Hi Tomasz,
On 29/09/15 06:25, Tomasz Figa wrote:
Currently the IOMMU subsystem provides 3 basic operations: iommu_map(),
iommu_map_sg() and iommu_unmap(). iommu_map() can be used to map memory
page by page, however it involves flushing the caches (CPU and IOMMU) for
every mapped page separately, which is unsuitable for use cases that
require low mapping latency. Similarly iommu_unmap(), even though it
takes a full IOVA range as its argument, performs unmapping in a page
by page manner.
This isn't necessarily the general case, though. If the IOMMU has a
coherent page table walk interface and its architecture prohibits
caching invalid PTEs, then the overhead on an unmap is only a TLB
invalidation and the overhead on a map is nothing.
To make mapping operation more suitable for such use cases, iommu_map_sg()
and .map_sg() callback in iommu_ops struct were introduced, which allowed
particular IOMMU drivers to directly iterate over SG entries, create
necessary mappings and flush everything in one go.
This approach, however, has two drawbacks:
1) it does not do anything about unmap performance,
2) it requires each driver willing to have fast map to implement its
own SG iteration code, even though this is a mostly generic operation.
This series tries to mitigate the two issues above, while acknowledging
the fact that the .map_sg() callback might be still necessary for some
specific platforms, which could have the need to iterate over SG elements
inside driver code. Proposed solution introduces a new .flush() callback,
which expects IOVA range as its argument and is expected to flush all
respective caches (be it CPU, IOMMU TLB or whatever) to make the given
IOVA area mapping change visible to IOMMU clients. Then all the 3 basic
map/unmap operations are modified to call the .flush() callback at the end
of the operation.
Advantages of proposed approach include:
1) ability to use default_iommu_map_sg() helper if all the driver needs
for performance optimization is batching the flush,
2) completely no effect on existing code - the .flush() callback is made
optional and if it isn't implemented drivers are expected to do
necessary flushes on a page by page basis in respective (un)mapping
callbakcs,
3) possibility of exporting the iommu_flush() operation and providing
unsynchronized map/unmap operations for subsystems with even higher
requirements for performance (e.g. drivers/gpu/drm).
A single callback doesn't really generalise well enough: If we wanted to
implement this in the ARM SMMU drivers to optimise the unmap() case [ask
Will how long he spends waiting for a software model to tear down an
entire VFIO domain invalidating one page at a time ;)], then we'd either
regress performance in the map() case with an unnecessary TLB flush, or
have to do a table walk in every flush() call to infer what actually
needs doing.
Personally I think it would be nicest to have two separate callbacks,
e.g. .map_sync/.unmap_sync, but at the very least some kind of
additional 'direction' kind of parameter would be necessary.
The series includes a generic patch implementing necessary changes in
IOMMU API and two Tegra-specific patches that demonstrate implementation
on driver side and which can be used for further testing.
Last, but not least, some performance numbers on Tegra210:
+-----------+--------------+-------------+------------+
| Operation | Size [bytes] | Before [us] | After [us] |
+-----------+--------------+-------------+------------+
| Map | 128K | 139 | 40 |
| | | 136 | 34 |
| | | 137 | 38 |
| | | 136 | 36 |
| | 4M | 3939 | 1163 |
| | | 3730 | 2389 |
| | | 3613 | 997 |
| | | 3622 | 1620 |
| | ~18M | 18635 | 4741 |
| | | 19261 | 6550 |
| | | 18473 | 9304 |
| | | 18125 | 5120 |
| Unmap | 128K | 128 | 7 |
| | | 122 | 8 |
| | | 119 | 10 |
| | | 123 | 12 |
| | 4M | 3829 | 151 |
| | | 3964 | 150 |
| | | 3908 | 145 |
| | | 3875 | 155 |
| | ~18M | 18570 | 683 |
| | | 18473 | 806 |
| | | 21020 | 643 |
| | | 21764 | 652 |
+-----------+--------------+-------------+------------+
The values are obtained by surrounding the calls to iommu_map_sg()
(with default_iommu_map_sg() helper used as .map_sg() callback) and
iommu_unmap() with ktime-based time measurement code. Taken 4 samples
of every buffer size. ~18M means around 17-19M due do the variance
in requested buffer sizes.
It would be interesting to know how much of the gain here is due to
batching up the TLB maintenance vs. doing the DMA sync in fewer total
pieces (with correspondingly fewer barriers). For the drivers which use
the io-pgtable framework, trying to do anything about the latter (where
it's relevant) looks essentially impossible without rewriting the whole
thing, but the former is definitely something we should be able to
handle and benefit from.
It certainly seems like a reasonable way to get closer to the kind of
iommu_map_range()/iommu_unmap_range() operations proposed before, but
with less trampling on the external API. Plus it's nicer than the
alternative workaround of having the driver claim anything larger than
your basic page size is valid so you can batch up most of your pagetable
updates behind the API's back.
Robin.
Tomasz Figa (2):
iommu: Add support for out of band flushing
iommu/tegra-smmu: Make the driver use out of band flushing
Vince Hsu (1):
memory: tegra: add TLB cache line size
drivers/iommu/iommu.c | 33 +++++++++++++--
drivers/iommu/tegra-smmu.c | 91 +++++++++++++++++++++++++++++++++++++----
drivers/memory/tegra/tegra114.c | 1 +
drivers/memory/tegra/tegra124.c | 3 ++
drivers/memory/tegra/tegra210.c | 1 +
drivers/memory/tegra/tegra30.c | 1 +
include/linux/iommu.h | 2 +
include/soc/tegra/mc.h | 1 +
8 files changed, 122 insertions(+), 11 deletions(-)
--
To unsubscribe from this list: send the line "unsubscribe linux-tegra" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
