> iirc nbd doesn't need it from an interrupt but just to avoid deadlock
> on low memory situations when you swap over it.
Hmm... seems, I do not understand. Are such deadlocks prevented
by some special GFP values without any inheriting? This sounds strange.
It is evident candidate to an attribute in task struct, at least.
This was something sort of PF_MEMALLOC, if I remember correclty.
BTW we had problem with kmalloc (that which was fixed by your patch)
exactly because it avoids deadlocks and does not respect reliability
of GFP_KERNEL. So, does it mean that we have not only unreliable
GFP_KERNEL but also inclined to deadlocks in any case? Fuuunny situation.