Re: [PATCH v2] mmc: mtk-sd: reduce CIT for better performance

From: AngeloGioacchino Del Regno
Date: Thu Jun 01 2023 - 06:02:04 EST


Il 01/06/23 05:16, Wenbin Mei (梅文彬) ha scritto:
On Wed, 2023-05-31 at 10:18 +0200, AngeloGioacchino Del Regno wrote:
External email : Please do not click links or open attachments until you have verified the sender or the content.

Il 31/05/23 09:32, Wenbin Mei (梅文彬) ha scritto:

On Thu, 2023-05-18 at 11:13 +0200, AngeloGioacchino Del Regno wrote:

External email : Please do not click links or open attachments until

you have verified the sender or the content.





Il 10/05/23 03:58, Wenbin Mei ha scritto:

CQHCI_SSC1 indicates to CQE the polling period to use when using

periodic

SEND_QUEUE_STATUS(CMD13) polling.

The default value 0x1000 that corresponds to 150us, let's decrease

it to



The default value 0x1000 (4096) corresponds to 4096 * 52.08uS =

231.33uS

...so the default is not 150uS.



If I'm wrong, this means that the CQCAP field is not 0, which would

mean

that the expected 3uS would be wrong.



Also, since the calculation can be done dynamically, this is what we

should

actually do in the driver, as this gives information to the next

engineer

checking this piece of code.



Apart from this, by just writing 0x40 to the CQHCI_SSC1 register, you

are

assuming that the CQCAP value requirement is fullfilled, but you

cannot

assume that the bootloader has set the CQCAP's ITCFVAL and ITCFMUL

fields

as you expect on all platforms: this means that implementing this

takes

a little more effort.



You have two ways to implement this:

*** First ***

1. Read ITCFMUL and ITCFVAL, then:

tclk_mul = itcfmul_to_mhz(ITCFMUL); /* pseudo function

interprets reg value*/

tclk = ITCFVAL * tclk_mul;



2. Set SSC1 so that we get 3nS:

#define CQHCI_SSC1_CIT GENMASK(15, 0)

poll_time = cit_time_ns_to_regval(3);

sscit = FIELD_PREP(CQHCI_SSC1_CIT, poll_time)

cqhci_writel( ... )



*** Second **



1. Pre-set ITCFMUL and ITCFVAL to

ITCFVAL = 192 (decimal)

ITCFMUL = 2 (where 2 == 0.1MHz)



2. Set SSC1 so that we get 3nS:

#define CQHCI_SSC1_CIT GENMASK(15, 0)

poll_time = cit_time_ns_to_regval(3);

sscit = FIELD_PREP(CQHCI_SSC1_CIT, poll_time)

cqhci_writel( ... )



I would implement the first way, as it paves the way to extend this

to different

tclk values if needed in the future.



Regards,

Angelo

Hi Angelo,



Sorry for lately reply.



For Mediatek mmc host IP, ITCFMUL is 0x2(0x1MHz), ITVFVAL reports 182,

and these fields are the same and are readonly for all IC, but since

Mediatek CQE uses msdc_hclk(273MHz), CMD13'interval calculation driver

should use 273MHz to get the actual time, so the actual clock is

27.3MHz.




You're right, I've misread the datasheet, just rechecked and it reports RO.


If CIT is 0x1000 by default, CMD idle time: 0x1000 * 1 / 27.3MHz =

around 150us.



In addition the bootloader will not set the CQCAP's ITCFVAL and ITCFMUL

fields, because these fields of CQCAP register is RO(readonly), so we

can ignore the change for the CQCAP's ITCFVAL and ITCFMUL fields.




Yes, that's right, again - this means that you should go for the first

proposed implementation, as future MediaTek SoCs may (or may not) change

that: if you implement as proposed, this is going to be a one-time thing

and future SoCs won't need specific changes.


That implementation also documents the flow about how we're getting to

the actual value, which is important for community people reading this

driver in the future for debugging purposes.


Regards,

Angelo



Thanks for your proposal.


I have discussed with our designer, and this fields of CQCAP's ITCFVAL and ITCFMUL will not change.
If we add more code for it, these codes will also affect the execution efficiency, even if it has a very
small effect.
I think if it's just for reading convenience, we can add mode comments to make it easier to read the code.
Do you think it's okay to add more comments?


This isn't a performance path, but anyway, if you think that it will be at some
point, you can read the two registers at probe time as part of the MMC_CAP2_CQE
if branch, and then cache the invariable values to `struct msdc_host`: this
will make you able to never perform register reads for ITCFVAL/FMUL in
msdc_cqe_enable(), resolving the efficiency issue.

Even better, instead of caching ITCFVAL/FMUL to two variables, since the idle
timer value likely won't ever change during runtime, you can directly perform
the calculation for SSC1 at probe time and cache that value instead, so that
in msdc_cqe_enable() you will have something like...

/* Set the send status command idle timer */
cqhci_writel(cq_host, host->cq_ssc1_time, CQHCI_SSC1);

where cq_ssc1_time is
struct msdc_host {
.......
u32 cq_ssc1_time;
....
}

and where your probe function is

static int msdc_drv_probe(struct platform_device *pdev)
{
......

if (mmc->caps2 & MMC_CAP2_CQE) {
host->cq_host = ......
........
read itcfval;
read itcfmul;
host->cq_ssc1_time = calculated-value;
........
}

.......
}

Regards,
Angelo


Begards,
Wenbin

Thanks

Wenbin



0x40 that corresponds to 3us, which can improve the performance of

some

eMMC devices.



Signed-off-by: Wenbin Mei <wenbin.mei@xxxxxxxxxxxx>

---

drivers/mmc/host/mtk-sd.c | 4 ++++

1 file changed, 4 insertions(+)



diff --git a/drivers/mmc/host/mtk-sd.c b/drivers/mmc/host/mtk-sd.c

index edade0e54a0c..ffeccddcd028 100644

--- a/drivers/mmc/host/mtk-sd.c

+++ b/drivers/mmc/host/mtk-sd.c

@@ -2453,6 +2453,7 @@ static void msdc_hs400_enhanced_strobe(struct

mmc_host *mmc,

static void msdc_cqe_enable(struct mmc_host *mmc)

{

struct msdc_host *host = mmc_priv(mmc);

+ struct cqhci_host *cq_host = mmc->cqe_private;



/* enable cmdq irq */

writel(MSDC_INT_CMDQ, host->base + MSDC_INTEN);

@@ -2462,6 +2463,9 @@ static void msdc_cqe_enable(struct mmc_host

*mmc)

msdc_set_busy_timeout(host, 20 * 1000000000ULL, 0);

/* default read data timeout 1s */

msdc_set_timeout(host, 1000000000ULL, 0);

+

+ /* decrease the send status command idle timer to 3us */

+ cqhci_writel(cq_host, 0x40, CQHCI_SSC1);

}



static void msdc_cqe_disable(struct mmc_host *mmc, bool recovery)