Re: DMA APIs gumble grumble

[David Miller <davem@davemloft.net>]

From: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Date: Fri, 10 Nov 2006 12:02:04 +1100

> > Please just mirror what I did on sparc64 for sparc32, see changeset
> > 42f142371e48fbc44956d57b4e506bb6ce673cd7, with followup bug fixes
> > in 36321426e320c2c6bc2f8a1587d6f4d695fca84c and
> > 7233589d77fdb593b482a8b7ee867e901f54b593.
> 
> Question about sparc. It's implementation of pci_alloc_consistent(),
> unlike the other ones from before we had a GFP mask massed, does
> GFP_KERNEL allocations and not GFP_ATOMIC. Thus it's never expected to
> be called in atomic context. In fact, it does various other things like
> calling allocate_resource which is not something you ever want to be
> called from interrupt context.

pci_alloc_consistent() is not allowed from atomic contexts.

> I'm splitting it into a pci_do_alloc_consistent that takes a gfp arg,
> and a pair of pci_alloc_consistent & dma_alloc_consistent wrappers.
> 
> Do you think I should have the former pass GFP_KERNEL like the current
> implementation does or switch it to GFP_ATOMIC like everybody does ? In
> this case, should I also change the kmalloc done in there to allocate a
> struct resource to use the gfp argument ? (It's currently doing
> GFP_KERNEL).

pci_alloc_consistent() really cannot be allowed to use GFP_ATOMIC.