Re: Intel IOMMU (and IOMMU for Virtualization) performances

["Grant Grundler" <grundler@google.com>]

On Thu, Jun 5, 2008 at 9:44 PM, FUJITA Tomonori
<fujita.tomonori@lab.ntt.co.jp> wrote:
...
> The current Intel IOMMU scheme is a bit unbalanced. It invalidates the
> translation table every time dma_unmap_* is called, but it does the
> batching of the TLB flushes. But it's what the most of Linux's IOMMU
> code does.
>
> I think that only PARISC (and IA64, of course) IOMMUs do the batching
> of invalidating the translation table entries.

1/2 correct. PARISC and IA64 could be the same in this regard but are not.
See where sba_mark_invalid() is called in the respective sba_iommu.c.
PARISC invalidates the IO Pdir entry immediately but batches the
IO TLB shootdown and resource "free". IA64 could (and probably should)
do the same. Added Alex Williamson and Bjorn Helgaas to CC list.
Not an urgent issue though unless they are doing perf measurements
with SSDs or other block device with equivalent IOPS.

Since parisc-linux is unlikely to ever run VM's and the IOMMU has a
very limited number of IO TLB entries (8 or 16 about), I'm thinking the
batching is a waste of time and parisc should follow SPARC behavior.
I'll chat more with jejb/kyle/willy about it sometime.

thanks,
grant