) { > if (ipsec->xs == xs) { > list_del(&ipsec->list); >@@ -557,8 +553,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) > break; > } > } >- spin_unlock_bh(&bond->ipsec_lock); >- rcu_read_unlock(); >+ mutex_unlock(&bond->ipsec_lock); > } > > static void bond_ipsec_del_sa_all(struct bonding *bond) >@@ -568,15 +563,12 @@ static void bond_ipsec_del_sa_all(struct bonding *bond) > struct bond_ipsec *ipsec; > struct slave *slave; > >- rcu_read_lock(); >- slave = rcu_dereference(bond->curr_active_slave); >- if (!slave) { >- rcu_read_unlock(); >+ slave = rtnl_dereference(bond->curr_active_slave); >+ real_dev = slave ? slave->dev : NULL; >+ if (!real_dev) > return; >- } > >- real_dev = slave->dev; >- spin_lock_bh(&bond->ipsec_lock); >+ mutex_lock(&bond->ipsec_lock); > list_for_each_entry(ipsec, &bond->ipsec_list, list) { > if (!ipsec->xs->xso.real_dev) > continue; >@@ -593,8 +585,7 @@ static void bond_ipsec_del_sa_all(struct bonding *bond) > real_dev->xfrmdev_ops->xdo_dev_state_free(ipsec->xs); > } > } >- spin_unlock_bh(&bond->ipsec_lock); >- rcu_read_unlock(); >+ mutex_unlock(&bond->ipsec_lock); > } > > static void bond_ipsec_free_sa(struct xfrm_state *xs) >@@ -5917,7 +5908,7 @@ void bond_setup(struct net_device *bond_dev) > /* set up xfrm device ops (only supported in active-backup right now) */ > bond_dev->xfrmdev_ops = &bond_xfrmdev_ops; > INIT_LIST_HEAD(&bond->ipsec_list); >- spin_lock_init(&bond->ipsec_lock); >+ mutex_init(&bond->ipsec_lock); > #endif /* CONFIG_XFRM_OFFLOAD */ > > /* don't acquire bond device's netif_tx_lock when transmitting */ >@@ -5966,6 +5957,10 @@ static void bond_uninit(struct net_device *bond_dev) > __bond_release_one(bond_dev, slave->dev, true, true); > netdev_info(bond_dev, "Released all slaves\n"); > >+#ifdef CONFIG_XFRM_OFFLOAD >+ mutex_destroy(&bond->ipsec_lock); >+#endif /* CONFIG_XFRM_OFFLOAD */ >+ > bond_set_slave_arr(bond, NULL, NULL); > > list_del_rcu(&bond->bond_list); >diff --git a/include/net/bonding.h b/include/net/bonding.h >index b61fb1aa3a56..8bb5f016969f 100644 >--- a/include/net/bonding.h >+++ b/include/net/bonding.h >@@ -260,7 +260,7 @@ struct bonding { > #ifdef CONFIG_XFRM_OFFLOAD > struct list_head ipsec_list; > /* protecting ipsec_list */ >- spinlock_t ipsec_lock; >+ struct mutex ipsec_lock; > #endif /* CONFIG_XFRM_OFFLOAD */ > struct bpf_prog *xdp_prog; > }; >-- >2.21.0 > --- -Jay Vosburgh, jv@jvosburgh.net
On Wed, 21 Aug 2024 09:00:30 -0700 Jay Vosburgh wrote: > Is it really safe to access real_dev once we've left the rcu > critical section? What prevents the device referenced by real_dev from > being deleted as soon as rcu_read_unlock() completes? Hah, I asked them this question at least 2 times. Let's see if your superior communication skills help :)
On Wed, 2024-08-21 at 09:00 -0700, Jay Vosburgh wrote: > Jianbo Liu <jianbol@nvidia.com> wrote: > > > In the cited commit, bond->ipsec_lock is added to protect > > ipsec_list, > > hence xdo_dev_state_add and xdo_dev_state_delete are called inside > > this lock. As ipsec_lock is a spin lock and such xfrmdev ops may > > sleep, > > "scheduling while atomic" will be triggered when changing bond's > > active slave. > > > > [ 101.055189] BUG: scheduling while atomic: bash/902/0x00000200 > > [ 101.055726] Modules linked in: > > [ 101.058211] CPU: 3 PID: 902 Comm: bash Not tainted 6.9.0-rc4+ #1 > > [ 101.058760] Hardware name: > > [ 101.059434] Call Trace: > > [ 101.059436] <TASK> > > [ 101.060873] dump_stack_lvl+0x51/0x60 > > [ 101.061275] __schedule_bug+0x4e/0x60 > > [ 101.061682] __schedule+0x612/0x7c0 > > [ 101.062078] ? __mod_timer+0x25c/0x370 > > [ 101.062486] schedule+0x25/0xd0 > > [ 101.062845] schedule_timeout+0x77/0xf0 > > [ 101.063265] ? asm_common_interrupt+0x22/0x40 > > [ 101.063724] ? __bpf_trace_itimer_state+0x10/0x10 > > [ 101.064215] __wait_for_common+0x87/0x190 > > [ 101.064648] ? usleep_range_state+0x90/0x90 > > [ 101.065091] cmd_exec+0x437/0xb20 [mlx5_core] > > [ 101.065569] mlx5_cmd_do+0x1e/0x40 [mlx5_core] > > [ 101.066051] mlx5_cmd_exec+0x18/0x30 [mlx5_core] > > [ 101.066552] mlx5_crypto_create_dek_key+0xea/0x120 [mlx5_core] > > [ 101.067163] ? bonding_sysfs_store_option+0x4d/0x80 [bonding] > > [ 101.067738] ? kmalloc_trace+0x4d/0x350 > > [ 101.068156] mlx5_ipsec_create_sa_ctx+0x33/0x100 [mlx5_core] > > [ 101.068747] mlx5e_xfrm_add_state+0x47b/0xaa0 [mlx5_core] > > [ 101.069312] bond_change_active_slave+0x392/0x900 [bonding] > > [ 101.069868] bond_option_active_slave_set+0x1c2/0x240 [bonding] > > [ 101.070454] __bond_opt_set+0xa6/0x430 [bonding] > > [ 101.070935] __bond_opt_set_notify+0x2f/0x90 [bonding] > > [ 101.071453] bond_opt_tryset_rtnl+0x72/0xb0 [bonding] > > [ 101.071965] bonding_sysfs_store_option+0x4d/0x80 [bonding] > > [ 101.072567] kernfs_fop_write_iter+0x10c/0x1a0 > > [ 101.073033] vfs_write+0x2d8/0x400 > > [ 101.073416] ? alloc_fd+0x48/0x180 > > [ 101.073798] ksys_write+0x5f/0xe0 > > [ 101.074175] do_syscall_64+0x52/0x110 > > [ 101.074576] entry_SYSCALL_64_after_hwframe+0x4b/0x53 > > > > As bond_ipsec_add_sa_all and bond_ipsec_del_sa_all are only called > > from bond_change_active_slave, which requires holding the RTNL > > lock. > > And bond_ipsec_add_sa and bond_ipsec_del_sa are xfrm state > > xdo_dev_state_add and xdo_dev_state_delete APIs, which are in user > > context. So ipsec_lock doesn't have to be spin lock, change it to > > mutex, and thus the above issue can be resolved. > > > > Fixes: 9a5605505d9c ("bonding: Add struct bond_ipesc to manage SA") > > Signed-off-by: Jianbo Liu <jianbol@nvidia.com> > > Signed-off-by: Tariq Toukan <tariqt@nvidia.com> > > Reviewed-by: Hangbin Liu <liuhangbin@gmail.com> > > --- > > drivers/net/bonding/bond_main.c | 67 +++++++++++++++--------------- > > --- > > include/net/bonding.h | 2 +- > > 2 files changed, 32 insertions(+), 37 deletions(-) > > > > diff --git a/drivers/net/bonding/bond_main.c > > b/drivers/net/bonding/bond_main.c > > index 0d1129eaf47b..f20f6d83ad54 100644 > > --- a/drivers/net/bonding/bond_main.c > > +++ b/drivers/net/bonding/bond_main.c > > @@ -439,38 +439,33 @@ static int bond_ipsec_add_sa(struct > > xfrm_state *xs, > > rcu_read_lock(); > > bond = netdev_priv(bond_dev); > > slave = rcu_dereference(bond->curr_active_slave); > > - if (!slave) { > > - rcu_read_unlock(); > > + real_dev = slave ? slave->dev : NULL; > > + rcu_read_unlock(); > > + if (!real_dev) > > return -ENODEV; > > In reading these, I was confused as to why some changes use > rcu_read_lock(), rcu_dereference() and others use rtnl_dereference(); > I > think it's because bond_ipsec_{add,del}_sa_all() are guaranteed to be > called under RTNL, while the bond_ipsec_{add,del}_sa() functions are > do > not have that guarantee. Am I understanding correctly? > Right. bond_ipsec_{add,del}_sa_all() are called by bond_change_active_slave() which has ASSERT_RTNL(), so I think they are under RTNL. > > - } > > > > - real_dev = slave->dev; > > if (!real_dev->xfrmdev_ops || > > !real_dev->xfrmdev_ops->xdo_dev_state_add || > > netif_is_bond_master(real_dev)) { > > NL_SET_ERR_MSG_MOD(extack, "Slave does not support > > ipsec offload"); > > - rcu_read_unlock(); > > return -EINVAL; > > } > > > > - ipsec = kmalloc(sizeof(*ipsec), GFP_ATOMIC); > > - if (!ipsec) { > > - rcu_read_unlock(); > > + ipsec = kmalloc(sizeof(*ipsec), GFP_KERNEL); > > + if (!ipsec) > > return -ENOMEM; > > Presumably the switch from ATOMIC to KERNEL is safe because > this > is only called under RTNL (and therefore always has a process > context), > i.e., this change is independent of any other changes in the patch. > Correct? > No. And it's RCU here, not RTNL. We are safe to use KERNEL after it's out of the RCU context, right? And this was suggested by Paolo after he reviewd the first version. > > - } > > > > xs->xso.real_dev = real_dev; > > err = real_dev->xfrmdev_ops->xdo_dev_state_add(xs, extack); > > if (!err) { > > ipsec->xs = xs; > > INIT_LIST_HEAD(&ipsec->list); > > - spin_lock_bh(&bond->ipsec_lock); > > + mutex_lock(&bond->ipsec_lock); > > list_add(&ipsec->list, &bond->ipsec_list); > > - spin_unlock_bh(&bond->ipsec_lock); > > + mutex_unlock(&bond->ipsec_lock); > > } else { > > kfree(ipsec); > > } > > - rcu_read_unlock(); > > return err; > > } > > > > @@ -481,35 +476,35 @@ static void bond_ipsec_add_sa_all(struct > > bonding *bond) > > struct bond_ipsec *ipsec; > > struct slave *slave; > > > > - rcu_read_lock(); > > - slave = rcu_dereference(bond->curr_active_slave); > > - if (!slave) > > - goto out; > > + slave = rtnl_dereference(bond->curr_active_slave); > > + real_dev = slave ? slave->dev : NULL; > > + if (!real_dev) > > + return; > > > > - real_dev = slave->dev; > > + mutex_lock(&bond->ipsec_lock); > > if (!real_dev->xfrmdev_ops || > > !real_dev->xfrmdev_ops->xdo_dev_state_add || > > netif_is_bond_master(real_dev)) { > > - spin_lock_bh(&bond->ipsec_lock); > > if (!list_empty(&bond->ipsec_list)) > > slave_warn(bond_dev, real_dev, > > "%s: no slave > > xdo_dev_state_add\n", > > __func__); > > - spin_unlock_bh(&bond->ipsec_lock); > > goto out; > > } > > > > - spin_lock_bh(&bond->ipsec_lock); > > list_for_each_entry(ipsec, &bond->ipsec_list, list) { > > + /* If new state is added before ipsec_lock acquired > > */ > > + if (ipsec->xs->xso.real_dev == real_dev) > > + continue; > > + > > ipsec->xs->xso.real_dev = real_dev; > > if (real_dev->xfrmdev_ops->xdo_dev_state_add(ipsec- > > >xs, NULL)) { > > slave_warn(bond_dev, real_dev, "%s: failed > > to add SA\n", __func__); > > ipsec->xs->xso.real_dev = NULL; > > } > > } > > - spin_unlock_bh(&bond->ipsec_lock); > > out: > > - rcu_read_unlock(); > > + mutex_unlock(&bond->ipsec_lock); > > } > > > > /** > > @@ -530,6 +525,8 @@ static void bond_ipsec_del_sa(struct xfrm_state > > *xs) > > rcu_read_lock(); > > bond = netdev_priv(bond_dev); > > slave = rcu_dereference(bond->curr_active_slave); > > + real_dev = slave ? slave->dev : NULL; > > + rcu_read_unlock(); > > Is it really safe to access real_dev once we've left the rcu > critical section? What prevents the device referenced by real_dev > from > being deleted as soon as rcu_read_unlock() completes? > I am not sure. But RCU protects accessing of the context pointed by curr_active_slave, not slave->dev itself. I wrong about this? I can move rcu_read_unlock after xdo_dev_state_delete(). And do the same change for bond_ipsec_add_sa and bond_ipsec_free_sa. What do you think? Thanks! Jianbo > -J > > > > > if (!slave) > > goto out; > > @@ -537,7 +534,6 @@ static void bond_ipsec_del_sa(struct xfrm_state > > *xs) > > if (!xs->xso.real_dev) > > goto out; > > > > - real_dev = slave->dev; > > WARN_ON(xs->xso.real_dev != real_dev); > > > > if (!real_dev->xfrmdev_ops || > > @@ -549,7 +545,7 @@ static void bond_ipsec_del_sa(struct xfrm_state > > *xs) > > > > real_dev->xfrmdev_ops->xdo_dev_state_delete(xs); > > out: > > - spin_lock_bh(&bond->ipsec_lock); > > + mutex_lock(&bond->ipsec_lock); > > list_for_each_entry(ipsec, &bond->ipsec_list, list) { > > if (ipsec->xs == xs) { > > list_del(&ipsec->list); > > @@ -557,8 +553,7 @@ static void bond_ipsec_del_sa(struct xfrm_state > > *xs) > > break; > > } > > } > > - spin_unlock_bh(&bond->ipsec_lock); > > - rcu_read_unlock(); > > + mutex_unlock(&bond->ipsec_lock); > > } > > > > static void bond_ipsec_del_sa_all(struct bonding *bond) > > @@ -568,15 +563,12 @@ static void bond_ipsec_del_sa_all(struct > > bonding *bond) > > struct bond_ipsec *ipsec; > > struct slave *slave; > > > > - rcu_read_lock(); > > - slave = rcu_dereference(bond->curr_active_slave); > > - if (!slave) { > > - rcu_read_unlock(); > > + slave = rtnl_dereference(bond->curr_active_slave); > > + real_dev = slave ? slave->dev : NULL; > > + if (!real_dev) > > return; > > - } > > > > - real_dev = slave->dev; > > - spin_lock_bh(&bond->ipsec_lock); > > + mutex_lock(&bond->ipsec_lock); > > list_for_each_entry(ipsec, &bond->ipsec_list, list) { > > if (!ipsec->xs->xso.real_dev) > > continue; > > @@ -593,8 +585,7 @@ static void bond_ipsec_del_sa_all(struct > > bonding *bond) > > real_dev->xfrmdev_ops- > > >xdo_dev_state_free(ipsec->xs); > > } > > } > > - spin_unlock_bh(&bond->ipsec_lock); > > - rcu_read_unlock(); > > + mutex_unlock(&bond->ipsec_lock); > > } > > > > static void bond_ipsec_free_sa(struct xfrm_state *xs) > > @@ -5917,7 +5908,7 @@ void bond_setup(struct net_device *bond_dev) > > /* set up xfrm device ops (only supported in active-backup > > right now) */ > > bond_dev->xfrmdev_ops = &bond_xfrmdev_ops; > > INIT_LIST_HEAD(&bond->ipsec_list); > > - spin_lock_init(&bond->ipsec_lock); > > + mutex_init(&bond->ipsec_lock); > > #endif /* CONFIG_XFRM_OFFLOAD */ > > > > /* don't acquire bond device's netif_tx_lock when > > transmitting */ > > @@ -5966,6 +5957,10 @@ static void bond_uninit(struct net_device > > *bond_dev) > > __bond_release_one(bond_dev, slave->dev, true, > > true); > > netdev_info(bond_dev, "Released all slaves\n"); > > > > +#ifdef CONFIG_XFRM_OFFLOAD > > + mutex_destroy(&bond->ipsec_lock); > > +#endif /* CONFIG_XFRM_OFFLOAD */ > > + > > bond_set_slave_arr(bond, NULL, NULL); > > > > list_del_rcu(&bond->bond_list); > > diff --git a/include/net/bonding.h b/include/net/bonding.h > > index b61fb1aa3a56..8bb5f016969f 100644 > > --- a/include/net/bonding.h > > +++ b/include/net/bonding.h > > @@ -260,7 +260,7 @@ struct bonding { > > #ifdef CONFIG_XFRM_OFFLOAD > > struct list_head ipsec_list; > > /* protecting ipsec_list */ > > - spinlock_t ipsec_lock; > > + struct mutex ipsec_lock; > > #endif /* CONFIG_XFRM_OFFLOAD */ > > struct bpf_prog *xdp_prog; > > }; > > -- > > 2.21.0 > > > > --- > -Jay Vosburgh, jv@jvosburgh.net
On Wed, 2024-08-21 at 17:11 -0700, Jakub Kicinski wrote: > On Wed, 21 Aug 2024 09:00:30 -0700 Jay Vosburgh wrote: > > Is it really safe to access real_dev once we've left the > > rcu > > critical section? What prevents the device referenced by real_dev > > from > > being deleted as soon as rcu_read_unlock() completes? > > Hah, I asked them this question at least 2 times. > Let's see if your superior communication skills help :) Sorry, I maybe misunderstood. I really need your suggestion about how to change it. Always waiting for you, hangbin or anyother's confirm before do any update. V5 is the rebase. If you and Jay want to move rcu_read_unlock after xfrm callbacks, I will send new version soon. Thanks! Jianbo
Jianbo Liu <jianbol@nvidia.com> wrote: >On Wed, 2024-08-21 at 09:00 -0700, Jay Vosburgh wrote: >> Jianbo Liu <jianbol@nvidia.com> wrote: >> >> > In the cited commit, bond->ipsec_lock is added to protect >> > ipsec_list, >> > hence xdo_dev_state_add and xdo_dev_state_delete are called inside >> > this lock. As ipsec_lock is a spin lock and such xfrmdev ops may >> > sleep, >> > "scheduling while atomic" will be triggered when changing bond's >> > active slave. >> > >> > [ 101.055189] BUG: scheduling while atomic: bash/902/0x00000200 >> > [ 101.055726] Modules linked in: >> > [ 101.058211] CPU: 3 PID: 902 Comm: bash Not tainted 6.9.0-rc4+ #1 >> > [ 101.058760] Hardware name: >> > [ 101.059434] Call Trace: >> > [ 101.059436] <TASK> >> > [ 101.060873] dump_stack_lvl+0x51/0x60 >> > [ 101.061275] __schedule_bug+0x4e/0x60 >> > [ 101.061682] __schedule+0x612/0x7c0 >> > [ 101.062078] ? __mod_timer+0x25c/0x370 >> > [ 101.062486] schedule+0x25/0xd0 >> > [ 101.062845] schedule_timeout+0x77/0xf0 >> > [ 101.063265] ? asm_common_interrupt+0x22/0x40 >> > [ 101.063724] ? __bpf_trace_itimer_state+0x10/0x10 >> > [ 101.064215] __wait_for_common+0x87/0x190 >> > [ 101.064648] ? usleep_range_state+0x90/0x90 >> > [ 101.065091] cmd_exec+0x437/0xb20 [mlx5_core] >> > [ 101.065569] mlx5_cmd_do+0x1e/0x40 [mlx5_core] >> > [ 101.066051] mlx5_cmd_exec+0x18/0x30 [mlx5_core] >> > [ 101.066552] mlx5_crypto_create_dek_key+0xea/0x120 [mlx5_core] >> > [ 101.067163] ? bonding_sysfs_store_option+0x4d/0x80 [bonding] >> > [ 101.067738] ? kmalloc_trace+0x4d/0x350 >> > [ 101.068156] mlx5_ipsec_create_sa_ctx+0x33/0x100 [mlx5_core] >> > [ 101.068747] mlx5e_xfrm_add_state+0x47b/0xaa0 [mlx5_core] >> > [ 101.069312] bond_change_active_slave+0x392/0x900 [bonding] >> > [ 101.069868] bond_option_active_slave_set+0x1c2/0x240 [bonding] >> > [ 101.070454] __bond_opt_set+0xa6/0x430 [bonding] >> > [ 101.070935] __bond_opt_set_notify+0x2f/0x90 [bonding] >> > [ 101.071453] bond_opt_tryset_rtnl+0x72/0xb0 [bonding] >> > [ 101.071965] bonding_sysfs_store_option+0x4d/0x80 [bonding] >> > [ 101.072567] kernfs_fop_write_iter+0x10c/0x1a0 >> > [ 101.073033] vfs_write+0x2d8/0x400 >> > [ 101.073416] ? alloc_fd+0x48/0x180 >> > [ 101.073798] ksys_write+0x5f/0xe0 >> > [ 101.074175] do_syscall_64+0x52/0x110 >> > [ 101.074576] entry_SYSCALL_64_after_hwframe+0x4b/0x53 >> > >> > As bond_ipsec_add_sa_all and bond_ipsec_del_sa_all are only called >> > from bond_change_active_slave, which requires holding the RTNL >> > lock. >> > And bond_ipsec_add_sa and bond_ipsec_del_sa are xfrm state >> > xdo_dev_state_add and xdo_dev_state_delete APIs, which are in user >> > context. So ipsec_lock doesn't have to be spin lock, change it to >> > mutex, and thus the above issue can be resolved. >> > >> > Fixes: 9a5605505d9c ("bonding: Add struct bond_ipesc to manage SA") >> > Signed-off-by: Jianbo Liu <jianbol@nvidia.com> >> > Signed-off-by: Tariq Toukan <tariqt@nvidia.com> >> > Reviewed-by: Hangbin Liu <liuhangbin@gmail.com> >> > --- >> > drivers/net/bonding/bond_main.c | 67 +++++++++++++++--------------- >> > --- >> > include/net/bonding.h | 2 +- >> > 2 files changed, 32 insertions(+), 37 deletions(-) >> > >> > diff --git a/drivers/net/bonding/bond_main.c >> > b/drivers/net/bonding/bond_main.c >> > index 0d1129eaf47b..f20f6d83ad54 100644 >> > --- a/drivers/net/bonding/bond_main.c >> > +++ b/drivers/net/bonding/bond_main.c >> > @@ -439,38 +439,33 @@ static int bond_ipsec_add_sa(struct >> > xfrm_state *xs, >> > rcu_read_lock(); >> > bond = netdev_priv(bond_dev); >> > slave = rcu_dereference(bond->curr_active_slave); >> > - if (!slave) { >> > - rcu_read_unlock(); >> > + real_dev = slave ? slave->dev : NULL; >> > + rcu_read_unlock(); >> > + if (!real_dev) >> > return -ENODEV; >> >> In reading these, I was confused as to why some changes use >> rcu_read_lock(), rcu_dereference() and others use rtnl_dereference(); >> I >> think it's because bond_ipsec_{add,del}_sa_all() are guaranteed to be >> called under RTNL, while the bond_ipsec_{add,del}_sa() functions are >> do >> not have that guarantee. Am I understanding correctly? >> > >Right. bond_ipsec_{add,del}_sa_all() are called by >bond_change_active_slave() which has ASSERT_RTNL(), so I think they are >under RTNL. Yes. >> > - } >> > >> > - real_dev = slave->dev; >> > if (!real_dev->xfrmdev_ops || >> > !real_dev->xfrmdev_ops->xdo_dev_state_add || >> > netif_is_bond_master(real_dev)) { >> > NL_SET_ERR_MSG_MOD(extack, "Slave does not support >> > ipsec offload"); >> > - rcu_read_unlock(); >> > return -EINVAL; >> > } >> > >> > - ipsec = kmalloc(sizeof(*ipsec), GFP_ATOMIC); >> > - if (!ipsec) { >> > - rcu_read_unlock(); >> > + ipsec = kmalloc(sizeof(*ipsec), GFP_KERNEL); >> > + if (!ipsec) >> > return -ENOMEM; >> >> Presumably the switch from ATOMIC to KERNEL is safe because >> this >> is only called under RTNL (and therefore always has a process >> context), >> i.e., this change is independent of any other changes in the patch. >> Correct? >> > >No. And it's RCU here, not RTNL. We are safe to use KERNEL after it's >out of the RCU context, right? And this was suggested by Paolo after he >reviewd the first version. Ok, I think I follow now. And, yes, KERNEL is ok when outside the RCU critical section, but not inside of it (because sleeping is not permitted within the critical section). >> > - } >> > >> > xs->xso.real_dev = real_dev; >> > err = real_dev->xfrmdev_ops->xdo_dev_state_add(xs, extack); >> > if (!err) { >> > ipsec->xs = xs; >> > INIT_LIST_HEAD(&ipsec->list); >> > - spin_lock_bh(&bond->ipsec_lock); >> > + mutex_lock(&bond->ipsec_lock); >> > list_add(&ipsec->list, &bond->ipsec_list); >> > - spin_unlock_bh(&bond->ipsec_lock); >> > + mutex_unlock(&bond->ipsec_lock); >> > } else { >> > kfree(ipsec); >> > } >> > - rcu_read_unlock(); >> > return err; >> > } >> > >> > @@ -481,35 +476,35 @@ static void bond_ipsec_add_sa_all(struct >> > bonding *bond) >> > struct bond_ipsec *ipsec; >> > struct slave *slave; >> > >> > - rcu_read_lock(); >> > - slave = rcu_dereference(bond->curr_active_slave); >> > - if (!slave) >> > - goto out; >> > + slave = rtnl_dereference(bond->curr_active_slave); >> > + real_dev = slave ? slave->dev : NULL; >> > + if (!real_dev) >> > + return; >> > >> > - real_dev = slave->dev; >> > + mutex_lock(&bond->ipsec_lock); >> > if (!real_dev->xfrmdev_ops || >> > !real_dev->xfrmdev_ops->xdo_dev_state_add || >> > netif_is_bond_master(real_dev)) { >> > - spin_lock_bh(&bond->ipsec_lock); >> > if (!list_empty(&bond->ipsec_list)) >> > slave_warn(bond_dev, real_dev, >> > "%s: no slave >> > xdo_dev_state_add\n", >> > __func__); >> > - spin_unlock_bh(&bond->ipsec_lock); >> > goto out; >> > } >> > >> > - spin_lock_bh(&bond->ipsec_lock); >> > list_for_each_entry(ipsec, &bond->ipsec_list, list) { >> > + /* If new state is added before ipsec_lock acquired >> > */ >> > + if (ipsec->xs->xso.real_dev == real_dev) >> > + continue; >> > + >> > ipsec->xs->xso.real_dev = real_dev; >> > if (real_dev->xfrmdev_ops->xdo_dev_state_add(ipsec- >> > >xs, NULL)) { >> > slave_warn(bond_dev, real_dev, "%s: failed >> > to add SA\n", __func__); >> > ipsec->xs->xso.real_dev = NULL; >> > } >> > } >> > - spin_unlock_bh(&bond->ipsec_lock); >> > out: >> > - rcu_read_unlock(); >> > + mutex_unlock(&bond->ipsec_lock); >> > } >> > >> > /** >> > @@ -530,6 +525,8 @@ static void bond_ipsec_del_sa(struct xfrm_state >> > *xs) >> > rcu_read_lock(); >> > bond = netdev_priv(bond_dev); >> > slave = rcu_dereference(bond->curr_active_slave); >> > + real_dev = slave ? slave->dev : NULL; >> > + rcu_read_unlock(); >> >> Is it really safe to access real_dev once we've left the rcu >> critical section? What prevents the device referenced by real_dev >> from >> being deleted as soon as rcu_read_unlock() completes? >> > >I am not sure. But RCU protects accessing of the context pointed by >curr_active_slave, not slave->dev itself. I wrong about this? No, you're not wrong: RCU does indeed protect the "slave" pointer in the above code while inside the RCU read-side critical section. However, we also know that as long as we're within that critical section, whatever the "slave" pointer points to will remain valid, because what curr_active_slave points to is also RCU protected (not only the pointer itself). For the interface behind slave->dev specifically, any attempt to delete an interface that is a member of a bond must pass through __bond_release_one() first, and it calls synchronize_rcu() as part of its processing (which will wait for active read-side critical sections to complete). Therefore, the bond member interface behind slave->dev here cannot simply vanish while execution is within this critical section. >I can move rcu_read_unlock after xdo_dev_state_delete(). And do the >same change for bond_ipsec_add_sa and bond_ipsec_free_sa. >What do you think? The original issue was that the xfrm callback within mlx5 could sleep while a spin lock was held. However, sleeping is not permitted within an RCU read-side critical section, either, so would this simply reintroduce the original problem from a different angle? Assuming that's correct, I think one way around that is to acquire a reference (via dev_hold or netdev_hold) to the interface (i.e., real_dev) within the minimal rcu_read_lock / rcu_read_unlock, do the xfrm magic, and then release the reference when finished. That won't prevent the interface from being removed from the bond and the struct slave being freed outside of the RCU critical section, so the code would also need to use only real_dev after rcu_read_unlock() is called. -J >Thanks! >Jianbo > >> -J >> >> > >> > if (!slave) >> > goto out; >> > @@ -537,7 +534,6 @@ static void bond_ipsec_del_sa(struct xfrm_state >> > *xs) >> > if (!xs->xso.real_dev) >> > goto out; >> > >> > - real_dev = slave->dev; >> > WARN_ON(xs->xso.real_dev != real_dev); >> > >> > if (!real_dev->xfrmdev_ops || >> > @@ -549,7 +545,7 @@ static void bond_ipsec_del_sa(struct xfrm_state >> > *xs) >> > >> > real_dev->xfrmdev_ops->xdo_dev_state_delete(xs); >> > out: >> > - spin_lock_bh(&bond->ipsec_lock); >> > + mutex_lock(&bond->ipsec_lock); >> > list_for_each_entry(ipsec, &bond->ipsec_list, list) { >> > if (ipsec->xs == xs) { >> > list_del(&ipsec->list); >> > @@ -557,8 +553,7 @@ static void bond_ipsec_del_sa(struct xfrm_state >> > *xs) >> > break; >> > } >> > } >> > - spin_unlock_bh(&bond->ipsec_lock); >> > - rcu_read_unlock(); >> > + mutex_unlock(&bond->ipsec_lock); >> > } >> > >> > static void bond_ipsec_del_sa_all(struct bonding *bond) >> > @@ -568,15 +563,12 @@ static void bond_ipsec_del_sa_all(struct >> > bonding *bond) >> > struct bond_ipsec *ipsec; >> > struct slave *slave; >> > >> > - rcu_read_lock(); >> > - slave = rcu_dereference(bond->curr_active_slave); >> > - if (!slave) { >> > - rcu_read_unlock(); >> > + slave = rtnl_dereference(bond->curr_active_slave); >> > + real_dev = slave ? slave->dev : NULL; >> > + if (!real_dev) >> > return; >> > - } >> > >> > - real_dev = slave->dev; >> > - spin_lock_bh(&bond->ipsec_lock); >> > + mutex_lock(&bond->ipsec_lock); >> > list_for_each_entry(ipsec, &bond->ipsec_list, list) { >> > if (!ipsec->xs->xso.real_dev) >> > continue; >> > @@ -593,8 +585,7 @@ static void bond_ipsec_del_sa_all(struct >> > bonding *bond) >> > real_dev->xfrmdev_ops- >> > >xdo_dev_state_free(ipsec->xs); >> > } >> > } >> > - spin_unlock_bh(&bond->ipsec_lock); >> > - rcu_read_unlock(); >> > + mutex_unlock(&bond->ipsec_lock); >> > } >> > >> > static void bond_ipsec_free_sa(struct xfrm_state *xs) >> > @@ -5917,7 +5908,7 @@ void bond_setup(struct net_device *bond_dev) >> > /* set up xfrm device ops (only supported in active-backup >> > right now) */ >> > bond_dev->xfrmdev_ops = &bond_xfrmdev_ops; >> > INIT_LIST_HEAD(&bond->ipsec_list); >> > - spin_lock_init(&bond->ipsec_lock); >> > + mutex_init(&bond->ipsec_lock); >> > #endif /* CONFIG_XFRM_OFFLOAD */ >> > >> > /* don't acquire bond device's netif_tx_lock when >> > transmitting */ >> > @@ -5966,6 +5957,10 @@ static void bond_uninit(struct net_device >> > *bond_dev) >> > __bond_release_one(bond_dev, slave->dev, true, >> > true); >> > netdev_info(bond_dev, "Released all slaves\n"); >> > >> > +#ifdef CONFIG_XFRM_OFFLOAD >> > + mutex_destroy(&bond->ipsec_lock); >> > +#endif /* CONFIG_XFRM_OFFLOAD */ >> > + >> > bond_set_slave_arr(bond, NULL, NULL); >> > >> > list_del_rcu(&bond->bond_list); >> > diff --git a/include/net/bonding.h b/include/net/bonding.h >> > index b61fb1aa3a56..8bb5f016969f 100644 >> > --- a/include/net/bonding.h >> > +++ b/include/net/bonding.h >> > @@ -260,7 +260,7 @@ struct bonding { >> > #ifdef CONFIG_XFRM_OFFLOAD >> > struct list_head ipsec_list; >> > /* protecting ipsec_list */ >> > - spinlock_t ipsec_lock; >> > + struct mutex ipsec_lock; >> > #endif /* CONFIG_XFRM_OFFLOAD */ >> > struct bpf_prog *xdp_prog; >> > }; >> > -- >> > 2.21.0 --- -Jay Vosburgh, jv@jvosburgh.net
On 8/22/2024 2:05 PM, Jay Vosburgh wrote: > Jianbo Liu <jianbol@nvidia.com> wrote: > >> On Wed, 2024-08-21 at 09:00 -0700, Jay Vosburgh wrote: >>> Jianbo Liu <jianbol@nvidia.com> wrote: >>> >>>> In the cited commit, bond->ipsec_lock is added to protect >>>> ipsec_list, >>>> hence xdo_dev_state_add and xdo_dev_state_delete are called inside >>>> this lock. As ipsec_lock is a spin lock and such xfrmdev ops may >>>> sleep, >>>> "scheduling while atomic" will be triggered when changing bond's >>>> active slave. >>>> >>>> [ 101.055189] BUG: scheduling while atomic: bash/902/0x00000200 >>>> [ 101.055726] Modules linked in: >>>> [ 101.058211] CPU: 3 PID: 902 Comm: bash Not tainted 6.9.0-rc4+ #1 >>>> [ 101.058760] Hardware name: >>>> [ 101.059434] Call Trace: >>>> [ 101.059436] <TASK> >>>> [ 101.060873] dump_stack_lvl+0x51/0x60 >>>> [ 101.061275] __schedule_bug+0x4e/0x60 >>>> [ 101.061682] __schedule+0x612/0x7c0 >>>> [ 101.062078] ? __mod_timer+0x25c/0x370 > >> > [ 101.062486] schedule+0x25/0xd0 >>>> [ 101.062845] schedule_timeout+0x77/0xf0 >>>> [ 101.063265] ? asm_common_interrupt+0x22/0x40 >>>> [ 101.063724] ? __bpf_trace_itimer_state+0x10/0x10 >>>> [ 101.064215] __wait_for_common+0x87/0x190 >>>> [ 101.064648] ? usleep_range_state+0x90/0x90 >>>> [ 101.065091] cmd_exec+0x437/0xb20 [mlx5_core] >>>> [ 101.065569] mlx5_cmd_do+0x1e/0x40 [mlx5_core] >>>> [ 101.066051] mlx5_cmd_exec+0x18/0x30 [mlx5_core] >>>> [ 101.066552] mlx5_crypto_create_dek_key+0xea/0x120 [mlx5_core] >>>> [ 101.067163] ? bonding_sysfs_store_option+0x4d/0x80 [bonding] >>>> [ 101.067738] ? kmalloc_trace+0x4d/0x350 >>>> [ 101.068156] mlx5_ipsec_create_sa_ctx+0x33/0x100 [mlx5_core] >>>> [ 101.068747] mlx5e_xfrm_add_state+0x47b/0xaa0 [mlx5_core] >>>> [ 101.069312] bond_change_active_slave+0x392/0x900 [bonding] >>>> [ 101.069868] bond_option_active_slave_set+0x1c2/0x240 [bonding] >>>> [ 101.070454] __bond_opt_set+0xa6/0x430 [bonding] >>>> [ 101.070935] __bond_opt_set_notify+0x2f/0x90 [bonding] >>>> [ 101.071453] bond_opt_tryset_rtnl+0x72/0xb0 [bonding] >>>> [ 101.071965] bonding_sysfs_store_option+0x4d/0x80 [bonding] >>>> [ 101.072567] kernfs_fop_write_iter+0x10c/0x1a0 >>>> [ 101.073033] vfs_write+0x2d8/0x400 >>>> [ 101.073416] ? alloc_fd+0x48/0x180 >>>> [ 101.073798] ksys_write+0x5f/0xe0 >>>> [ 101.074175] do_syscall_64+0x52/0x110 >>>> [ 101.074576] entry_SYSCALL_64_after_hwframe+0x4b/0x53 >>>> >>>> As bond_ipsec_add_sa_all and bond_ipsec_del_sa_all are only called >>>> from bond_change_active_slave, which requires holding the RTNL >>>> lock. >>>> And bond_ipsec_add_sa and bond_ipsec_del_sa are xfrm state >>>> xdo_dev_state_add and xdo_dev_state_delete APIs, which are in user >>>> context. So ipsec_lock doesn't have to be spin lock, change it to >>>> mutex, and thus the above issue can be resolved. >>>> >>>> Fixes: 9a5605505d9c ("bonding: Add struct bond_ipesc to manage SA") >>>> Signed-off-by: Jianbo Liu <jianbol@nvidia.com> >>>> Signed-off-by: Tariq Toukan <tariqt@nvidia.com> >>>> Reviewed-by: Hangbin Liu <liuhangbin@gmail.com> >>>> --- >>>> drivers/net/bonding/bond_main.c | 67 +++++++++++++++--------------- >>>> --- >>>> include/net/bonding.h | 2 +- >>>> 2 files changed, 32 insertions(+), 37 deletions(-) >>>> >>>> diff --git a/drivers/net/bonding/bond_main.c >>>> b/drivers/net/bonding/bond_main.c >>>> index 0d1129eaf47b..f20f6d83ad54 100644 >>>> --- a/drivers/net/bonding/bond_main.c >>>> +++ b/drivers/net/bonding/bond_main.c >>>> @@ -439,38 +439,33 @@ static int bond_ipsec_add_sa(struct >>>> xfrm_state *xs, >>>> rcu_read_lock(); >>>> bond = netdev_priv(bond_dev); >>>> slave = rcu_dereference(bond->curr_active_slave); >>>> - if (!slave) { >>>> - rcu_read_unlock(); >>>> + real_dev = slave ? slave->dev : NULL; >>>> + rcu_read_unlock(); >>>> + if (!real_dev) >>>> return -ENODEV; >>> >>> In reading these, I was confused as to why some changes use >>> rcu_read_lock(), rcu_dereference() and others use rtnl_dereference(); >>> I >>> think it's because bond_ipsec_{add,del}_sa_all() are guaranteed to be >>> called under RTNL, while the bond_ipsec_{add,del}_sa() functions are >>> do >>> not have that guarantee. Am I understanding correctly? >>> >> >> Right. bond_ipsec_{add,del}_sa_all() are called by >> bond_change_active_slave() which has ASSERT_RTNL(), so I think they are >> under RTNL. > > Yes. > >>>> - } >>>> >>>> - real_dev = slave->dev; >>>> if (!real_dev->xfrmdev_ops || >>>> !real_dev->xfrmdev_ops->xdo_dev_state_add || >>>> netif_is_bond_master(real_dev)) { >>>> NL_SET_ERR_MSG_MOD(extack, "Slave does not support >>>> ipsec offload"); >>>> - rcu_read_unlock(); >>>> return -EINVAL; >>>> } >>>> >>>> - ipsec = kmalloc(sizeof(*ipsec), GFP_ATOMIC); >>>> - if (!ipsec) { >>>> - rcu_read_unlock(); >>>> + ipsec = kmalloc(sizeof(*ipsec), GFP_KERNEL); >>>> + if (!ipsec) >>>> return -ENOMEM; >>> >>> Presumably the switch from ATOMIC to KERNEL is safe because >>> this >>> is only called under RTNL (and therefore always has a process >>> context), >>> i.e., this change is independent of any other changes in the patch. >>> Correct? >>> >> >> No. And it's RCU here, not RTNL. We are safe to use KERNEL after it's >> out of the RCU context, right? And this was suggested by Paolo after he >> reviewd the first version. > > Ok, I think I follow now. And, yes, KERNEL is ok when outside > the RCU critical section, but not inside of it (because sleeping is not > permitted within the critical section). > Yes. >>>> - } >>>> >>>> xs->xso.real_dev = real_dev; >>>> err = real_dev->xfrmdev_ops->xdo_dev_state_add(xs, extack); >>>> if (!err) { >>>> ipsec->xs = xs; >>>> INIT_LIST_HEAD(&ipsec->list); >>>> - spin_lock_bh(&bond->ipsec_lock); >>>> + mutex_lock(&bond->ipsec_lock); >>>> list_add(&ipsec->list, &bond->ipsec_list); >>>> - spin_unlock_bh(&bond->ipsec_lock); >>>> + mutex_unlock(&bond->ipsec_lock); >>>> } else { >>>> kfree(ipsec); >>>> } >>>> - rcu_read_unlock(); >>>> return err; >>>> } >>>> >>>> @@ -481,35 +476,35 @@ static void bond_ipsec_add_sa_all(struct >>>> bonding *bond) >>>> struct bond_ipsec *ipsec; >>>> struct slave *slave; >>>> >>>> - rcu_read_lock(); >>>> - slave = rcu_dereference(bond->curr_active_slave); >>>> - if (!slave) >>>> - goto out; >>>> + slave = rtnl_dereference(bond->curr_active_slave); >>>> + real_dev = slave ? slave->dev : NULL; >>>> + if (!real_dev) >>>> + return; >>>> >>>> - real_dev = slave->dev; >>>> + mutex_lock(&bond->ipsec_lock); >>>> if (!real_dev->xfrmdev_ops || >>>> !real_dev->xfrmdev_ops->xdo_dev_state_add || >>>> netif_is_bond_master(real_dev)) { >>>> - spin_lock_bh(&bond->ipsec_lock); >>>> if (!list_empty(&bond->ipsec_list)) >>>> slave_warn(bond_dev, real_dev, >>>> "%s: no slave >>>> xdo_dev_state_add\n", >>>> __func__); >>>> - spin_unlock_bh(&bond->ipsec_lock); >>>> goto out; >>>> } >>>> >>>> - spin_lock_bh(&bond->ipsec_lock); >>>> list_for_each_entry(ipsec, &bond->ipsec_list, list) { >>>> + /* If new state is added before ipsec_lock acquired >>>> */ >>>> + if (ipsec->xs->xso.real_dev == real_dev) >>>> + continue; >>>> + >>>> ipsec->xs->xso.real_dev = real_dev; >>>> if (real_dev->xfrmdev_ops->xdo_dev_state_add(ipsec- >>>>> xs, NULL)) { >>>> slave_warn(bond_dev, real_dev, "%s: failed >>>> to add SA\n", __func__); >>>> ipsec->xs->xso.real_dev = NULL; >>>> } >>>> } >>>> - spin_unlock_bh(&bond->ipsec_lock); >>>> out: >>>> - rcu_read_unlock(); >>>> + mutex_unlock(&bond->ipsec_lock); >>>> } >>>> >>>> /** >>>> @@ -530,6 +525,8 @@ static void bond_ipsec_del_sa(struct xfrm_state >>>> *xs) >>>> rcu_read_lock(); >>>> bond = netdev_priv(bond_dev); >>>> slave = rcu_dereference(bond->curr_active_slave); >>>> + real_dev = slave ? slave->dev : NULL; >>>> + rcu_read_unlock(); >>> >>> Is it really safe to access real_dev once we've left the rcu >>> critical section? What prevents the device referenced by real_dev >>> from >>> being deleted as soon as rcu_read_unlock() completes? >>> >> >> I am not sure. But RCU protects accessing of the context pointed by >> curr_active_slave, not slave->dev itself. I wrong about this? > > No, you're not wrong: RCU does indeed protect the "slave" > pointer in the above code while inside the RCU read-side critical > section. > > However, we also know that as long as we're within that critical > section, whatever the "slave" pointer points to will remain valid, > because what curr_active_slave points to is also RCU protected (not only > the pointer itself). > > For the interface behind slave->dev specifically, any attempt to > delete an interface that is a member of a bond must pass through > __bond_release_one() first, and it calls synchronize_rcu() as part of > its processing (which will wait for active read-side critical sections > to complete). Therefore, the bond member interface behind slave->dev > here cannot simply vanish while execution is within this critical > section. > Many thanks for you clarification. It helps me a lot. >> I can move rcu_read_unlock after xdo_dev_state_delete(). And do the >> same change for bond_ipsec_add_sa and bond_ipsec_free_sa. >> What do you think? > > The original issue was that the xfrm callback within mlx5 could > sleep while a spin lock was held. However, sleeping is not permitted But, can we assume no sleep in xfrm callbacks, considering drivers need to interact with hardware? It's not related to mlx5 only, but others from different vendors. > within an RCU read-side critical section, either, so would this simply > reintroduce the original problem from a different angle? > I moved the callbacks out of RCU critical section, because we never know what drivers will do. I thought bond may (should) hold the netdev, but I was wrong as you explained. > Assuming that's correct, I think one way around that is to > acquire a reference (via dev_hold or netdev_hold) to the interface > (i.e., real_dev) within the minimal rcu_read_lock / rcu_read_unlock, do That's what I did, trying to reduce critical section, unlock immediately after get real_dev. > the xfrm magic, and then release the reference when finished. That > won't prevent the interface from being removed from the bond and the > struct slave being freed outside of the RCU critical section, so the > code would also need to use only real_dev after rcu_read_unlock() is > called. > I think it's good solution. So I need to add the dev_hold/dev_put as following, for example, for bond_ipsec_del_sa, right? @@ -526,6 +534,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) bond = netdev_priv(bond_dev); slave = rcu_dereference(bond->curr_active_slave); real_dev = slave ? slave->dev : NULL; + dev_hold(real_dev); rcu_read_unlock(); if (!slave) @@ -545,6 +554,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) real_dev->xfrmdev_ops->xdo_dev_state_delete(xs); out: + dev_put(real_dev); mutex_lock(&bond->ipsec_lock); list_for_each_entry(ipsec, &bond->ipsec_list, list) { if (ipsec->xs == xs) { If you are ok with that, I will add the same for bond_ipsec_add_sa/bond_ipsec_free_sa, and send new version. Please confirm. Thanks! Jianbo > -J > >> Thanks! >> Jianbo >> >>> -J >>> >>>> >>>> if (!slave) >>>> goto out; >>>> @@ -537,7 +534,6 @@ static void bond_ipsec_del_sa(struct xfrm_state >>>> *xs) >>>> if (!xs->xso.real_dev) >>>> goto out; >>>> >>>> - real_dev = slave->dev; >>>> WARN_ON(xs->xso.real_dev != real_dev); >>>> >>>> if (!real_dev->xfrmdev_ops || >>>> @@ -549,7 +545,7 @@ static void bond_ipsec_del_sa(struct xfrm_state >>>> *xs) >>>> >>>> real_dev->xfrmdev_ops->xdo_dev_state_delete(xs); >>>> out: >>>> - spin_lock_bh(&bond->ipsec_lock); >>>> + mutex_lock(&bond->ipsec_lock); >>>> list_for_each_entry(ipsec, &bond->ipsec_list, list) { >>>> if (ipsec->xs == xs) { >>>> list_del(&ipsec->list); >>>> @@ -557,8 +553,7 @@ static void bond_ipsec_del_sa(struct xfrm_state >>>> *xs) >>>> break; >>>> } >>>> } >>>> - spin_unlock_bh(&bond->ipsec_lock); >>>> - rcu_read_unlock(); >>>> + mutex_unlock(&bond->ipsec_lock); >>>> } >>>> >>>> static void bond_ipsec_del_sa_all(struct bonding *bond) >>>> @@ -568,15 +563,12 @@ static void bond_ipsec_del_sa_all(struct >>>> bonding *bond) >>>> struct bond_ipsec *ipsec; >>>> struct slave *slave; >>>> >>>> - rcu_read_lock(); >>>> - slave = rcu_dereference(bond->curr_active_slave); >>>> - if (!slave) { >>>> - rcu_read_unlock(); >>>> + slave = rtnl_dereference(bond->curr_active_slave); >>>> + real_dev = slave ? slave->dev : NULL; >>>> + if (!real_dev) >>>> return; >>>> - } >>>> >>>> - real_dev = slave->dev; >>>> - spin_lock_bh(&bond->ipsec_lock); >>>> + mutex_lock(&bond->ipsec_lock); >>>> list_for_each_entry(ipsec, &bond->ipsec_list, list) { >>>> if (!ipsec->xs->xso.real_dev) >>>> continue; >>>> @@ -593,8 +585,7 @@ static void bond_ipsec_del_sa_all(struct >>>> bonding *bond) >>>> real_dev->xfrmdev_ops- >>>>> xdo_dev_state_free(ipsec->xs); >>>> } >>>> } >>>> - spin_unlock_bh(&bond->ipsec_lock); >>>> - rcu_read_unlock(); >>>> + mutex_unlock(&bond->ipsec_lock); >>>> } >>>> >>>> static void bond_ipsec_free_sa(struct xfrm_state *xs) >>>> @@ -5917,7 +5908,7 @@ void bond_setup(struct net_device *bond_dev) >>>> /* set up xfrm device ops (only supported in active-backup >>>> right now) */ >>>> bond_dev->xfrmdev_ops = &bond_xfrmdev_ops; >>>> INIT_LIST_HEAD(&bond->ipsec_list); >>>> - spin_lock_init(&bond->ipsec_lock); >>>> + mutex_init(&bond->ipsec_lock); >>>> #endif /* CONFIG_XFRM_OFFLOAD */ >>>> >>>> /* don't acquire bond device's netif_tx_lock when >>>> transmitting */ >>>> @@ -5966,6 +5957,10 @@ static void bond_uninit(struct net_device >>>> *bond_dev) >>>> __bond_release_one(bond_dev, slave->dev, true, >>>> true); >>>> netdev_info(bond_dev, "Released all slaves\n"); >>>> >>>> +#ifdef CONFIG_XFRM_OFFLOAD >>>> + mutex_destroy(&bond->ipsec_lock); >>>> +#endif /* CONFIG_XFRM_OFFLOAD */ >>>> + >>>> bond_set_slave_arr(bond, NULL, NULL); >>>> >>>> list_del_rcu(&bond->bond_list); >>>> diff --git a/include/net/bonding.h b/include/net/bonding.h >>>> index b61fb1aa3a56..8bb5f016969f 100644 >>>> --- a/include/net/bonding.h >>>> +++ b/include/net/bonding.h >>>> @@ -260,7 +260,7 @@ struct bonding { >>>> #ifdef CONFIG_XFRM_OFFLOAD >>>> struct list_head ipsec_list; >>>> /* protecting ipsec_list */ >>>> - spinlock_t ipsec_lock; >>>> + struct mutex ipsec_lock; >>>> #endif /* CONFIG_XFRM_OFFLOAD */ >>>> struct bpf_prog *xdp_prog; >>>> }; >>>> -- >>>> 2.21.0 > > --- > -Jay Vosburgh, jv@jvosburgh.net
On Thu, 22 Aug 2024 19:15:24 +0800 Jianbo Liu wrote:
> + dev_hold(real_dev);
netdev_hold(), please, so that we can find the location of the
reference leak if a bug sneaks in.
Jianbo Liu <jianbol@nvidia.com> wrote: [...] >I think it's good solution. >So I need to add the dev_hold/dev_put as following, for example, for >bond_ipsec_del_sa, right? > >@@ -526,6 +534,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) > bond = netdev_priv(bond_dev); > slave = rcu_dereference(bond->curr_active_slave); > real_dev = slave ? slave->dev : NULL; >+ dev_hold(real_dev); > rcu_read_unlock(); > > if (!slave) >@@ -545,6 +554,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) > > real_dev->xfrmdev_ops->xdo_dev_state_delete(xs); > out: >+ dev_put(real_dev); > mutex_lock(&bond->ipsec_lock); > list_for_each_entry(ipsec, &bond->ipsec_list, list) { > if (ipsec->xs == xs) { > >If you are ok with that, I will add the same for >bond_ipsec_add_sa/bond_ipsec_free_sa, and send new version. Yes, I think that will work, but please use netdev_hold() as Jakub requested. -J --- -Jay Vosburgh, jv@jvosburgh.net
diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c index 0d1129eaf47b..f20f6d83ad54 100644 --- a/drivers/net/bonding/bond_main.c +++ b/drivers/net/bonding/bond_main.c @@ -439,38 +439,33 @@ static int bond_ipsec_add_sa(struct xfrm_state *xs, rcu_read_lock(); bond = netdev_priv(bond_dev); slave = rcu_dereference(bond->curr_active_slave); - if (!slave) { - rcu_read_unlock(); + real_dev = slave ? slave->dev : NULL; + rcu_read_unlock(); + if (!real_dev) return -ENODEV; - } - real_dev = slave->dev; if (!real_dev->xfrmdev_ops || !real_dev->xfrmdev_ops->xdo_dev_state_add || netif_is_bond_master(real_dev)) { NL_SET_ERR_MSG_MOD(extack, "Slave does not support ipsec offload"); - rcu_read_unlock(); return -EINVAL; } - ipsec = kmalloc(sizeof(*ipsec), GFP_ATOMIC); - if (!ipsec) { - rcu_read_unlock(); + ipsec = kmalloc(sizeof(*ipsec), GFP_KERNEL); + if (!ipsec) return -ENOMEM; - } xs->xso.real_dev = real_dev; err = real_dev->xfrmdev_ops->xdo_dev_state_add(xs, extack); if (!err) { ipsec->xs = xs; INIT_LIST_HEAD(&ipsec->list); - spin_lock_bh(&bond->ipsec_lock); + mutex_lock(&bond->ipsec_lock); list_add(&ipsec->list, &bond->ipsec_list); - spin_unlock_bh(&bond->ipsec_lock); + mutex_unlock(&bond->ipsec_lock); } else { kfree(ipsec); } - rcu_read_unlock(); return err; } @@ -481,35 +476,35 @@ static void bond_ipsec_add_sa_all(struct bonding *bond) struct bond_ipsec *ipsec; struct slave *slave; - rcu_read_lock(); - slave = rcu_dereference(bond->curr_active_slave); - if (!slave) - goto out; + slave = rtnl_dereference(bond->curr_active_slave); + real_dev = slave ? slave->dev : NULL; + if (!real_dev) + return; - real_dev = slave->dev; + mutex_lock(&bond->ipsec_lock); if (!real_dev->xfrmdev_ops || !real_dev->xfrmdev_ops->xdo_dev_state_add || netif_is_bond_master(real_dev)) { - spin_lock_bh(&bond->ipsec_lock); if (!list_empty(&bond->ipsec_list)) slave_warn(bond_dev, real_dev, "%s: no slave xdo_dev_state_add\n", __func__); - spin_unlock_bh(&bond->ipsec_lock); goto out; } - spin_lock_bh(&bond->ipsec_lock); list_for_each_entry(ipsec, &bond->ipsec_list, list) { + /* If new state is added before ipsec_lock acquired */ + if (ipsec->xs->xso.real_dev == real_dev) + continue; + ipsec->xs->xso.real_dev = real_dev; if (real_dev->xfrmdev_ops->xdo_dev_state_add(ipsec->xs, NULL)) { slave_warn(bond_dev, real_dev, "%s: failed to add SA\n", __func__); ipsec->xs->xso.real_dev = NULL; } } - spin_unlock_bh(&bond->ipsec_lock); out: - rcu_read_unlock(); + mutex_unlock(&bond->ipsec_lock); } /** @@ -530,6 +525,8 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) rcu_read_lock(); bond = netdev_priv(bond_dev); slave = rcu_dereference(bond->curr_active_slave); + real_dev = slave ? slave->dev : NULL; + rcu_read_unlock(); if (!slave) goto out; @@ -537,7 +534,6 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) if (!xs->xso.real_dev) goto out; - real_dev = slave->dev; WARN_ON(xs->xso.real_dev != real_dev); if (!real_dev->xfrmdev_ops || @@ -549,7 +545,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) real_dev->xfrmdev_ops->xdo_dev_state_delete(xs); out: - spin_lock_bh(&bond->ipsec_lock); + mutex_lock(&bond->ipsec_lock); list_for_each_entry(ipsec, &bond->ipsec_list, list) { if (ipsec->xs == xs) { list_del(&ipsec->list); @@ -557,8 +553,7 @@ static void bond_ipsec_del_sa(struct xfrm_state *xs) break; } } - spin_unlock_bh(&bond->ipsec_lock); - rcu_read_unlock(); + mutex_unlock(&bond->ipsec_lock); } static void bond_ipsec_del_sa_all(struct bonding *bond) @@ -568,15 +563,12 @@ static void bond_ipsec_del_sa_all(struct bonding *bond) struct bond_ipsec *ipsec; struct slave *slave; - rcu_read_lock(); - slave = rcu_dereference(bond->curr_active_slave); - if (!slave) { - rcu_read_unlock(); + slave = rtnl_dereference(bond->curr_active_slave); + real_dev = slave ? slave->dev : NULL; + if (!real_dev) return; - } - real_dev = slave->dev; - spin_lock_bh(&bond->ipsec_lock); + mutex_lock(&bond->ipsec_lock); list_for_each_entry(ipsec, &bond->ipsec_list, list) { if (!ipsec->xs->xso.real_dev) continue; @@ -593,8 +585,7 @@ static void bond_ipsec_del_sa_all(struct bonding *bond) real_dev->xfrmdev_ops->xdo_dev_state_free(ipsec->xs); } } - spin_unlock_bh(&bond->ipsec_lock); - rcu_read_unlock(); + mutex_unlock(&bond->ipsec_lock); } static void bond_ipsec_free_sa(struct xfrm_state *xs) @@ -5917,7 +5908,7 @@ void bond_setup(struct net_device *bond_dev) /* set up xfrm device ops (only supported in active-backup right now) */ bond_dev->xfrmdev_ops = &bond_xfrmdev_ops; INIT_LIST_HEAD(&bond->ipsec_list); - spin_lock_init(&bond->ipsec_lock); + mutex_init(&bond->ipsec_lock); #endif /* CONFIG_XFRM_OFFLOAD */ /* don't acquire bond device's netif_tx_lock when transmitting */ @@ -5966,6 +5957,10 @@ static void bond_uninit(struct net_device *bond_dev) __bond_release_one(bond_dev, slave->dev, true, true); netdev_info(bond_dev, "Released all slaves\n"); +#ifdef CONFIG_XFRM_OFFLOAD + mutex_destroy(&bond->ipsec_lock); +#endif /* CONFIG_XFRM_OFFLOAD */ + bond_set_slave_arr(bond, NULL, NULL); list_del_rcu(&bond->bond_list); diff --git a/include/net/bonding.h b/include/net/bonding.h index b61fb1aa3a56..8bb5f016969f 100644 --- a/include/net/bonding.h +++ b/include/net/bonding.h @@ -260,7 +260,7 @@ struct bonding { #ifdef CONFIG_XFRM_OFFLOAD struct list_head ipsec_list; /* protecting ipsec_list */ - spinlock_t ipsec_lock; + struct mutex ipsec_lock; #endif /* CONFIG_XFRM_OFFLOAD */ struct bpf_prog *xdp_prog; };
In the cited commit, bond->ipsec_lock is added to protect ipsec_list, hence xdo_dev_state_add and xdo_dev_state_delete are called inside this lock. As ipsec_lock is a spin lock and such xfrmdev ops may sleep, "scheduling while atomic" will be triggered when changing bond's active slave. [ 101.055189] BUG: scheduling while atomic: bash/902/0x00000200 [ 101.055726] Modules linked in: [ 101.058211] CPU: 3 PID: 902 Comm: bash Not tainted 6.9.0-rc4+ #1 [ 101.058760] Hardware name: [ 101.059434] Call Trace: [ 101.059436] <TASK> [ 101.060873] dump_stack_lvl+0x51/0x60 [ 101.061275] __schedule_bug+0x4e/0x60 [ 101.061682] __schedule+0x612/0x7c0 [ 101.062078] ? __mod_timer+0x25c/0x370 [ 101.062486] schedule+0x25/0xd0 [ 101.062845] schedule_timeout+0x77/0xf0 [ 101.063265] ? asm_common_interrupt+0x22/0x40 [ 101.063724] ? __bpf_trace_itimer_state+0x10/0x10 [ 101.064215] __wait_for_common+0x87/0x190 [ 101.064648] ? usleep_range_state+0x90/0x90 [ 101.065091] cmd_exec+0x437/0xb20 [mlx5_core] [ 101.065569] mlx5_cmd_do+0x1e/0x40 [mlx5_core] [ 101.066051] mlx5_cmd_exec+0x18/0x30 [mlx5_core] [ 101.066552] mlx5_crypto_create_dek_key+0xea/0x120 [mlx5_core] [ 101.067163] ? bonding_sysfs_store_option+0x4d/0x80 [bonding] [ 101.067738] ? kmalloc_trace+0x4d/0x350 [ 101.068156] mlx5_ipsec_create_sa_ctx+0x33/0x100 [mlx5_core] [ 101.068747] mlx5e_xfrm_add_state+0x47b/0xaa0 [mlx5_core] [ 101.069312] bond_change_active_slave+0x392/0x900 [bonding] [ 101.069868] bond_option_active_slave_set+0x1c2/0x240 [bonding] [ 101.070454] __bond_opt_set+0xa6/0x430 [bonding] [ 101.070935] __bond_opt_set_notify+0x2f/0x90 [bonding] [ 101.071453] bond_opt_tryset_rtnl+0x72/0xb0 [bonding] [ 101.071965] bonding_sysfs_store_option+0x4d/0x80 [bonding] [ 101.072567] kernfs_fop_write_iter+0x10c/0x1a0 [ 101.073033] vfs_write+0x2d8/0x400 [ 101.073416] ? alloc_fd+0x48/0x180 [ 101.073798] ksys_write+0x5f/0xe0 [ 101.074175] do_syscall_64+0x52/0x110 [ 101.074576] entry_SYSCALL_64_after_hwframe+0x4b/0x53 As bond_ipsec_add_sa_all and bond_ipsec_del_sa_all are only called from bond_change_active_slave, which requires holding the RTNL lock. And bond_ipsec_add_sa and bond_ipsec_del_sa are xfrm state xdo_dev_state_add and xdo_dev_state_delete APIs, which are in user context. So ipsec_lock doesn't have to be spin lock, change it to mutex, and thus the above issue can be resolved. Fixes: 9a5605505d9c ("bonding: Add struct bond_ipesc to manage SA") Signed-off-by: Jianbo Liu <jianbol@nvidia.com> Signed-off-by: Tariq Toukan <tariqt@nvidia.com> Reviewed-by: Hangbin Liu <liuhangbin@gmail.com> --- drivers/net/bonding/bond_main.c | 67 +++++++++++++++------------------ include/net/bonding.h | 2 +- 2 files changed, 32 insertions(+), 37 deletions(-)