99 lines · plain
1========================2MMC Asynchronous Request3========================4 5Rationale6=========7 8How significant is the cache maintenance overhead?9 10It depends. Fast eMMC and multiple cache levels with speculative cache11pre-fetch makes the cache overhead relatively significant. If the DMA12preparations for the next request are done in parallel with the current13transfer, the DMA preparation overhead would not affect the MMC performance.14 15The intention of non-blocking (asynchronous) MMC requests is to minimize the16time between when an MMC request ends and another MMC request begins.17 18Using mmc_wait_for_req(), the MMC controller is idle while dma_map_sg and19dma_unmap_sg are processing. Using non-blocking MMC requests makes it20possible to prepare the caches for next job in parallel with an active21MMC request.22 23MMC block driver24================25 26The mmc_blk_issue_rw_rq() in the MMC block driver is made non-blocking.27 28The increase in throughput is proportional to the time it takes to29prepare (major part of preparations are dma_map_sg() and dma_unmap_sg())30a request and how fast the memory is. The faster the MMC/SD is the31more significant the prepare request time becomes. Roughly the expected32performance gain is 5% for large writes and 10% on large reads on a L2 cache33platform. In power save mode, when clocks run on a lower frequency, the DMA34preparation may cost even more. As long as these slower preparations are run35in parallel with the transfer performance won't be affected.36 37Details on measurements from IOZone and mmc_test38================================================39 40https://wiki.linaro.org/WorkingGroups/Kernel/Specs/StoragePerfMMC-async-req41 42MMC core API extension43======================44 45There is one new public function mmc_start_req().46 47It starts a new MMC command request for a host. The function isn't48truly non-blocking. If there is an ongoing async request it waits49for completion of that request and starts the new one and returns. It50doesn't wait for the new request to complete. If there is no ongoing51request it starts the new request and returns immediately.52 53MMC host extensions54===================55 56There are two optional members in the mmc_host_ops -- pre_req() and57post_req() -- that the host driver may implement in order to move work58to before and after the actual mmc_host_ops.request() function is called.59 60In the DMA case pre_req() may do dma_map_sg() and prepare the DMA61descriptor, and post_req() runs the dma_unmap_sg().62 63Optimize for the first request64==============================65 66The first request in a series of requests can't be prepared in parallel67with the previous transfer, since there is no previous request.68 69The argument is_first_req in pre_req() indicates that there is no previous70request. The host driver may optimize for this scenario to minimize71the performance loss. A way to optimize for this is to split the current72request in two chunks, prepare the first chunk and start the request,73and finally prepare the second chunk and start the transfer.74 75Pseudocode to handle is_first_req scenario with minimal prepare overhead::76 77 if (is_first_req && req->size > threshold)78 /* start MMC transfer for the complete transfer size */79 mmc_start_command(MMC_CMD_TRANSFER_FULL_SIZE);80 81 /*82 * Begin to prepare DMA while cmd is being processed by MMC.83 * The first chunk of the request should take the same time84 * to prepare as the "MMC process command time".85 * If prepare time exceeds MMC cmd time86 * the transfer is delayed, guesstimate max 4k as first chunk size.87 */88 prepare_1st_chunk_for_dma(req);89 /* flush pending desc to the DMAC (dmaengine.h) */90 dma_issue_pending(req->dma_desc);91 92 prepare_2nd_chunk_for_dma(req);93 /*94 * The second issue_pending should be called before MMC runs out95 * of the first chunk. If the MMC runs out of the first data chunk96 * before this call, the transfer is delayed.97 */98 dma_issue_pending(req->dma_desc);99