https://community.nxp.com/t5/P-Series/What-is-max-P2020-PCI-DMA-window-size-for-Linux/m-p/238775#M371 <HTML><HEAD></HEAD><BODY><P>I have a P2020 based system with 4GB RAM.&nbsp; I have been testing mSATA SSD performance which involves a x4 PCIe connection to a SATA controller.&nbsp; If I run the board in 2GB RAM mode I get much better performance than in 4GB RAM mode, 381 MB/sec vs. only 219 MB/sec.&nbsp; I noticed that in 4GB mode, I get some extra kernel log entries that appear to originate from the linux source file arch/powerpc/sysdev/fsl_pci.c, "Outbound window cfg leaves gaps in memory map...", "DMA window size is 0x8000000" (2GB).&nbsp; I believe the mSATA performance degradation is due to the use of SWIOTLB bounce buffers for PCI DMA when in 4GB mode.</P><P></P><P>The current Linux BSP for the board is only at kernel version 3.0.3.&nbsp; I came across the following downstream patch for fsl_pci.c:</P><P></P><P><A href="http://linuxppc.10917.n7.nabble.com/PATCH-powerpc-fsl-pci-Allow-64-bit-PCIe-devices-to-DMA-to-any-memory-address-td20919.html">http://linuxppc.10917.n7.nabble.com/PATCH-powerpc-fsl-pci-Allow-64-bit-PCIe-devices-to-DMA-to-any-memory-address-td20919.html</A></P><P></P><P>Adding the code from this patch to the 3.0.3 version of fsl_pci.c significantly improved the 4GB mode mSATA performance to about 370 MB/sec, so not quite as good as when in 2GB mode.</P><P></P><P>I would like to know if it is possible to modify the memory map somehow so I end up with a 4GB PCI DMA window size instead of only 2GB.&nbsp; My assumption is that may provide 4GB mode performance as least as good as 2GB mode.</P></BODY></HTML> Fri, 02 Aug 2013 23:13:56 GMT rkj 2013-08-02T23:13:56Z https://community.nxp.com/t5/P-Series/What-is-max-P2020-PCI-DMA-window-size-for-Linux/m-p/238775#M371 <HTML><HEAD></HEAD><BODY><P>I have a P2020 based system with 4GB RAM.&nbsp; I have been testing mSATA SSD performance which involves a x4 PCIe connection to a SATA controller.&nbsp; If I run the board in 2GB RAM mode I get much better performance than in 4GB RAM mode, 381 MB/sec vs. only 219 MB/sec.&nbsp; I noticed that in 4GB mode, I get some extra kernel log entries that appear to originate from the linux source file arch/powerpc/sysdev/fsl_pci.c, "Outbound window cfg leaves gaps in memory map...", "DMA window size is 0x8000000" (2GB).&nbsp; I believe the mSATA performance degradation is due to the use of SWIOTLB bounce buffers for PCI DMA when in 4GB mode.</P><P></P><P>The current Linux BSP for the board is only at kernel version 3.0.3.&nbsp; I came across the following downstream patch for fsl_pci.c:</P><P></P><P><A href="http://linuxppc.10917.n7.nabble.com/PATCH-powerpc-fsl-pci-Allow-64-bit-PCIe-devices-to-DMA-to-any-memory-address-td20919.html">http://linuxppc.10917.n7.nabble.com/PATCH-powerpc-fsl-pci-Allow-64-bit-PCIe-devices-to-DMA-to-any-memory-address-td20919.html</A></P><P></P><P>Adding the code from this patch to the 3.0.3 version of fsl_pci.c significantly improved the 4GB mode mSATA performance to about 370 MB/sec, so not quite as good as when in 2GB mode.</P><P></P><P>I would like to know if it is possible to modify the memory map somehow so I end up with a 4GB PCI DMA window size instead of only 2GB.&nbsp; My assumption is that may provide 4GB mode performance as least as good as 2GB mode.</P></BODY></HTML> Fri, 02 Aug 2013 23:13:56 GMT https://community.nxp.com/t5/P-Series/What-is-max-P2020-PCI-DMA-window-size-for-Linux/m-p/238775#M371 rkj 2013-08-02T23:13:56Z