Re: [PATCH] net-sysfs: Fix memory leak in netdev_register_kobject

[Stephen Hemminger]

On Tue, 19 Mar 2019 11:03:54 +0800
"wanghai (M)" <wanghai26@xxxxxxxxxx> wrote:

> å 2019/3/18 23:57, Stephen Hemminger åé:
> > On Tue, 19 Mar 2019 01:06:57 -0400
> > Wang Hai <wanghai26@xxxxxxxxxx> wrote:
> >  
> >> When registering struct net_device, it will call
> >> 	register_netdevice ->
> >> 		netdev_register_kobject ->
> >> 			device_add(dev)
> >> 			register_queue_kobjects(ndev)
> >>
> >> If device_add(dev) or register_queue_kobjects(ndev) fails.
> >> Register_netdevice() will return error, causing netdev_freemem(ndev)
> >> to be called to free net_device, however (&ndev->dev)->kobj.name will
> >> not be freed, resulting in a memory leak.
> >>
> >> syzkaller report this:
> >> BUG: memory leak
> >> unreferenced object 0xffff8881f4fad168 (size 8):
> >> comm "syz-executor.0", pid 3575, jiffies 4294778002 (age 20.134s)
> >> hex dump (first 8 bytes):
> >>    77 70 61 6e 30 00 ff ff                          wpan0...
> >> backtrace:
> >>    [<000000006d2d91d7>] kstrdup_const+0x3d/0x50 mm/util.c:73
> >>    [<00000000ba9ff953>] kvasprintf_const+0x112/0x170 lib/kasprintf.c:48
> >>    [<000000005555ec09>] kobject_set_name_vargs+0x55/0x130 lib/kobject.c:281
> >>    [<0000000098d28ec3>] dev_set_name+0xbb/0xf0 drivers/base/core.c:1915
> >>    [<00000000b7553017>] netdev_register_kobject+0xc0/0x410 net/core/net-sysfs.c:1727
> >>    [<00000000c826a797>] register_netdevice+0xa51/0xeb0 net/core/dev.c:8711
> >>    [<00000000857bfcfd>] cfg802154_update_iface_num.isra.2+0x13/0x90 [ieee802154]
> >>    [<000000003126e453>] ieee802154_llsec_fill_key_id+0x1d5/0x570 [ieee802154]
> >>    [<00000000e4b3df51>] 0xffffffffc1500e0e
> >>    [<00000000b4319776>] platform_drv_probe+0xc6/0x180 drivers/base/platform.c:614
> >>    [<0000000037669347>] really_probe+0x491/0x7c0 drivers/base/dd.c:509
> >>    [<000000008fed8862>] driver_probe_device+0xdc/0x240 drivers/base/dd.c:671
> >>    [<00000000baf52041>] device_driver_attach+0xf2/0x130 drivers/base/dd.c:945
> >>    [<00000000c7cc8dec>] __driver_attach+0x10e/0x210 drivers/base/dd.c:1022
> >>    [<0000000057a757c2>] bus_for_each_dev+0x154/0x1e0 drivers/base/bus.c:304
> >>    [<000000005f5ae04b>] bus_add_driver+0x427/0x5e0 drivers/base/bus.c:645
> >>
> >> Reported-by: Hulk Robot <hulkci@xxxxxxxxxx>
> >> Fixes: 1d24eb4815d1 ("xps: Transmit Packet Steering")
> >> Signed-off-by: Wang Hai <wanghai26@xxxxxxxxxx>
> >> ---
> >>   net/core/net-sysfs.c | 15 ++++++++++-----
> >>   1 file changed, 10 insertions(+), 5 deletions(-)
> >>
> >> diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c
> >> index 4ff661f..f0e53dc 100644
> >> --- a/net/core/net-sysfs.c
> >> +++ b/net/core/net-sysfs.c
> >> @@ -1745,17 +1745,22 @@ int netdev_register_kobject(struct net_device *ndev)
> >>   
> >>   	error = device_add(dev);
> >>   	if (error)
> >> -		return error;
> >> +		goto device_add_error;
> >>   
> >>   	error = register_queue_kobjects(ndev);
> >> -	if (error) {
> >> -		device_del(dev);
> >> -		return error;
> >> -	}
> >> +	if (error)
> >> +		goto register_error;
> >>   
> >>   	pm_runtime_set_memalloc_noio(dev, true);
> >>   
> >> +out:
> >>   	return error;
> >> +
> >> +register_error:
> >> +	device_del(dev);
> >> +device_add_error:
> >> +	kfree_const(dev->kobj.name);  
> > This looks a bug in device_add() not here.
> > In general, it is better for an api to clean up after itself.
> > Since dev->kobj.name is created in device_add and normally freed
> > in device_del; why is device_add leaving it behind?\  
> 
> When registering struct net_device, it will call
> register_netdevice ->
>  ÂÂÂ netdev_register_kobject ->
>  ÂÂÂ ÂÂÂ dev_set_name(dev, "%s", ndev->name)
>  ÂÂÂÂ ÂÂÂ device_add(dev)
>  ÂÂÂÂ ÂÂÂ register_queue_kobjects(ndev)
> 
> The dev->kobj.name that causes the memory leak is created in 
> dev_set_name(dev, "%s", ndev-> name) in the function 
> netdev_register_kobject(), not in device_add(dev). If device_add(dev) or 
> register_queue_kobjects(ndev) fails, it should release dev-> kobj.name 
> in netdev_register_kobject()

Good catch, dev_set_name is using asprintf to allocate new memory
for the name. The problem with your patch is that device_del()
already cleans up the device (including name). The device object
should really not be touched after device_del.

Where is the name freed in the normal case?