Re: [PATCH 3/3] fs: fat: Limit transfer size to prevent SDHCI controller timeout

[Balaji Selvanathan <balaji.selvanathan@oss.qualcomm.com>]

On 3/24/2026 5:32 PM, Jonas Karlman wrote:
> Hi,
>
> On 3/24/2026 6:52 AM, Balaji Selvanathan wrote:
>> Some SDHCI controllers have limitations on maximum transfer sizes
>> and can timeout on large block operations.
>>
>> Break large read and write operations into chunks of 16384 blocks
>> to stay within controller limits. This prevents timeout errors during
>> large file transfers on FAT filesystems while maintaining
>> compatibility with all platforms.
>>
>> This patch builds on top of the work from:
>> https://lore.kernel.org/u-boot/20260224035000.1617869-1-varadarajan.narayanan@oss.qualcomm.com/
>>
>> Signed-off-by: Balaji Selvanathan <balaji.selvanathan@oss.qualcomm.com>
>> ---
>>   fs/fat/fat.c | 29 ++++++++++++++++++++---------
>>   1 file changed, 20 insertions(+), 9 deletions(-)
>>
>> diff --git a/fs/fat/fat.c b/fs/fat/fat.c
>> index 85b511f75af..6ec241af584 100644
>> --- a/fs/fat/fat.c
>> +++ b/fs/fat/fat.c
>> @@ -30,6 +30,9 @@
>>   /* maximum number of clusters for FAT12 */
>>   #define MAX_FAT12	0xFF4
>>   
>> +/* maximum blocks per read/write to avoid SDHCI timeout */
>> +#define MAX_BLOCKS_PER_TRANSFER 16384
>> +
>>   /*
>>    * Convert a string to lowercase.  Converts at most 'len' characters,
>>    * 'len' may be larger than the length of 'str' if 'str' is NULL
>> @@ -136,18 +139,26 @@ static int disk_rw(__u32 sect, __u32 nr_sect, void *buf, bool read)
>>   	if (rem > blksz) {
>>   		n = rem / blksz;
>>   
>> -		if (read)
>> -			ret = blk_dread(cur_dev, start + s, n, buf);
>> -		else
>> -			ret = blk_dwrite(cur_dev, start + s, n, buf);
>> +		/* Break large reads/writes into chunks */
>> +		while (n > 0) {
>> +			__u32 transfer_count = (n > MAX_BLOCKS_PER_TRANSFER) ? MAX_BLOCKS_PER_TRANSFER : n;
> Have you tried to configure SYS_MMC_MAX_BLK_COUNT=16384 for your
> platform, or otherwise set b_max in your mmc driver?
>
> That should cause similar max blocks handling at a the mmc-uclass level.
>
> Regards,
> Jonas

Hi Jonas,

Thanks for the feedback. Addressed this in v2: 
https://lore.kernel.org/u-boot/20260421-emmc-v2-0-0ebd3322b676@oss.qualcomm.com/.

Thanks,

Balaji

>
>>   
>> -		if (ret != n) {
>> -			ret = -1;
>> -			goto exit;
>> +			if (read)
>> +				ret = blk_dread(cur_dev, start + s, transfer_count, buf);
>> +			else
>> +				ret = blk_dwrite(cur_dev, start + s, transfer_count, buf);
>> +
>> +			if (ret != transfer_count) {
>> +				ret = -1;
>> +				goto exit;
>> +			}
>> +
>> +			buf += transfer_count * blksz;
>> +			s += transfer_count;
>> +			n -= transfer_count;
>>   		}
>> -		buf += n * blksz;
>> +
>>   		rem = rem % blksz;
>> -		s += n;
>>   	}
>>   
>>   	/* Do part 3, read a block and copy the trailing sectors */
>>