Re: More serial ports in EFI bootstrap?

[Emmanuel Dreyfus <manu%netbsd.org@localhost>]

On Fri, May 16, 2025 at 07:39:31AM +0000, Emmanuel Dreyfus wrote:
> I have a Dell server with no serial port, and a serial console is highly
> desirable for system administration. I wonder if EFI bootstrap could
> be modified to use a serial port from a PCIe to serial card, or from a
> USB to serial adapter.

Replying to myself to tell some succes.

The UEFI bootstrap uses UEFI for serial I/O, hence we need UEFI to 
have a driver for the serial port in order to use it in NetBSD EFI
bootstrap.

There is such a driver for a unique USB-to-serial chip (VendorID
0x0403, productId 0x6001). The FtdiUsbSerialDxe driver can be (painfully)
built from Tianocore/EDK2. The instruction are here. I only built
OvmfPkg:
https://github.com/tianocore/edk2-platforms

Once done, I have:
Build/OvmfX64/DEBUG_GCC/X64/Shell.efi
Build/OptionRomPkg/DEBUG_GCC/X64/FtdiUsbSerialDxe.efi

Shell.efi goes in EFI/boot on the EFI partition. FtdiUsbSerialDxe.efi
goes wherever you want, for instance in EFI/EDK2
Next, I tell the machine to boot from file using EFI/boot/Shell.efi
In the UEFI shell I type:
FS0:
load \EFI\EDK2\FtdiUsbSerialDxe.efi

That causes a new serial port to be reported by the sermode command
in the UEFI shell. I should be able to connect it to the UEFI console, 
but I got no success here. By typing bootx64.efi I exit the UEFI
shell and load the NetBSD EFI bootstrap (the setps in UEFI shell 
can be made automatic using /EFI/boot/startup.nsh)

The NetBSD EFI bootstrap looks up devices implementing the SerialIO
protocol, and then filters out anything not on an EISA bus. I removed
that filtering:
--- eficons.c   14 Sep 2023 03:05:15 -0000      1.14
+++ eficons.c   4 Jun 2025 06:59:24 -0000
@@ -819,9 +819,13 @@

                        if (DevicePathType(dp) == ACPI_DEVICE_PATH &&
                            DevicePathSubType(dp) == ACPI_DP) {
                                dpp = (EFI_DEV_PATH_PTR)dp;
+#if 0
                                if (dpp.Acpi->HID == EISA_PNP_ID(0x0501)) {
+#else
+                               {
+#endif
                                        uid = dpp.Acpi->UID;
                                        break;
                                }
                        }

consdev now show only "pc com0" while the unpatched boostrap
did show "pc com0 com1". But this single com0 is the right one. 
If I type consdev com0, I get a working serial console at 9600 bps.

This is great, but the next step is booting the kernel. It does not
support the console on the USB-to-serial adapter and I boot to 
multiuser without a console.

Hence the new question is: do we have any hope of using a USB-to-serial
adapter for kernel console? I have not looked at it yet, but I understand
there will be the issue that the USB-to-serial adapter cannot work 
before a lot of devices initialization. It would be nice to have it 
working for single-user mode, and that should be possible.

-- 
Emmanuel Dreyfus
manu%netbsd.org@localhost