Hi,
I found if enabled multicore and used a Sandisk Extreme U3 A2 card as
root file system(ext4). Then, running 'stress-ng-aiol' (async I/O
test) would randomly cause SPI crc error in
./drivers/mmc/host/mmc_spi.c and ./drivers/spi/spi-sifive.c.
"mmcblk0: error -84 transferring data end_request: I/O error, dev
mmcblk0, sector"
After looking into the issue, I found the issue is the same as the
commit, "3a855b7ac7d5021674aa3e1cc9d3bfd6b604e9c0".(enable stable
write) The tx_buf is modified during the SPI transfer and causes the
SD card to fail to check CRC.
There are 2 ways I've tested:
a) Passing 'mmc_core.use_spi_crc=0' to the kernel
This way can workaround the issue by disabling CRC check. It can
pass 'stress-ng-aiol' consistently.
b) Copy tx_buf before generating the crc and SPI transfer
/* memcpy sg to another buffer in mmc_spi_data_do()*/
memcpy(page_address(host->backup_pg) + sg->offset, sg_virt(sg), sg->length);
This way also can fix the issue and pass the test. But it has low
performance.
Currently, neither of them is a reasonable fix to me. The previous
commit, 'enabling stable write', seems to fix some cases. Maybe it
means we should have other places being blocked before transfer is
done? I would appreciate any thoughts about where to go next.
Thanks,
-
Green
