Peter Zijlstra wrote:

On Wed, 2006-08-09 at 02:30 -0400, Jeff Garzik wrote:

quoted

David Miller wrote:

quoted

From: Daniel Phillips <redacted>
Date: Tue, 08 Aug 2006 22:51:20 -0700

quoted

Elaborate please.  Do you think that all drivers should be updated to
fix the broken blockdev semantics, making NETIF_F_MEMALLOC redundant?
If so, I trust you will help audit for it?

I think he's saying that he doesn't think your code is yet a
reasonable way to solve the problem, and therefore doesn't belong
upstream.

Pretty much.  It is completely non-sensical to add NETIF_F_MEMALLOC, 
when it should be blindingly obvious that every net driver will be 
allocating memory, and every net driver could potentially be used with 
NBD and similar situations.

Sure, but until every single driver is converted I'd like to warn people
about the fact that their setups is not up to expectations. Iff all
drivers are converted I'll be the forst to submit a patch that removes
the feature flag.

A temporary-for-years flag is not a good approach.  The flag is not 
_needed_ for technical reasons, but for supposed user expectation reasons.

Rather, just go ahead and convert drivers to netdev_alloc_skb() where 
people care.  If someone suddenly gets a burr up their ass about the 
sunlance or epic100 driver deadlocking on NBD, then they can convert it 
or complain loudly themselves.

Overall, a good solution needs to be uniform across all net drivers. 
NETIF_F_MEMALLOC is just _encouraging_ people to be slackers and delay 
converting other drivers, creating two classes of drivers, the "haves" 
and the "have nots".

Just make a big netdev_alloc_skb() patch that converts most users. 
netdev_alloc_skb() is a good thing to use, because it builds an 
association with struct net_device and the allocation.

	Jeff



P.S.  Since netdev_alloc_skb() calls skb_reserve(), you need to take 
that into account.  That's a bug in current patches.