On Sat, Jan 3, 2015 at 4:51 PM, Alim Akhtar [off-list ref] wrote:

Hi Doug / Sonny
New year greetings!!

On Sat, Jan 3, 2015 at 4:37 AM, Doug Anderson [off-list ref] wrote:

quoted

Alim,

On Mon, Dec 22, 2014 at 4:12 AM, Alim Akhtar [off-list ref] wrote:

quoted

From: Seungwon Jeon <redacted>

Even though 1MB is reserved for descriptor table in IDMAC,
the dw_mmc host driver is allowed to receive only maximum
128KB block length in one request. This is caused by setting
improper max_blk_count. It needs to be e adjusted so that
descriptor table is used fully. It is found that the performance
is improved with the increased the max_blk_count.

Signed-off-by: Seungwon Jeon <redacted>
Acked-by: Jaehoon Chung <jh80.chung@samsung.com>
Signed-off-by: Alim Akhtar <alim.akhtar@samsung.com>
---
 drivers/mmc/host/dw_mmc.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/mmc/host/dw_mmc.c b/drivers/mmc/host/dw_mmc.c
index 64ea042..a1b80e5 100644
--- a/drivers/mmc/host/dw_mmc.c
+++ b/drivers/mmc/host/dw_mmc.c

@@ -2332,9 +2332,9 @@ static int dw_mci_init_slot(struct dw_mci *host, unsigned int id)
 #ifdef CONFIG_MMC_DW_IDMAC
                mmc->max_segs = host->ring_size;
                mmc->max_blk_size = 65536;
-               mmc->max_blk_count = host->ring_size;
                mmc->max_seg_size = 0x1000;
-               mmc->max_req_size = mmc->max_seg_size * mmc->max_blk_count;
+               mmc->max_req_size = mmc->max_seg_size * host->ring_size;
+               mmc->max_blk_count = mmc->max_req_size / 512;

The numbers are a little different than the ones you and Sonny arrived
at about a year ago at
<https://chromium-review.googlesource.com/#/c/176488/>

You end up with the same max_req_size I think, but it looks as if
max_blk_count is 0x800 in your series and ix 0xffff in the one from a
year ago.  Do you have any explanation for why it is max_req_size /
512?

Before taking this patch, I went over all the discussions that had
happened in past
http://marc.info/?l=linux-mmc&m=133039478518090&w=2 (way back in 2012)
http://www.spinics.net/lists/linux-mmc/msg25549.html
and
http://www.spinics.net/lists/linux-mmc/msg25797.html

and the last one was having an ACK from maintainer.

Having said that, I did check the performance number with current
patch which has landed and with one which sets max_blk_count as
0xFFFF.
And both given exactly the same numbers on peach-pi (Write: ~52MB/s
and Read: ~146MB/s (HS200) using iozone),

I checked it a bit more and found that in both the case
max_hw_sectors_kb is set to 1024 (max request size to block layer) and
max_hw_sectors_kb is the one which actually affect the Performance
numbers and max_hw_sectors_kb is set as

blk_queue_max_hw_sectors(mq->queue, min(host->max_blk_count,
host->max_req_size / 512));

In case of 64bit IDMAC (one on exynos7) which given a different value
for "ring_size", both the patches set the same max_hw_sectors_kb (512
in this case)

Do you see some performance degradation with some other variants of
dw_mmc host controller with current ToT?

Looks like Doug test it and saw a performance improvement on rk3288
Thanks!

Sonny,
I ran iozone like: iozone -az -i0 -i1 -I -s 10M -f /home/test (on eMMC)

quoted

-Doug

--
Regards,
Alim