lkml.org 
[lkml]   [2011]   [Jan]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[PATCH 0/5] mmc: add double buffering for mmc block requests
Date
Add support to prepare one MMC request while another is active on
the host. This is done by making the issue_rw_rq() asynchronous.
The increase in throughput is proportional to the time it takes to
prepare a request and how fast the memory is. The faster the MMC/SD is
the more significant the prepare request time becomes. Measurements on U5500
and U8500 on eMMC shows significant performance gain for DMA on MMC for large
reads. In the PIO case there is some gain in performance for large reads too.
There seems to be no or small performance gain for write, don't have a good
explanation for this yet.

There are two optional hooks pre_req() and post_req() that the host driver
may implement in order to improve double buffering. In the DMA case pre_req()
may do dma_map_sg() and prepare the dma descriptor and post_req runs the
dma_unmap_sg.

The mmci host driver implementation for double buffering is not intended
nor ready for mainline yet. It is only an example of how to implement
pre_req() and post_req(). The reason for this is that the basic DMA support
for MMCI is not complete yet. The mmci patches are sent in a separate patch
series "[FYI 0/4] arm: mmci: example implementation of double buffering".

Issues/Questions for issue_rw_rq() in block.c:
* Is it safe to claim the host for the first MMC request and wait to release
it until the MMC queue is empty again? Or must the host be claimed and
released for every request?
* Is it possible to predict the result from __blk_end_request().
If there are no errors for a completed MMC request and the
blk_rq_bytes(req) == data.bytes_xfered, will it be guaranteed that
__blk_end_request will return 0?

Here follows the IOZone results for u8500 v1.1 on eMMC.
The numbers for DMA are a bit to good here due to the fact that the
CPU speed is decreased compared to u8500 v2. This makes the cache handling
even more significant.

Command line used: ./iozone -az -i0 -i1 -i2 -s 50m -I -f /iozone.tmp -e -R -+u

Relative diff: VANILLA-MMC-PIO -> 2BUF-MMC-PIO
cpu load is abs diff
random random
KB reclen write rewrite read reread read write
51200 4 +0% +0% +0% +0% +0% +0%
cpu: +0.1 -0.1 -0.5 -0.3 -0.1 -0.0

51200 8 +0% +0% +6% +6% +8% +0%
cpu: +0.1 -0.1 -0.3 -0.4 -0.8 +0.0

51200 16 +0% -2% +0% +0% -3% +0%
cpu: +0.0 -0.2 +0.0 +0.0 -0.2 +0.0

51200 32 +0% +1% +0% +0% +0% +0%
cpu: +0.1 +0.0 -0.3 +0.0 +0.0 +0.0

51200 64 +0% +0% +0% +0% +0% +0%
cpu: +0.1 +0.0 +0.0 +0.0 +0.0 +0.0

51200 128 +0% +1% +1% +1% +1% +0%
cpu: +0.0 +0.2 +0.1 -0.3 +0.4 +0.0

51200 256 +0% +0% +1% +1% +1% +0%
cpu: +0.0 -0.0 +0.1 +0.1 +0.1 +0.0

51200 512 +0% +1% +2% +2% +2% +0%
cpu: +0.1 +0.0 +0.2 +0.2 +0.2 +0.1

51200 1024 +0% +2% +2% +2% +3% +0%
cpu: +0.2 +0.1 +0.2 +0.5 -0.8 +0.0

51200 2048 +0% +2% +3% +3% +3% +0%
cpu: +0.0 -0.2 +0.4 +0.8 -0.5 +0.2

51200 4096 +0% +1% +3% +3% +3% +1%
cpu: +0.2 +0.1 +0.9 +0.9 +0.5 +0.1

51200 8192 +1% +0% +3% +3% +3% +1%
cpu: +0.2 +0.2 +1.3 +1.3 +1.0 +0.0

51200 16384 +0% +1% +3% +3% +3% +1%
cpu: +0.2 +0.1 +1.0 +1.3 +1.0 +0.5

Relative diff: VANILLA-MMC-DMA -> 2BUF-MMC-MMCI-DMA
cpu load is abs diff
random random
KB reclen write rewrite read reread read write
51200 4 +0% -3% +6% +5% +5% +0%
cpu: +0.0 -0.2 -0.6 -0.1 +0.3 +0.0

51200 8 +0% +0% +7% +7% +7% +0%
cpu: +0.0 +0.1 +0.8 +0.6 +0.9 +0.0

51200 16 +0% +0% +7% +7% +8% +0%
cpu: +0.0 -0.0 +0.7 +0.7 +0.8 +0.0

51200 32 +0% +0% +8% +8% +9% +0%
cpu: +0.0 +0.1 +0.7 +0.7 +0.3 +0.0

51200 64 +0% +1% +9% +9% +9% +0%
cpu: +0.0 +0.0 +0.8 +0.7 +0.8 +0.0

51200 128 +1% +0% +13% +13% +14% +0%
cpu: +0.2 +0.0 +1.0 +1.0 +1.1 +0.0

51200 256 +1% +2% +8% +8% +11% +0%
cpu: +0.0 +0.3 +0.0 +0.7 +1.5 +0.0

51200 512 +1% +2% +16% +16% +17% +0%
cpu: +0.2 +0.2 +2.2 +2.1 +2.2 +0.1

51200 1024 +1% +2% +20% +20% +20% +1%
cpu: +0.2 +0.1 +2.6 +1.9 +2.6 +0.0

51200 2048 +0% +2% +22% +22% +21% +0%
cpu: +0.0 +0.3 +2.3 +2.9 +2.1 -0.0

51200 4096 +1% +2% +23% +23% +23% +1%
cpu: +0.2 +0.1 +2.0 +3.2 +3.1 +0.0

51200 8192 +1% +5% +24% +24% +24% +1%
cpu: +1.4 -0.0 +4.2 +3.0 +2.8 +0.1

51200 16384 +1% +3% +24% +24% +24% +2%
cpu: +0.0 +0.3 +3.4 +3.8 +3.7 +0.1

Here follows the IOZone results for u5500 on eMMC.
These numbers for DMA are more as expected.

Command line used: ./iozone -az -i0 -i1 -i2 -s 50m -I -f /iozone.tmp -e -R -+u

Relative diff: VANILLA-MMC-DMA -> 2BUF-MMC-MMCI-DMA
cpu load is abs diff
random random
KB reclen write rewrite read reread read write
51200 128 +1% +1% +10% +9% +10% +0%
cpu: +0.1 +0.0 +1.3 +0.1 +0.8 +0.1

51200 256 +2% +2% +7% +7% +9% +0%
cpu: +0.1 +0.4 +0.5 +0.6 +0.7 +0.0

51200 512 +2% +2% +12% +12% +12% +1%
cpu: +0.4 +0.6 +1.8 +2.4 +2.4 +0.2

51200 1024 +2% +3% +14% +14% +14% +0%
cpu: +0.3 +0.1 +2.1 +1.4 +1.4 +0.2

51200 2048 +3% +3% +16% +16% +16% +1%
cpu: +0.2 +0.2 +2.5 +1.8 +2.4 -0.2

51200 4096 +3% +3% +17% +17% +18% +3%
cpu: +0.1 -0.1 +2.7 +2.0 +2.7 -0.1

51200 8192 +3% +3% +18% +18% +18% +3%
cpu: -0.1 +0.2 +3.0 +2.3 +2.2 +0.2

51200 16384 +3% +3% +18% +18% +18% +4%
cpu: +0.2 +0.2 +2.8 +3.5 +2.4 -0.0

Per Forlin (5):
mmc: add member in mmc queue struct to hold request data
mmc: Add a block request prepare function
mmc: Add a second mmc queue request member
mmc: Store the mmc block request struct in mmc queue
mmc: Add double buffering for mmc block requests

drivers/mmc/card/block.c | 337 ++++++++++++++++++++++++++++++----------------
drivers/mmc/card/queue.c | 171 +++++++++++++++---------
drivers/mmc/card/queue.h | 31 +++-
drivers/mmc/core/core.c | 77 +++++++++--
include/linux/mmc/core.h | 7 +-
include/linux/mmc/host.h | 8 +
6 files changed, 432 insertions(+), 199 deletions(-)



\
 
 \ /
  Last update: 2011-01-12 19:17    [W:0.108 / U:2.156 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site