[U-Boot] [PATCH 1/2] arm: socfpga: mmc: Enable calibration for drvsel and smpsel

[Marek Vasut <marex@denx.de>]

On Wednesday, August 19, 2015 at 10:21:17 AM, Chin Liang See wrote:
> Hi,

Hi again,

> On Wed, 2015-08-19 at 02:41 +0000, marex at denx.de wrote:
> > On Wednesday, August 19, 2015 at 07:54:50 AM, Chin Liang See wrote:
> > > Enable SDMMC calibration to determine the best setting for
> > > drvsel and smpsel. It will be triggered whenever there is
> > > a change of card frequency and bus width. This is to ensure
> > > reliable transmission between the controller and the card.
> > > 
> > > Signed-off-by: Chin Liang See <clsee@altera.com>
> > > Cc: Dinh Nguyen <dinguyen@opensource.altera.com>
> > > Cc: Pavel Machek <pavel@denx.de>
> > > Cc: Marek Vasut <marex@denx.de>
> > > Cc: Wolfgang Denk <wd@denx.de>
> > > Cc: Stefan Roese <sr@denx.de>
> > > Cc: Tom Rini <trini@konsulko.com>
> > > ---
> > > 
> > >  drivers/mmc/socfpga_dw_mmc.c |  187
> > > 
> > > ++++++++++++++++++++++++++++++++++++++++-- 1 files changed, 180
> > > insertions(+), 7 deletions(-)
> > > 
> > > diff --git a/drivers/mmc/socfpga_dw_mmc.c
> > > b/drivers/mmc/socfpga_dw_mmc.c index eb69aed..15e537c 100644
> > > --- a/drivers/mmc/socfpga_dw_mmc.c
> > > +++ b/drivers/mmc/socfpga_dw_mmc.c
> > > @@ -11,25 +11,140 @@
> > > 
> > >  #include <asm/arch/dwmmc.h>
> > >  #include <asm/arch/clock_manager.h>
> > >  #include <asm/arch/system_manager.h>
> > > 
> > > +#include "mmc_private.h"
> > > 
> > >  static const struct socfpga_clock_manager *clock_manager_base =
> > >  
> > >  		(void *)SOCFPGA_CLKMGR_ADDRESS;
> > >  
> > >  static const struct socfpga_system_manager *system_manager_base =
> > >  
> > >  		(void *)SOCFPGA_SYSMGR_ADDRESS;
> > > 
> > > -static void socfpga_dwmci_clksel(struct dwmci_host *host)
> > > +#define CAL_ROWS     7
> > > +#define CAL_COLS     8
> > > +
> > > +/*
> > > + * This function determines the largest rectangle filled with 1's and
> > > returns + * the middle. This functon can be optimized, for example
> > > using the algorithm + * from
> > > http://www.drdobbs.com/database/the-maximal-rectangle-problem/184410529
> > > + * It currently takes less than 1ms, while creating the input data
> > > takes ~5ms + * so there is not a real need to optimize it.
> > > + */
> > > +int find_calibration_point(unsigned char
> > > cal_results[CAL_ROWS][CAL_COLS], +unsigned int sum, unsigned int *
> > > cal_row, unsigned int * cal_col)
> > > 
> > >  {
> > > 
> > > -	unsigned int drvsel;
> > > -	unsigned int smplsel;
> > > +	/* Structure containing a rectangle candidate */
> > > +	typedef struct
> > > +	{
> > > +		unsigned char height;
> > > +		unsigned char width;
> > > +		unsigned short area;
> > > +	} rect_cand_t;
> > > +
> > > +	/* Array with the rectangle candidates */
> > > +	rect_cand_t rect_cands[CAL_ROWS * CAL_COLS];
> > > +	unsigned char cr_rect_cand = 0;
> > > +	unsigned char height, width, k;
> > > +
> > > +	/* No solution if all combinations fail */
> > > +	if(sum == 0)
> > > +		return 1;
> > > +
> > > +	/* Simple solution if all combinations pass */
> > > +	if(sum == (CAL_ROWS * CAL_COLS)) {
> > > +		*cal_row = (CAL_ROWS - 1) / 2;
> > > +		*cal_col = (CAL_COLS - 1) / 2;
> > > +		return 0;
> > > +	}
> > > +
> > > +	/*
> > > +	 * Create list of all possible sub-rectangles, in descending area
> > > +	 * order
> > > +	 */
> > > +	for(height = CAL_ROWS; height >= 1; height--) {
> > > +		for(width = CAL_COLS; width >= 1 ; width--){
> > > +			/* Add a new rectangle candidate */
> > > +			rect_cands[cr_rect_cand].height = height;
> > > +			rect_cands[cr_rect_cand].width = width;
> > > +			rect_cands[cr_rect_cand].area = height * width;
> > > +			cr_rect_cand++;
> > > +
> > > +			/* First candidate it always in the right position */
> > > +			if(cr_rect_cand == 1)
> > > +				continue;
> > > +
> > > +			/*
> > > +			 * Put the candidate in right location to maintain
> > > +			 * descending order
> > > +			 */
> > > +			for(k = cr_rect_cand - 1; k > 1; k--){
> > > +				if(rect_cands[k-1].area < rect_cands[k].area){
> > > +					rect_cand_t tmp = rect_cands[k-1];
> > > +					rect_cands[k-1] = rect_cands[k];
> > > +					rect_cands[k] = tmp;
> > > +				} else
> > > +					break;
> > > +			}
> > > +		}
> > > +	}
> > > +
> > > +	/* Try to fit the rectangle candidates, in descending area order */
> > > +	for(k = 0; k < CAL_ROWS * CAL_COLS; k++) {
> > > +		unsigned char start_row, start_col;
> > > +		unsigned rect_width = rect_cands[k].width;
> > > +		unsigned rect_height = rect_cands[k].height;
> > > +
> > > +		/* Find the row and column where the candidate fits */
> > > +		for (start_col = 0; start_col < (CAL_COLS - rect_width + 1);
> > > +		     start_col++) {
> > > +			for (start_row = 0; start_row < (CAL_ROWS -
> > > +			     rect_height + 1); start_row++) {
> > > +				unsigned ok = 1;
> > > +				unsigned add_col, add_row;
> > > +
> > > +				/* Determine if the rectangle fits here */
> > > +				for (add_col = 0; (add_col < rect_width) && ok;
> > > +				     add_col++) {
> > > +					for (add_row = 0; add_row < rect_height;
> > > +					     add_row++) {
> > > +						if (!cal_results
> > > +						   [start_row+add_row][start_col
> > > +						   + add_col]) {
> > > +							ok = 0;
> > > +							break;
> > > +						}
> > > +					}
> > > +				}
> > > +
> > > +				/*
> > > +				 * Return 'middle' of rectangle in case of
> > > +				 * success
> > > +				 */
> > > +				if (ok) {
> > > +					if(rect_width > 1)
> > > +						rect_width--;
> > > +
> > > +					if(rect_height > 1)
> > > +						rect_height--;
> > > +
> > > +					*cal_row = start_row +
> > > +						   (rect_height / 2);
> > > +					*cal_col = start_col + (rect_width / 2);
> > > +
> > > +					return 0;
> > > +				}
> > > +			}
> > > +		}
> > > +	}
> > > +
> > > +	/* We could not fit any rectangle - return failure */
> > > +	return 1;
> > 
> > Oh btw. please make sure to use errno.h and standard return codes ;
> > -EINVAL in this case I believe.
> 
> Good suggestion, will fix in v2.
> Thanks

Thanks. I was curious, is this really socfpga specific or can this calibration
be used on dwmmc in general -- thus on exynos and rockchip systems -- as well?
If it's generic to dwmmc, this should be moved into the dwmmc core code. Also,
I am CCing Simon, he's been plumbing in the DWMMC recently.

Marek Vasut