netdev sysfs failure handling
Subsystems (2)
RFC 1/1 netdev sysfs failure handling
In case of sysfs failure, don't let device be brought up. It can be cleared by unregister_netdevice so module can be unloaded normally.
net/core/dev.c +8 -1
@@ -3043,10 +3043,17 @@switch (dev->reg_state) {case NETREG_REGISTERING : +/* Can't do proper error handling here because +* this is a delayed call after register_netdevice +* so no way to tell device driver what is wrong. +*/ err =netdev_register_sysfs (dev); -if (err) +if (err) {printk (KERN_ERR "%s: failed sysfs registration (%d)\n" , dev->name, err); +/* Don't let device be brought up */ +clear_bit (__LINK_STATE_PRESENT, &dev->state); + } dev->reg_state =NETREG_REGISTERED ;break ;
David S. Miller
maintainerFrom: Stephen Hemminger <[email protected]>
Date: Fri, 21 Apr 2006 13:42:05 -0700
Quoted 5 lines
In case of sysfs failure, don't let device be brought up. It can be cleared by
unregister_netdeviceso module can be unloaded normally.Signed-off-by Stephen Hemminger <[email protected]>
I'm not so sure about this, a hot plug event could clear that bit too.
The problem I think is that here we've structured things such that we can't handle the error properly and pass it back to the register_netdevice() caller because we do the sysfs registry call in the rtnl_unlock() todo list execution.
Next, even if you prevent the device from being brought up, people can still assign IP addresses and do other stuff to the device so it still sort of behaves as if it is there.
It would therefore be the best if we can do this stuff inside of register_netdevice(), then handle and propagate any errors correctly.
Something like this would handle errors better, but introduce possible problems for drivers that call register_netdevice with irq's disabled. There was some comment about racing with linkwatch, but don't see how that could happen during creation.
For 2.6.18?
include/linux/netdevice.h +1 -2
@@ -433,8 +433,7 @@/* register/unregister state machine */ enum {NETREG_UNINITIALIZED =0 , -NETREG_REGISTERING ,/* called register_netdevice */ -NETREG_REGISTERED ,/* completed register todo */ +NETREG_REGISTERED ,/* completed register_netdevice */ NETREG_UNREGISTERING ,/* called unregister_netdevice */ NETREG_UNREGISTERED ,/* completed unregister todo */ NETREG_RELEASED ,/* called free_netdev */
net/core/dev.c +28 -37
109 lines not on this page · raw message
David S. Miller
maintainerFrom: Stephen Hemminger <[email protected]>
Date: Tue, 9 May 2006 12:01:07 -0700
Quoted 6 lines
Something like this would handle errors better, but introduce possible problems for drivers that call
register_netdevicewith irq's disabled. There was some comment about racing with linkwatch, but don't see how that could happen during creation.For 2.6.18?
I've been thinking about this a bit more.
How can anyone be using this with IRQ's disabled if we have an ASSERT_RTNL() there?
Quoted 8 lines · On Tue, 09 May 2006 14:05:01 -0700 (PDT) "David S. Miller"…
From: Stephen Hemminger <[email protected]>
Date: Tue, 9 May 2006 12:01:07 -0700Quoted 5 lines
Something like this would handle errors better, but introduce possible problems for drivers that call
register_netdevicewith irq's disabled. There was some comment about racing with linkwatch, but don't see how that could happen during creation.
Quoted 6 lines
Quoted 1 line
For 2.6.18?
I've been thinking about this a bit more.
How can anyone be using this with IRQ's disabled if we have an
ASSERT_RTNL()there?
Agreed, especially since rtnl is now a real mutex. The case, that I was worried about:
rtnl_lock() spin_lock_irq(&mylock); x = register_netdevice();
...
Doesn't show up in any current code, even for the pseudo devices and funny virtualized interfaces.
David S. Miller
maintainerFrom: Stephen Hemminger <[email protected]>
Date: Tue, 9 May 2006 14:40:49 -0700
Quoted 9 lines
Agreed, especially since rtnl is now a real mutex. The case, that I was worried about:
rtnl_lock() spin_lock_irq(&mylock); x = register_netdevice();...
Doesn't show up in any current code, even for the pseudo devices and funny virtualized interfaces.
Right, therefore I think we should put something like your patch in there now.... perhaps.
The case where we really needed the todo list is unregister, so that we can safely wait for all references to the net device to go away.
I still wonder about those mentioned hotplug races wrt. linkwatch in the comment above netdev_run_todo().
Linkwatch is such a nuissance because it combines asynchronous link state change processing with keventd and RTNL locking. It sleeps waiting for __LINK_STATE_SCHED to clear with the RTNL held (via dev_deactivate()). But then again dev_close() code paths do this too, so the dev_deactivate() bit should be OK.
Linkwatch, after doing the dev_activate(), emits a NETDEV_CHANGE notifier on netdev_chain and also sends out an RTM_NETLINK message. This is for the case where IFF_UP is set.
Until we release the RTNL semaphore, during netdev register, nobody can go in an inspect the state of a net device. So doing the sysfs node creation in register_netdevice() should be OK as far as I can tell.
Can anyone find a problem with this?
Quoted 32 lines · On Tue, 09 May 2006 15:43:22 -0700 (PDT) "David S. Miller"…
From: Stephen Hemminger <[email protected]>
Date: Tue, 9 May 2006 14:40:49 -0700Quoted 9 lines
Agreed, especially since rtnl is now a real mutex. The case, that I was worried about:
rtnl_lock() spin_lock_irq(&mylock); x = register_netdevice();...
Doesn't show up in any current code, even for the pseudo devices and funny virtualized interfaces.
Right, therefore I think we should put something like your patch in there now.... perhaps.
The case where we really needed the todo list is unregister, so that we can safely wait for all references to the net device to go away.
I still wonder about those mentioned hotplug races wrt. linkwatch in the comment above
netdev_run_todo().Linkwatch is such a nuissance because it combines asynchronous link state change processing with keventd and RTNL locking. It sleeps waiting for
__LINK_STATE_SCHEDto clear with the RTNL held (viadev_deactivate()). But then againdev_close()code paths do this too, so thedev_deactivate()bit should be OK.Linkwatch, after doing the
dev_activate(), emits aNETDEV_CHANGEnotifier onnetdev_chainand also sends out anRTM_NETLINKmessage. This is for the case whereIFF_UPis set.
Quoted 6 lines
Until we release the RTNL semaphore, during netdev register, nobody can go in an inspect the state of a net device. So doing the sysfs node creation in
register_netdevice()should be OK as far as I can tell.Can anyone find a problem with this?
Also, by getting the netdevice fully in sysfs under RTNL, we are safe from races with the hotplug uevent that occurs.
Right now, it might be possible on SMP for the hotplug to happen after register_netdevice, but before the device shows up in sysfs.