hv_netvsc: fix schedule in RCU context

When netvsc device is removed it can call reschedule in RCU context.
This happens because canceling the subchannel setup work could (in theory)
cause a reschedule when manipulating the timer.

To reproduce, run with lockdep enabled kernel and unbind
a network device from hv_netvsc (via sysfs).

[  160.682011] WARNING: suspicious RCU usage
[  160.707466] 4.19.0-rc3-uio+ #2 Not tainted
[  160.709937] -----------------------------
[  160.712352] ./include/linux/rcupdate.h:302 Illegal context switch in RCU read-side critical section!
[  160.723691]
[  160.723691] other info that might help us debug this:
[  160.723691]
[  160.730955]
[  160.730955] rcu_scheduler_active = 2, debug_locks = 1
[  160.762813] 5 locks held by rebind-eth.sh/1812:
[  160.766851]  #0: 000000008befa37a (sb_writers#6){.+.+}, at: vfs_write+0x184/0x1b0
[  160.773416]  #1: 00000000b097f236 (&of->mutex){+.+.}, at: kernfs_fop_write+0xe2/0x1a0
[  160.783766]  #2: 0000000041ee6889 (kn->count#3){++++}, at: kernfs_fop_write+0xeb/0x1a0
[  160.787465]  #3: 0000000056d92a74 (&dev->mutex){....}, at: device_release_driver_internal+0x39/0x250
[  160.816987]  #4: 0000000030f6031e (rcu_read_lock){....}, at: netvsc_remove+0x1e/0x250 [hv_netvsc]
[  160.828629]
[  160.828629] stack backtrace:
[  160.831966] CPU: 1 PID: 1812 Comm: rebind-eth.sh Not tainted 4.19.0-rc3-uio+ #2
[  160.832952] Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS Hyper-V UEFI Release v1.0 11/26/2012
[  160.832952] Call Trace:
[  160.832952]  dump_stack+0x85/0xcb
[  160.832952]  ___might_sleep+0x1a3/0x240
[  160.832952]  __flush_work+0x57/0x2e0
[  160.832952]  ? __mutex_lock+0x83/0x990
[  160.832952]  ? __kernfs_remove+0x24f/0x2e0
[  160.832952]  ? __kernfs_remove+0x1b2/0x2e0
[  160.832952]  ? mark_held_locks+0x50/0x80
[  160.832952]  ? get_work_pool+0x90/0x90
[  160.832952]  __cancel_work_timer+0x13c/0x1e0
[  160.832952]  ? netvsc_remove+0x1e/0x250 [hv_netvsc]
[  160.832952]  ? __lock_is_held+0x55/0x90
[  160.832952]  netvsc_remove+0x9a/0x250 [hv_netvsc]
[  160.832952]  vmbus_remove+0x26/0x30
[  160.832952]  device_release_driver_internal+0x18a/0x250
[  160.832952]  unbind_store+0xb4/0x180
[  160.832952]  kernfs_fop_write+0x113/0x1a0
[  160.832952]  __vfs_write+0x36/0x1a0
[  160.832952]  ? rcu_read_lock_sched_held+0x6b/0x80
[  160.832952]  ? rcu_sync_lockdep_assert+0x2e/0x60
[  160.832952]  ? __sb_start_write+0x141/0x1a0
[  160.832952]  ? vfs_write+0x184/0x1b0
[  160.832952]  vfs_write+0xbe/0x1b0
[  160.832952]  ksys_write+0x55/0xc0
[  160.832952]  do_syscall_64+0x60/0x1b0
[  160.832952]  entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  160.832952] RIP: 0033:0x7fe48f4c8154

Resolve this by getting RTNL earlier. This is safe because the subchannel
work queue does trylock on RTNL and will detect the race.

Fixes: 7b2ee50c0c ("hv_netvsc: common detach logic")
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit is contained in:
Stephen Hemminger 2018-09-13 08:03:43 -07:00 committed by David S. Miller
parent f5b9bac745
commit 018349d70f

View File

@ -2272,17 +2272,15 @@ static int netvsc_remove(struct hv_device *dev)
cancel_delayed_work_sync(&ndev_ctx->dwork);
rcu_read_lock();
nvdev = rcu_dereference(ndev_ctx->nvdev);
if (nvdev)
rtnl_lock();
nvdev = rtnl_dereference(ndev_ctx->nvdev);
if (nvdev)
cancel_work_sync(&nvdev->subchan_work);
/*
* Call to the vsc driver to let it know that the device is being
* removed. Also blocks mtu and channel changes.
*/
rtnl_lock();
vf_netdev = rtnl_dereference(ndev_ctx->vf_netdev);
if (vf_netdev)
netvsc_unregister_vf(vf_netdev);
@ -2294,7 +2292,6 @@ static int netvsc_remove(struct hv_device *dev)
list_del(&ndev_ctx->list);
rtnl_unlock();
rcu_read_unlock();
hv_set_drvdata(dev, NULL);