[SRU][U/O][N][PATCH 0/1] IOMMU DMA mode changed in kernel config causes massive throughput degradation for PCI-related network workloads (LP: 2071471)

[frank.heimes at canonical.com]

BugLink: https://bugs.launchpad.net/bugs/2071471

SRU Justification:

[Impact]

 * With the introduction of c76c067e488c "s390/pci: Use dma-iommu layer"
   (upstream with since kernel v6.7-rc1) there was a move (on s390x only)
   to a different dma-iommu implementation.

 * And with 92bce97f0c34 "s390/pci: Fix reset of IOMMU software counters"
   (again upstream since 6.7(rc-1) the IOMMU_DEFAULT_DMA_LAZY kernel config
   option should now be set to 'yes' by default for s390x.

 * Since CONFIG_IOMMU_DEFAULT_DMA_STRICT and IOMMU_DEFAULT_DMA_LAZY
   are related to each other CONFIG_IOMMU_DEFAULT_DMA_STRICT needs to be
   set to "no" by default, which was upstream done by b2b97a62f055
   "Revert "s390: update defconfigs"".

 * These changes are all upstream, but were not picked up by the Ubuntu
   kernel config.

 * And not having these config options set properly is causing significant
   PCI-related network throughput degradation (up to -72%).

 * This shows for almost all workloads and numbers of connections,
   deteriorating with the number of connections increasing.

 * Especially drastic is the drop for a high number of parallel connections
   (50 and 250) and for small and medium-size transactional workloads.
   However, also for streaming-type workloads the degradation is clearly
   visible (up to 48% degradation).

[Fix]

 * The (upstream accepted) fix is to set
   IOMMU_DEFAULT_DMA_STRICT=no
   and
   IOMMU_DEFAULT_DMA_LAZY=y
   (which is needed for the changed DAM IOMMU implementation since v6.7).

[Test Case]

 * Setup two Ubuntu Server 24.04 LPARs (with kernel 6.8)
   (one acting as server and as client)
   that have (PCIe attached) RoCE Express devices attached
   and that are connected to each other.

 * Sample workload rr1c-200x1000-250 with rr1c-200x1000-250.xml:
   <?xml version="1.0"?>
   <profile name="TCP_RR">
           <group nprocs="250">
                   <transaction iterations="1">
                           <flowop type="connect" options="remotehost=<remote IP> protocol=tcp tcp_nodelay" />
                   </transaction>
                   <transaction duration="300">
                           <flowop type="write" options="size=200"/>
                           <flowop type="read" options="size=1000"/>
                   </transaction>
                   <transaction iterations="1">
                           <flowop type="disconnect" />
                   </transaction>
           </group>
   </profile>

 * Install uperf on both systems, client and server.

 * Start uperf at server: uperf -s

 * Start uperf at client: uperf -vai 5 -m uperf-profile.xml

 * Switch from strict to lazy mode 
   either using the new kernel (or the test build below)
   or using kernel cmd-line parameter iommu.strict=0.

 * Restart uperf on server and client, like before.

 * Verification will be performed by IBM.

[Regression Potential]

 * The is a certain regression potential, since the behavior with
   the two modified kernel config options will change significantly.

 * This may solve the (network) throughput issue with PCI devices,
   but may also come with side-effects on other PCIe based devices
   (the old compression adapters or the new NVMe carrier cards).

[Other]

 * CCW devices are not affected.

 * This is s390x-specific only, hence will not affect any other architecture.

Frank Heimes (1):
  UBUNTU: [Config] Set IOMMU_DEFAULT_DMA_STRICT=n and
    IOMMU_DEFAULT_DMA_LAZY=y for s390x

 debian.master/config/annotations | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

-- 
2.43.0