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

[David Brownell]

Adam J. Richter wrote:
> 	In practice, I think that if we just added one, maybe two,
> URB's by default for every endpoint when a device is added, that
> that would be enough to guarantee that would reduce the number of
> drivers that needed to reserve more URB's than that to few or none.

I seem to recall someone posted a patch to make non-iso URB allocation
use a mempool.

>>Hmm, I was unaware that anyone expected GFP_KERNEL (or rather,
>>__GFP_WAIT) to guarantee that memory was always returned.  It's
>>not called __GFP_NEVERFAIL, after all.
> 
> 
> 	mempool_alloc does.  That's the point of it.  You calculate
> how many objects you need in order to guarantee no deadlocks and
> reserve that number in advance (the initial reservation can fail).

To rephrase that so it illustrates my point:  the whole reason to
use mempool is to try adding __GFP_NEVERFAIL when __GFP_WAIT is
given ... because __GFP_WAIT doesn't otherwise mean NEVERFAIL.

- Dave