In order to do this, I modified the EEPROM bootloader bootloader according to the instructions in the Raspberry Pi documentation. That page is also on GitHub which might be a more stable location. On Raspbian on the RPi:
My updated bootconf.txt is:
[all] BOOT_UART=1 WAKE_ON_GPIO=0 POWER_OFF_ON_HALT=0 DHCP_TIMEOUT=45000 DHCP_REQ_TIMEOUT=4000 TFTP_FILE_TIMEOUT=30000 ENABLE_SELF_UPDATE=1 DISABLE_HDMI=0 BOOT_ORDER=0xf412
I enabled UART debugging, and set the boot order to be: network
0x2, SD card
0x1, USB mass storage
0x4, and finally reboot
0xf. These steps need to be
repeated if the bootloader is updated via apt.
I enabled the TFTP server on my Mac:
I’m not sure if the
enable command is actually necessary. This doesn’t
actually start the
tftpd daemon. Instead, macOS starts the daemon on demand
when it notices an incoming tftp request on the network. Don’t be alarmed!
The tftp server looks for files to serve out of /private/tftpboot, and those
things need to be world
777. By default (this is configurable) the
RPi queries for a directory named by its serial number.
RPi looks for files of various names in that directory, one in particular by the name of start.elf.
Next, I had to update my Ubiquiti router’s DHCP server configuration (on the
command line) to pass a
tftp-server parameter in the DHCP payload. This step
may be optional because you can also set
TFTP_IP in the bootconf.txt above
to specify the IP directly. On my router:
I also gave my Mac a static IP, and renewed the DHCP lease so it took the new IP to make the whole process a little more smooth. Now, it appears the RPi will attempt a TFTP boot, and I see queries in the logs on my Mac.