1Rationale
2=========
3
4How significant is the cache maintenance overhead?
5It depends. Fast eMMC and multiple cache levels with speculative cache
6pre-fetch makes the cache overhead relatively significant. If the DMA
7preparations for the next request are done in parallel with the current
8transfer, the DMA preparation overhead would not affect the MMC performance.
9The intention of non-blocking (asynchronous) MMC requests is to minimize the
10time between when an MMC request ends and another MMC request begins.
11Using mmc_wait_for_req(), the MMC controller is idle while dma_map_sg and
12dma_unmap_sg are processing. Using non-blocking MMC requests makes it
13possible to prepare the caches for next job in parallel with an active
14MMC request.
15
16MMC block driver
17================
18
19The mmc_blk_issue_rw_rq() in the MMC block driver is made non-blocking.
20The increase in throughput is proportional to the time it takes to
21prepare (major part of preparations are dma_map_sg() and dma_unmap_sg())
22a request and how fast the memory is. The faster the MMC/SD is the
23more significant the prepare request time becomes. Roughly the expected
24performance gain is 5% for large writes and 10% on large reads on a L2 cache
25platform. In power save mode, when clocks run on a lower frequency, the DMA
26preparation may cost even more. As long as these slower preparations are run
27in parallel with the transfer performance won't be affected.
28
29Details on measurements from IOZone and mmc_test
30================================================
31
32https://wiki.linaro.org/WorkingGroups/Kernel/Specs/StoragePerfMMC-async-req
33
34MMC core API extension
35======================
36
37There is one new public function mmc_start_req().
38It starts a new MMC command request for a host. The function isn't
39truly non-blocking. If there is an ongoing async request it waits
40for completion of that request and starts the new one and returns. It
41doesn't wait for the new request to complete. If there is no ongoing
42request it starts the new request and returns immediately.
43
44MMC host extensions
45===================
46
47There are two optional members in the mmc_host_ops -- pre_req() and
48post_req() -- that the host driver may implement in order to move work
49to before and after the actual mmc_host_ops.request() function is called.
50In the DMA case pre_req() may do dma_map_sg() and prepare the DMA
51descriptor, and post_req() runs the dma_unmap_sg().
52
53Optimize for the first request
54==============================
55
56The first request in a series of requests can't be prepared in parallel
57with the previous transfer, since there is no previous request.
58The argument is_first_req in pre_req() indicates that there is no previous
59request. The host driver may optimize for this scenario to minimize
60the performance loss. A way to optimize for this is to split the current
61request in two chunks, prepare the first chunk and start the request,
62and finally prepare the second chunk and start the transfer.
63
64Pseudocode to handle is_first_req scenario with minimal prepare overhead:
65
66if (is_first_req && req->size > threshold)
67   /* start MMC transfer for the complete transfer size */
68   mmc_start_command(MMC_CMD_TRANSFER_FULL_SIZE);
69
70   /*
71    * Begin to prepare DMA while cmd is being processed by MMC.
72    * The first chunk of the request should take the same time
73    * to prepare as the "MMC process command time".
74    * If prepare time exceeds MMC cmd time
75    * the transfer is delayed, guesstimate max 4k as first chunk size.
76    */
77    prepare_1st_chunk_for_dma(req);
78    /* flush pending desc to the DMAC (dmaengine.h) */
79    dma_issue_pending(req->dma_desc);
80
81    prepare_2nd_chunk_for_dma(req);
82    /*
83     * The second issue_pending should be called before MMC runs out
84     * of the first chunk. If the MMC runs out of the first data chunk
85     * before this call, the transfer is delayed.
86     */
87    dma_issue_pending(req->dma_desc);
88