For streaming DMA mappings involving an IOMMU and whose IOVA len regularly
exceeds the IOVA rcache upper limit (meaning that they are not cached),
performance can be reduced.
This may be much more pronounced from commit 4e89dce72521 ("iommu/iova:
Retry from last rb tree node if iova search fails"), as discussed at [0].
IOVAs which cannot be cached are highly involved in the IOVA ageing issue,
as discussed at [1].
This series allows the IOVA rcache range be configured, so that we may
cache all IOVAs per domain, thus improving performance.
A new IOMMU group sysfs file is added - max_opt_dma_size - which is used
indirectly to configure the IOVA rcache range:
/sys/kernel/iommu_groups/X/max_opt_dma_size
This file is updated same as how the IOMMU group default domain type is
updated, i.e. must unbind the only device in the group first.
The inspiration here comes from block layer request queue sysfs
"optimal_io_size" file, in /sys/block/sdX/queue/optimal_io_size
Some old figures* for storage scenario (when increasing IOVA rcache range
to cover all DMA mapping sizes from the LLD):
v5.13-rc1 baseline: 1200K IOPS
With series: 1800K IOPS
All above are for IOMMU strict mode. Non-strict mode gives ~1800K IOPS in
all scenarios.
Based on v5.17-rc4 + [2]
* I lost my high data throughout test setup
Differences to v4:
https://lore.kernel.org/linux-iommu/[email protected]/
- Major rebase
- Change the "Refactor iommu_group_store_type()" to not use a callback
and an op type enum instead
- I didn't pick up Will's Ack as it has changed so much
- Use a domain feature flag to keep same default group type
- Add wrapper for default IOVA rcache range
- Combine last 2x patches
[0] https://lore.kernel.org/linux-iommu/[email protected]/
[1] https://lore.kernel.org/linux-iommu/[email protected]/
[2] https://lore.kernel.org/linux-iommu/[email protected]/T/#m5b2b59576d35cad544314470f32e5f40ac5d1fe9
John Garry (5):
iommu: Refactor iommu_group_store_type()
iova: Allow rcache range upper limit to be flexible
iommu: Allow iommu_change_dev_def_domain() realloc same default domain
type
iommu: Allow max opt DMA len be set for a group via sysfs
iova: Add iova_len argument to iova_domain_init_rcaches()
.../ABI/testing/sysfs-kernel-iommu_groups | 16 ++
drivers/iommu/dma-iommu.c | 15 +-
drivers/iommu/iommu.c | 202 +++++++++++++-----
drivers/iommu/iova.c | 37 ++--
drivers/vdpa/vdpa_user/iova_domain.c | 4 +-
include/linux/iommu.h | 7 +
include/linux/iova.h | 6 +-
7 files changed, 212 insertions(+), 75 deletions(-)
--
2.26.2
On 14/02/2022 17:29, John Garry wrote:
Hi guys,
And a friendly reminder on this series also.
Cheers,
john
> For streaming DMA mappings involving an IOMMU and whose IOVA len regularly
> exceeds the IOVA rcache upper limit (meaning that they are not cached),
> performance can be reduced.
>
> This may be much more pronounced from commit 4e89dce72521 ("iommu/iova:
> Retry from last rb tree node if iova search fails"), as discussed at [0].
>
> IOVAs which cannot be cached are highly involved in the IOVA ageing issue,
> as discussed at [1].
>
> This series allows the IOVA rcache range be configured, so that we may
> cache all IOVAs per domain, thus improving performance.
>
> A new IOMMU group sysfs file is added - max_opt_dma_size - which is used
> indirectly to configure the IOVA rcache range:
> /sys/kernel/iommu_groups/X/max_opt_dma_size
>
> This file is updated same as how the IOMMU group default domain type is
> updated, i.e. must unbind the only device in the group first.
>
> The inspiration here comes from block layer request queue sysfs
> "optimal_io_size" file, in /sys/block/sdX/queue/optimal_io_size
>
> Some old figures* for storage scenario (when increasing IOVA rcache range
> to cover all DMA mapping sizes from the LLD):
> v5.13-rc1 baseline: 1200K IOPS
> With series: 1800K IOPS
>
> All above are for IOMMU strict mode. Non-strict mode gives ~1800K IOPS in
> all scenarios.
>
> Based on v5.17-rc4 + [2]
> * I lost my high data throughout test setup
>
> Differences to v4:
> https://lore.kernel.org/linux-iommu/[email protected]/
> - Major rebase
> - Change the "Refactor iommu_group_store_type()" to not use a callback
> and an op type enum instead
> - I didn't pick up Will's Ack as it has changed so much
> - Use a domain feature flag to keep same default group type
> - Add wrapper for default IOVA rcache range
> - Combine last 2x patches
>
> [0] https://lore.kernel.org/linux-iommu/[email protected]/
> [1] https://lore.kernel.org/linux-iommu/[email protected]/
> [2] https://lore.kernel.org/linux-iommu/[email protected]/T/#m5b2b59576d35cad544314470f32e5f40ac5d1fe9
>
> John Garry (5):
> iommu: Refactor iommu_group_store_type()
> iova: Allow rcache range upper limit to be flexible
> iommu: Allow iommu_change_dev_def_domain() realloc same default domain
> type
> iommu: Allow max opt DMA len be set for a group via sysfs
> iova: Add iova_len argument to iova_domain_init_rcaches()
>
> .../ABI/testing/sysfs-kernel-iommu_groups | 16 ++
> drivers/iommu/dma-iommu.c | 15 +-
> drivers/iommu/iommu.c | 202 +++++++++++++-----
> drivers/iommu/iova.c | 37 ++--
> drivers/vdpa/vdpa_user/iova_domain.c | 4 +-
> include/linux/iommu.h | 7 +
> include/linux/iova.h | 6 +-
> 7 files changed, 212 insertions(+), 75 deletions(-)
>