On Mon, May 23, 2022 at 02:19:34PM -0500, Zhi Li wrote: > Bjorn: Are you satisfied with the below message? I will fix the > other code in the next version. Looks good, minor questions and tweaks below: > Some PCI Endpoint controllers integrate an eDMA (embedded DMA). > The eDMA controller issues a single bus command per data transfer. Still not sure why this sentence is here. Is it something the patch relies on? Why does it matter how many bus commands there are? > eDMA can bypass the outbound memory address translation unit to > access all RC memory space. > > Add eDMA support for pci-epf-test. > > EPF test can use, depending on HW availability, eDMA or general system > DMA controllers to perform DMA. The test probes the EPF DMA channel > capabilities. Depending on HW availability, the EPF test can use either eDMA or general system DMA controllers to perform DMA. The test tries to use eDMA first and falls back to general system DMA controllers if there's no eDMA. > Separate dma_chan to dma_chan_tx and dma_chan_rx. Search eDMA channel > firstly, then search memory to memory DMA channel if eDMA does not exist. > If general memory to memory channels are used, dma_chan_rx = dma_chan_tx. Search for an eDMA channel first, then search for a memory-to-memory DMA channel ... > Add dma_addr_t dma_remote in pci_epf_test_data_transfer() > because eDMA uses remote RC physical address directly. > > Add enum dma_transfer_direction dir in pci_epf_test_data_transfer() > because eDMA chooses the correct RX/TX channel by dir. > > The overall steps are > > 1. Execute dma_request_channel() and filter function to find correct > eDMA RX and TX Channel. If a channel does not exist, fallback to try to > allocate general memory to memory DMA channel. > 2. Execute dmaengine_slave_config() to configure remote side physical > address. > 3. Execute dmaengine_prep_slave_single() to create transfer descriptor. > 4. Execute tx_submit(). > 5. Execute dma_async_issue_pending()
