Re: [PATCH net] net: use proper block lock instead of RTNL lock

[Vlad Buslov <vladbu@nvidia.com>]

On Mon 26 Oct 2020 at 10:28, Leon Romanovsky <leon@kernel.org> wrote:
> On Mon, Oct 26, 2020 at 10:10:31AM +0200, Vlad Buslov wrote:
>> On Mon 26 Oct 2020 at 08:04, Leon Romanovsky <leon@kernel.org> wrote:
>> > From: Leon Romanovsky <leonro@nvidia.com>
>> >
>> > The tcf_block_unbind() expects that the caller will take block->cb_lock
>> > before calling it, however the code took RTNL lock instead. This causes
>> > to the following kernel panic.
>> >
>> >  WARNING: CPU: 1 PID: 13524 at net/sched/cls_api.c:1488 tcf_block_unbind+0x2db/0x420
>> >  Modules linked in: mlx5_ib mlx5_core mlxfw ptp pps_core act_mirred act_tunnel_key cls_flower vxlan ip6_udp_tunnel udp_tunnel dummy sch_ingress openvswitch nsh xt_conntrack xt_MASQUERADE nf_conntrack_netlink nfnetlink xt_addrtype iptable_nat nf_nat nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 br_netfilter rpcrdma rdma_ucm ib_iser libiscsi scsi_transport_iscsi ib_umad ib_ipoib rdma_cm iw_cm ib_cm ib_uverbs ib_core overlay [last unloaded: mlxfw]
>> >  CPU: 1 PID: 13524 Comm: test-ecmp-add-v Tainted: G        W         5.9.0+ #1
>> >  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
>> >  RIP: 0010:tcf_block_unbind+0x2db/0x420
>> >  Code: ff 48 83 c4 40 5b 5d 41 5c 41 5d 41 5e 41 5f c3 49 8d bc 24 30 01 00 00 be ff ff ff ff e8 7d 7f 70 00 85 c0 0f 85 7b fd ff ff <0f> 0b e9 74 fd ff ff 48 c7 c7 dc 6a 24 84 e8 02 ec fe fe e9 55 fd
>> >  RSP: 0018:ffff888117d17968 EFLAGS: 00010246
>> >  RAX: 0000000000000000 RBX: ffff88812f713c00 RCX: 1ffffffff0848d5b
>> >  RDX: 0000000000000001 RSI: ffff88814fbc8130 RDI: ffff888107f2b878
>> >  RBP: 1ffff11022fa2f3f R08: 0000000000000000 R09: ffffffff84115a87
>> >  R10: fffffbfff0822b50 R11: ffff888107f2b898 R12: ffff88814fbc8000
>> >  R13: ffff88812f713c10 R14: ffff888117d17a38 R15: ffff88814fbc80c0
>> >  FS:  00007f6593d36740(0000) GS:ffff8882a4f00000(0000) knlGS:0000000000000000
>> >  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>> >  CR2: 00005607a00758f8 CR3: 0000000131aea006 CR4: 0000000000170ea0
>> >  Call Trace:
>> >   tc_block_indr_cleanup+0x3e0/0x5a0
>> >   ? tcf_block_unbind+0x420/0x420
>> >   ? __mutex_unlock_slowpath+0xe7/0x610
>> >   flow_indr_dev_unregister+0x5e2/0x930
>> >   ? mlx5e_restore_tunnel+0xdf0/0xdf0 [mlx5_core]
>> >   ? mlx5e_restore_tunnel+0xdf0/0xdf0 [mlx5_core]
>> >   ? flow_indr_block_cb_alloc+0x3c0/0x3c0
>> >   ? mlx5_db_free+0x37c/0x4b0 [mlx5_core]
>> >   mlx5e_cleanup_rep_tx+0x8b/0xc0 [mlx5_core]
>> >   mlx5e_detach_netdev+0xe5/0x120 [mlx5_core]
>> >   mlx5e_vport_rep_unload+0x155/0x260 [mlx5_core]
>> >   esw_offloads_disable+0x227/0x2b0 [mlx5_core]
>> >   mlx5_eswitch_disable_locked.cold+0x38e/0x699 [mlx5_core]
>> >   mlx5_eswitch_disable+0x94/0xf0 [mlx5_core]
>> >   mlx5_device_disable_sriov+0x183/0x1f0 [mlx5_core]
>> >   mlx5_core_sriov_configure+0xfd/0x230 [mlx5_core]
>> >   sriov_numvfs_store+0x261/0x2f0
>> >   ? sriov_drivers_autoprobe_store+0x110/0x110
>> >   ? sysfs_file_ops+0x170/0x170
>> >   ? sysfs_file_ops+0x117/0x170
>> >   ? sysfs_file_ops+0x170/0x170
>> >   kernfs_fop_write+0x1ff/0x3f0
>> >   ? rcu_read_lock_any_held+0x6e/0x90
>> >   vfs_write+0x1f3/0x620
>> >   ksys_write+0xf9/0x1d0
>> >   ? __x64_sys_read+0xb0/0xb0
>> >   ? lockdep_hardirqs_on_prepare+0x273/0x3f0
>> >   ? syscall_enter_from_user_mode+0x1d/0x50
>> >   do_syscall_64+0x2d/0x40
>> >   entry_SYSCALL_64_after_hwframe+0x44/0xa9
>> >
>> > <...>
>> >
>> >  ---[ end trace bfdd028ada702879 ]---
>> >
>> > Fixes: 0fdcf78d5973 ("net: use flow_indr_dev_setup_offload()")
>> > Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
>> > ---
>> >  net/sched/cls_api.c | 4 +---
>> >  1 file changed, 1 insertion(+), 3 deletions(-)
>> >
>> > diff --git a/net/sched/cls_api.c b/net/sched/cls_api.c
>> > index faeabff283a2..fca147386ad2 100644
>> > --- a/net/sched/cls_api.c
>> > +++ b/net/sched/cls_api.c
>> > @@ -655,10 +655,8 @@ static void tc_block_indr_cleanup(struct flow_block_cb *block_cb)
>> >  	down_write(&block->cb_lock);
>> >  	list_del(&block_cb->driver_list);
>> >  	list_move(&block_cb->list, &bo.cb_list);
>> > -	up_write(&block->cb_lock);
>> > -	rtnl_lock();
>> >  	tcf_block_unbind(block, &bo);
>> > -	rtnl_unlock();
>> > +	up_write(&block->cb_lock);
>>
>> Hi Leon,
>>
>> This fix probably breaks all other drivers besides mlx5 that require
>> rtnl lock on TC filter update path (called from tp->ops->reoffload()).
>> I suggest both rtnl lock and cb_lock are required here.
>
> You are not taking rtnl lock in other bind/unbind paths, why this is different?
> Are you taking rtnl lock in layers above?

Yes. Qdisc init that calls tcf_block_get_ext() is already protected by
rtnl lock.

>
> Thanks