Re: pci_pool_free from IRQ

[David Brownell <david-b@pacbell.net>]

>  > Pete's patch to pci_pool_free() is fine with me, and I'd be glad
>  > to see that bit of pci interface cleaned up.  Any changes needed
>  > other than the pci.txt doc update?
> 
> Ummm... What Alan's saying is:

(consistent with what I said -- those are two separate issues!)

> 1) Whatever driver is trying to shut down from IRQ context
>    is broken must be fixed.  pci_pool is fine.

In _that_ respect, yes.  pci_pool_destroy() called in shutdown
context "should" be OK.  Getting rid of pages then is fine.

> 2) The Documentation/ files which suggest that such device
>    removal from IRQs is "OK" must be fixed because it is not
>    "OK" to handle device removal from IRQ context.

All agreed.

> So Pete's change is not needed.  A fix for the documentation and
> broken drivers is needed instead.

Two issues are mixed up there.  Doc should address both:

Documentation/pci.txt

    ... that remove() is never called in_interrupt
        point (2) above

Documentation/DMA-mapping.txt

    pci_free_consistent() -- do not call in_interrupt
        ... unless we bugfix the ARM behavior "soon"

    pci_pool_destroy() -- do not call in_interrupt
        ... normally called on device remove()
        
    pci_pool_free() -- may be called in_interrupt
        ... often called in device interrupt handling

Pete's patch deferred some pci_free_consistent calls from
pci_pool_free() where they're unsafe (on ARM) to where
they're safe (all architectures, as discussed above).

- Dave