We only ever obtain this lock immediately before the iova_rbtree_lock,
and release it immediately after the iova_rbtree_lock. So ditch it and
just use iova_rbtree_lock.
[v2: Remove the lockdep bits this time too]
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
We did before, in the end -- but it was at the bottom of a long stack of
functions. Add an inline wrapper get_valid_domain_for_dev() which will
use the cached one _first_ and only make the out-of-line call if it's
not already set.
This takes the average time taken for a 1-page intel_map_sg() from 5961
cycles to 4812 cycles on my Lenovo x200s test box -- a modest 20%.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Our current strategy for pass-through mode is to put all devices into
the 1:1 domain at startup (which is before we know what their dma_mask
will be), and only _later_ take them out of that domain, if it turns out
that they really can't address all of memory.
However, when there are a bunch of PCI devices behind a bridge, they all
end up with the same source-id on their DMA transactions, and hence in
the same IOMMU domain. This means that we _can't_ easily move them from
the 1:1 domain into their own domain at runtime, because there might be DMA
in-flight from their siblings.
So we have to adjust our pass-through strategy: For PCI devices not on
the root bus, and for the bridges which will take responsibility for
their transactions, we have to start up _out_ of the 1:1 domain, just in
case.
This fixes the BUG() we see when we have 32-bit-capable devices behind a
PCI-PCI bridge, and use the software identity mapping.
It does mean that we might end up using 'normal' mapping mode for some
devices which could actually live with the faster 1:1 mapping -- but
this is only for PCI devices behind bridges, which presumably aren't the
devices for which people are most concerned about performance.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
At boot time, the dma_mask won't have been set on any devices, so we
assume that all devices will be 64-bit capable (and thus get a 1:1 map).
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
This should fix kernel.org bug #11821, where the dcdbas driver makes up
a platform device and then uses dma_alloc_coherent() on it, in an
attempt to get memory < 4GiB.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
We need to give people a little more time to fix the broken drivers.
Re-introduce this, but tied in properly with the 'iommu=pt' support this
time. Change the config option name and make it default to 'no' too.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
We do this twice, and it's about to get more complicated. This makes the
code slightly clearer about what it's doing, too.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
When we reattach a device to the si_domain (because it's been removed
from a VM), we weren't calling domain_context_mapping() to actually tell
the hardware about that.
We should really put the call to domain_context_mapping() into
domain_add_dev_info() -- we never call the latter without also doing the
former, and we can keep the error paths simple that way. But that's a
cleanup which can wait for 2.6.32 now.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
We should check iommu_dummy() _first_, because that means it's attached
to an iommu that we've just disabled completely. At the moment, we might
try to put the device into the identity mapping domain.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
The aligned_nrpages() function rounds up to the next VM page, but
returns its result as a number of DMA pages.
Purely theoretical except on IA64, which doesn't boot with VT-d right
now anyway.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Check dma_pte_present() and only free the page if there _is_ one.
Kind of surprising that there was no warning about this.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
On Wed, 2009-07-01 at 16:59 -0700, Linus Torvalds wrote:
> I also _really_ hate how you do
>
> (unsigned long)pte >> VTD_PAGE_SHIFT ==
> (unsigned long)first_pte >> VTD_PAGE_SHIFT
Kill this, in favour of just looking to see if the incremented pte
pointer has 'wrapped' onto the next page. Which means we have to check
it _after_ incrementing it, not before.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
This would have found the bug in i386 pci_unmap_addr() a long time ago.
We shouldn't just silently return without doing anything.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Since we're using cmpxchg64() anyway (because that's the only way to do
an atomic 64-bit store on i386), we might as well ditch the extra
locking and just use cmpxchg64() to ensure that we don't add the page
twice.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
As with other functions, batch the CPU data cache flushes and don't keep
recalculating PTE addresses.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
The loop condition was wrong -- we should free a PMD only if its
_entire_ range is within the range we're intending to clear. The
early-termination condition was right, but not the loop.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Instead of calling domain_pfn_mapping() repeatedly with single or
small numbers of pages, just pass the sglist in. It can optimise the
number of cache flushes like domain_pfn_mapping() does, and gives a huge
speedup for large scatterlists.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
There's no need for the separate iommu_alloc_iova() function, and
certainly not for it to be global. Remove the underscores while we're at
it.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
As with dma_pte_clear_range(), don't keep flushing a single PTE at a
time. And also micro-optimise the setting of PTE values rather than
using the helper functions to do all the masking.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
It's a bit silly to repeatedly call domain_flush_cache() for each PTE
individually, as we clear it. Instead, batch them up and flush a whole
range at a time. We might as well refrain from recalculating the PTE
address from scratch each time round the loop too.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
This is fairly broken anyway -- it doesn't take hotplug into account.
We should probably be checking page_is_ram() instead.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Most of its callers are having to shift for themselves anyway, so we might
as well do it in iommu_flush_iotlb_psi().
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
... and use it in the trivial cases; the other callers want individual
(and bisectable) attention, since I screwed them up the first time...
Make the BUG_ON() happen on too-large virtual address rather than
physical address, too. That's the one we care about.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Use unaligned address for domain->max_addr. That algorithm isn't ideal
anyway -- we should probably just look at the last iova in the tree.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
With some cleanup of intel_unmap_page(), intel_unmap_sg() and
vm_domain_exit() to no longer play with 64-bit addresses.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Add some helpers for converting between VT-d and normal system pfns,
since system pages can be larger than VT-d pages.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
There's no need for the GFX workaround now we have 'iommu=pt' for the
cases where people really care about performance. There's no need to
have a special case for just one type of device.
This also speeds up the iommu=pt path and reduces memory usage by
setting up the si_domain _once_ and then using it for all devices,
rather than giving each device its own private page tables.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
In caching mode, domain ID 0 is reserved for non-present to present
mapping flush. Device IOTLB doesn't need to be flushed in this case.
Previously we were avoiding the flush for domain zero, even if the IOMMU
wasn't in caching mode and domain zero wasn't special.
Signed-off-by: Yu Zhao <yu.zhao@intel.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Drop the e820 scanning and use existing function for finding valid
RAM regions to add to 1:1 mapping.
Signed-off-by: Chris Wright <chrisw@redhat.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Identity mapping for IOMMU defines a single domain to 1:1 map all PCI
devices to all usable memory.
This reduces map/unmap overhead in DMA API's and improve IOMMU
performance. On 10Gb network cards, Netperf shows no performance
degradation compared to non-IOMMU performance.
This method may lose some of DMA remapping benefits like isolation.
The patch sets up identity mapping for all PCI devices to all usable
memory. In the DMA API, there is no overhead to maintain page tables,
invalidate iotlb, flush cache etc.
32 bit DMA devices don't use identity mapping domain, in order to access
memory beyond 4GiB.
When kernel option iommu=pt, pass through is first tried. If pass
through succeeds, IOMMU goes to pass through. If pass through is not
supported in hw or fail for whatever reason, IOMMU goes to identity
mapping.
Signed-off-by: Fenghua Yu <fenghua.yu@intel.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
