Intel® Acceleration Stack for Intel® Xeon® CPU with FPGAs Version 1.2.1 Release Notes

ID 683480
Date 3/06/2020
Public

Known Issues for the Intel Acceleration Stack v1.2.1

Table 4.  Known Issues for the Intel Acceleration Stack v1.2.1
Known Issue Details
PCIe directed speed changes are not supported. Only automatic down-training at boot time is supported.
When an AFU workload exceeds board management controller sensor thresholds, the server may reboot.
  • An AFU workload may exceed the board management controller threshold before pacd can identify and act on the threshold violation. If this situation occurs, the board management controller removes the card from the PCIe bus and the server reboots before the pacd can perform a graceful shutdown.
  • Workaround: None.
  • Status: No planned fix.
Inconsistency in green_bs input port naming.
  • The names of the clock port inputs to the green_bs block do not correspond to resulting frequency provided from the dcp_io_pll block.
    dcp_io_pll port frequency provided to green_bs block clock input green_bs clock input port name
    50 MHz Clk_100
    100 MHz Clk_200
    200 MHz Clk_400
    The file can be found at a10_gx_pac_ias_1_2_1_pv/hw/lib/build/platform/green_bs.sv.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.

setup_permissions.sh in OpenCL* prints error when connected to a virtual function in a virtual machine.
  • When you run the setup_permissions.sh script in a virtual machine to set permissions and system parameters for OpenCL* , the following error displays:
    chmod: cannot access ‘/dev/intel-fpga-fme.*’: No such file or directory.
  • Virtual function does not have access to change permissions for FME sysfs entry files.
  • Workaround: None.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.

The fpgainfo error First Malformed Req returns 0xFFFFFFFFFFFFFFFF for streaming_dma_afu.gbs or dma_afu.gbs.
  • If you configure the streaming_dma_afu.gbs or dma_afu.gbs using fpgaconfig and then run fpgainfo errors, the command returns a First Malformed Req error of 0xFFFFFFFFFFFFFFFF. This error does not affect the functioning of the dma_afu or streaming_dma_afu block.
  • Workaround: None.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
Virtual Function (VF) may fail to attach or detach when using the Linux* Red Hat* 3.10 kernel. The VF failure to attach or detach is a known issue with qemu/kvm and libvirt. Refer to the Red Hat* website for more information about this issue.
Packager issue with Ubuntu and compiling OpenCL kernels.
  • Depending on the host environment setup, the packager tool may fail to run. This can prevent OpenCL kernel compiles on Ubuntu. Compiling OpenCL kernel in Ubuntu 18.04 fails with the following error:

    ERROR: packager tool failed to run. Check installation. Aborting compilation!

  • Workaround: Edit the following file:
    $OPAE_PLATFORM_ROOT/opencl/opencl_bsp/hardware/pac_a10/build/run.sh

    Change line: ADAPT_PACKAGER_BIN="python ./tools/packager.pyz" to ADAPT_PACKAGER_BIN="packager".

  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
Flash fallback does not meet PCIe timeout.
  • The host may hang or report a PCIe failure when the user image in flash is corrupted and the configuration subsystem loads the factory image into the FPGA.
  • Workaround: Run the following super-rsu command to update the FPGA User image and BMC bootloader and firmware:
    • For RHEL:
      sudo super-rsu /usr/share/opae/a10*/super*/*.json
    • For Ubuntu:
      sudo super-rsu /usr/share/opae/a10-gx-pac/super-rsu/base/rsu-09C4.json
  • Status: No planned fix.
Unsupported transaction layer packet types.
  • The Acceleration Stack FPGA Interface Manager (FIM) does not support PCIe* Memory Read Lock, Configuration Read Type 1, and Configuration Write Type 1 transaction layer packets (TLPs). If the device receives a PCIe* packet of this type, it does not respond with a Completion packet as expected.
  • Workaround: None.
  • Status: No planned fix.
Board management controller does not verify the CRC in a PLDM request.
  • The board management controller responds to PLDM packets regardless of the CRC value.
  • Workaround: None.
  • Status: No planned fix.
Occasionally, fpgainfo may not show the BMC version.
  • Workaround: Power cycle the server.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
Limitation after updating the BMC image for 1000 times.
  • After updating the BMC image in the on-chip flash for 1000 times, the subsequent use of fpgasupdate will result in an I/O error.
  • Workaround: Wait for 60 seconds before using the fpgasupdate to update the BMC image.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
When pacd is triggered due to any of the sensors crossing the threshold, an error message is displayed during pacd process.
  • Error messages:
    libopae-c sysfs.c:304:sysfs_write_u64() **ERROR** : Failed to write
    libopae-c reconf.c:217:clear_port_errors() **ERROR** : Failed to clear port errors
    libopae-c reconf.c:358:fpgaReconfigureSlot() **ERROR** : Failed to clear port errors
    
    You can ignore these error messages, they do no affect the functionality of pacd.
  • Workaround: None.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
OpenCL kernels compiled for the Intel FPGA PAC with Intel® Arria® 10 GX may contain hold time violations.
The UNR threshold for 1.2v current reported by pacd is greater than the actual UNR for the sensor.
  • This causes pacd to not handle graceful thermal shutdown as expected if UNR threshold for 1.2v current sensor is reached.
  • Workaround: None.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
Incorrect message in /var/log/messages.
  • You will observe the following messages:
    intel-pac-hssi intel-pac-hssi.0.auto: The HSSI config sysfs file is deprecated
    intel-pac-hssi intel-pac-hssi.0.auto: Please use the config_qsfp0 file instead
    These messages can be ignored.
  • Workaround: Continue to use hssi_mgmt/config sysfs file with the provided ethernet design examples.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
Graceful shutdown is not supported when using a virtual machine.
  • When a hypervisor is connected to a virtual function on a card, Intel® PAC daemon (pacd) cannot perform graceful shutdown when the sensor thresholds are exceeded. In this case, the hypervisor and server reboot. This is a known limitation.
  • Workaround: None.
  • Status: No future plan to fix.

For more information, refer to the Intel Acceleration Stack Quick Start Guide for Intel Programmable Acceleration Card with Intel Arria 10 GX FPGA.

When virtual functions (VF) or physical functions (PF) send multiple unaligned read or write requests to the Intel® PAC with Intel® Arria® 10 GX FPGA registers, the host reboots.
  • This issue only exists when using Red Hat* Enterprise Linux* (RHEL) version 7.6 Kernel 3.10.0-957.
  • Workaround: None available.
  • Status: This limitation is addressed in newer Linux* kernels. Intel recommends upgrading from the Red Hat* Enterprise Linux* (RHEL) version 7.6 Kernel 3.10.0-957 kernel. No fixes are planned for this version of the kernel.
A typo in board_env.xml, listing the Intel® Quartus® Prime software version as 17.1.
  • This issue does not cause any failures and can be ignored.
  • Workaround: None.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
FPGA Client Driver (FCD) file names for Intel® PAC with Intel® Arria® 10 GX FPGA and Intel FPGA PAC D5005 are identical.
  • This may cause an error in linking the memory-mapped device (MMD) libraries if you change the installation of the Intel® PAC with Intel® Arria® 10 GX FPGA to Intel FPGA PAC D5005 or vice versa.
    To avoid this issue, ensure that the FCD file points to the MMD library (libintel_opae_mmd.so) from the intended opencl_bsp directory. Run the following command to verify:
    cat /opt/Intel/OpenCL/Boards/dcp_bsp.fcd
    Desired output:
    $OPAE_PLATFORM_ROOT/opencl/opencl_bsp/linux64/lib/libintel_opae_mmd.so
    where the $OPAE_PLATFORM_ROOT must point to your intended installation directory. For more information about installing the OpenCL BSP and setting up FCD file, refer to the OpenCL on Intel Programmable Acceleration Card with Intel Arria 10 GX FPGA Quick Start User Guide.
  • Workaround: None.
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
The command aocl diagnose reports Installable Client Driver (ICD) test as failed.
  • Workaround: Manually run the following two commands to copy the ICD files:
    sudo cp $INTELFPGAOCLSDKROOT/Altera.icd /etc/OpenCL/vendors/
    sudo cp $INTELFPGAOCLSDKROOT/Intel_FPGA_SSG_Emulator.icd /etc/OpenCL/vendors/
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.
pacd occasionally fails to start after cold or warm reboot.
  • If pacd is started as a systemd service and followed by a warm or cold reboot, pacd can occasionally fail to start.
  • Workaround:
    1. If a warm or cold reboot is performed after setting up pacd as a systemd service, ensure pacd is running after the reboot. Otherwise, follow these steps (you must have root access to perform these steps):
      1. Remove the pacd lock file using:
        sudo rm –f /tmp/pacd.lock
      2. Restart the pacd service with systemctl start pacd.service
    2. Modify the pacd.service file:
      • Replace the line: RestartPreventExitStatus=1 with ExecStopPost=/bin/rm -f /tmp/pacd.lock
  • Status: Fix targeted for a future version of the Intel® Acceleration Stack.