bonding: fix destruction of bond with devices different from arphrd_ether
authorNikolay Aleksandrov <nikolay@cumulusnetworks.com>
Wed, 15 Jul 2015 19:52:51 +0000 (21:52 +0200)
committerDavid S. Miller <davem@davemloft.net>
Mon, 20 Jul 2015 19:56:11 +0000 (12:56 -0700)
commit06f6d1094aa0992432b1e2a0920b0ee86ccd83bf
treedb7819dadfadcb63bb708efdc681ecabfcadd3ae
parent40a7166044a6fce3dbcaa8b5c89845980c218c98
bonding: fix destruction of bond with devices different from arphrd_ether

When the bonding is being unloaded and the netdevice notifier is
unregistered it executes NETDEV_UNREGISTER for each device which should
remove the bond's proc entry but if the device enslaved is not of
ARPHRD_ETHER type and is in front of the bonding, it may execute
bond_release_and_destroy() first which would release the last slave and
destroy the bond device leaving the proc entry and thus we will get the
following error (with dynamic debug on for bond_netdev_event to see the
events order):
[  908.963051] eql: event: 9
[  908.963052] eql: IFF_SLAVE
[  908.963054] eql: event: 2
[  908.963056] eql: IFF_SLAVE
[  908.963058] eql: event: 6
[  908.963059] eql: IFF_SLAVE
[  908.963110] bond0: Releasing active interface eql
[  908.976168] bond0: Destroying bond bond0
[  908.976266] bond0 (unregistering): Released all slaves
[  908.984097] ------------[ cut here ]------------
[  908.984107] WARNING: CPU: 0 PID: 1787 at fs/proc/generic.c:575
remove_proc_entry+0x112/0x160()
[  908.984110] remove_proc_entry: removing non-empty directory
'net/bonding', leaking at least 'bond0'
[  908.984111] Modules linked in: bonding(-) eql(O) 9p nfsd auth_rpcgss
oid_registry nfs_acl nfs lockd grace fscache sunrpc crct10dif_pclmul
crc32_pclmul crc32c_intel ghash_clmulni_intel ppdev qxl drm_kms_helper
snd_hda_codec_generic aesni_intel ttm aes_x86_64 glue_helper pcspkr lrw
gf128mul ablk_helper cryptd snd_hda_intel virtio_console snd_hda_codec
psmouse serio_raw snd_hwdep snd_hda_core 9pnet_virtio 9pnet evdev joydev
drm virtio_balloon snd_pcm snd_timer snd soundcore i2c_piix4 i2c_core
pvpanic acpi_cpufreq parport_pc parport processor thermal_sys button
autofs4 ext4 crc16 mbcache jbd2 hid_generic usbhid hid sg sr_mod cdrom
ata_generic virtio_blk virtio_net floppy ata_piix e1000 libata ehci_pci
virtio_pci scsi_mod uhci_hcd ehci_hcd virtio_ring virtio usbcore
usb_common [last unloaded: bonding]

[  908.984168] CPU: 0 PID: 1787 Comm: rmmod Tainted: G        W  O
4.2.0-rc2+ #8
[  908.984170] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[  908.984172]  0000000000000000 ffffffff81732d41 ffffffff81525b34
ffff8800358dfda8
[  908.984175]  ffffffff8106c521 ffff88003595af78 ffff88003595af40
ffff88003e3a4280
[  908.984178]  ffffffffa058d040 0000000000000000 ffffffff8106c59a
ffffffff8172ebd0
[  908.984181] Call Trace:
[  908.984188]  [<ffffffff81525b34>] ? dump_stack+0x40/0x50
[  908.984193]  [<ffffffff8106c521>] ? warn_slowpath_common+0x81/0xb0
[  908.984196]  [<ffffffff8106c59a>] ? warn_slowpath_fmt+0x4a/0x50
[  908.984199]  [<ffffffff81218352>] ? remove_proc_entry+0x112/0x160
[  908.984205]  [<ffffffffa05850e6>] ? bond_destroy_proc_dir+0x26/0x30
[bonding]
[  908.984208]  [<ffffffffa057540e>] ? bond_net_exit+0x8e/0xa0 [bonding]
[  908.984217]  [<ffffffff8142f407>] ? ops_exit_list.isra.4+0x37/0x70
[  908.984225]  [<ffffffff8142f52d>] ?
unregister_pernet_operations+0x8d/0xd0
[  908.984228]  [<ffffffff8142f58d>] ?
unregister_pernet_subsys+0x1d/0x30
[  908.984232]  [<ffffffffa0585269>] ? bonding_exit+0x23/0xdba [bonding]
[  908.984236]  [<ffffffff810e28ba>] ? SyS_delete_module+0x18a/0x250
[  908.984241]  [<ffffffff81086f99>] ? task_work_run+0x89/0xc0
[  908.984244]  [<ffffffff8152b732>] ?
entry_SYSCALL_64_fastpath+0x16/0x75
[  908.984247] ---[ end trace 7c006ed4abbef24b ]---

Thus remove the proc entry manually if bond_release_and_destroy() is
used. Because of the checks in bond_remove_proc_entry() it's not a
problem for a bond device to change namespaces (the bug fixed by the
Fixes commit) but since commit
f9399814927ad ("bonding: Don't allow bond devices to change network
namespaces.") that can't happen anyway.

Reported-by: Carol Soto <clsoto@linux.vnet.ibm.com>
Signed-off-by: Nikolay Aleksandrov <nikolay@cumulusnetworks.com>
Fixes: a64d49c3dd50 ("bonding: Manage /proc/net/bonding/ entries from
                      the netdev events")
Tested-by: Carol L Soto <clsoto@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/bonding/bond_main.c