Re: [PATCH] generic device DMA (dma_pool update)

[David Brownell]

James Bottomley wrote:
> [e-mail] said:
> 
>>	Can someone show me or invent an example of two different uses of
>>dma_alloc_coherent that really should use different policies on
>>whether to block or not? 
> 
> 
> The obvious one is allocations from interrupt routines, which must be 
> GFP_ATOMIC (ignoring the issue of whether a driver should be doing a memory 
> allocation in an interrupt).  Allocating large pools at driver initialisation 
> should probably be GFP_KERNEL as you say.

More:  not just "from interrupt routines", also "when a spinlock is held".
Though one expects that if a dma_pool were in use, it'd probably have one
of that kind of chunk already available (ideally, even initialized).

Many task-initiated allocations would use GFP_KERNEL; not just major driver
activation points like open().

- Dave