Re: [PATCH] mv643xx_eth: Fix a possible deadlock upon ifdown

On Fri, Mar 01, 2013 at 06:30:13PM +0100, Alexander Holler wrote:

quoted

quoted

From: Lubomir Rintel <redacted>

=================================
[ INFO: inconsistent lock state ]
3.7.0-6.luboskovo.fc19.armv5tel.kirkwood #1 Tainted: G        W
---------------------------------
inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage.
NetworkManager/337 [HC0[0]:SC0[0]:HE1:SE1] takes:
 (_xmit_ETHER#2){+.?...}, at: [<bf07adfc>] txq_reclaim+0x54/0x264 [mv643xx_eth]

I get the same annoying warning when the MTU gets changed (through dhcp).

That is actually an issue.

quoted

Maybe I'm not reading it right, but I doubt that this is an actual
deadlock or that the patch is needed.

txq_reclaim() indeed doesn't disable BHs, but that's because it's
always called in BH context.  Almost always -- the only exception is
txq_deinit(), called from ->ndo_stop(), but by that time we've
already napi_disable()'d and netif_carrier_off()'d and free_irq()'d.

Agreed. I've just read me through that too and don't think a
deadlock is possible.

quoted

How to explain that to lockdep, though, I don't know.

The patch helps with that. ;)

It fixes a bug (the MTU change thing) and a non-bug (the lockdep
warning) at the expense of slowing down the much more common path,
and I don't like it for that reason.

Can you make a __txq_reclaim() which is basically txq_reclaim()
without grabbing the tx queue lock, and then move the lock grabbing
to the caller?

E.g. make __txq_reclaim() have two callers, txq_reclaim() and
txq_reclaim_bh(), and then use the appropriate wrapper depending on
the context.  (tx queue lock but no BH disable when called from
mv643xx_eth_poll(), tx queue lock plus BH disable for MTU change,
and no locking at all when called from ->ndo_stop().  Something
like that.)