UEFI: TC2: booting from A7 core causes boot failures due to MMC/SD timeouts

[Ryan Harkin]

On switching from A15 to A17 booting, UEFI on TC2 become unreliable due to MMC/SD card timeouts.

Testing shows that the board will boot 100% of the time from A15, and approximately 80% of the time from A7.

This was discovered after updating the firmware in LAVA:

BUG: "vexpress-tc2 target boards need a change in firmware to boot on to A7"
https://bugs.launchpad.net/lava-lab/+bug/1171829

[Ryan Harkin (ryanharkin)]

The following timeout message is also thrown intermittently:

MciReadBlockData(): TIMEOUT! Response:0xA404012 Status:0x48

[Ryan Harkin (ryanharkin)]

I have a patch in uefi-next.git that makes this a little better, but still doesn't solve the problem.

Investigating the MMC stack in UEFI, if a block read times out, the stack will abort immediately with no retries. I added a retry loop in a couple of different placed, however, neither worked.

The code in question is:

EmbeddedPkg/Universal/MmcDxe/MmcBlockIo.c, circa line 659:

    if (Transfer == MMC_IOBLOCKS_READ) {
      // Read one block of Data
      Status = MmcHost->ReadBlockData (MmcHost, Lba, This->Media->BlockSize, Buffer);
      gBS->RestoreTPL(Tpl) ;
      if (EFI_ERROR (Status)) {
        DEBUG ((EFI_D_BLKIO, "MmcIoBlocks(): Error Read Block Data and Status = %r\n", Status));
        MmcStopTransmission(MmcHost) ;
        return Status;
      }

And MmcHost->ReadBlockData() is a call into:

ArmPlatformPkg/Drivers/PL180MciDxe/PL180Mci.c, circa line 208, with the timeout code here:

      //Check for error conditions and timeouts
      if (Status & MCI_STATUS_CMD_DATATIMEOUT) {
        DEBUG ((EFI_D_ERROR, "MciReadBlockData(): TIMEOUT! Response:0x%X Status:0x%x\n", MmioRead32 (MCI_RESPONSE0_REG), Status));
        RetVal = EFI_TIMEOUT;
        break;
      } ...

Note, both places make no attempt to handle the timeout, they simply drop out in error.

[Fathi Boudra (fboudra)]

As discussed with Ryan, set priority to medium and remove the targeted milestone.

[Ryan Harkin (ryanharkin)]

This problem no longer happens with the latest code.