Coral is a Chromebook (or really about 20 different Chromebooks) which use the Intel Apollo Lake platform (APL). The ‘reef’ Chromebooks use the same APL SoC so should also work. Some later ones based on Glacier Lake (GLK) need various changes in GPIOs, etc. but are very similar.
It is hoped that this port can enable ports to embedded APL boards which are starting to appear.
Note that booting U-Boot on APL is already supported by coreboot and Slim Bootloader. This documentation refers to a ‘bare metal’ port.
Boot flow - TPL¶
Apollo Lake boots via an IFWI (Integrated Firmware Image). TPL is placed in this, in the IBBL entry.
On boot, an on-chip microcontroller called the CSE (Converged Security Engine) sets up some SDRAM at ffff8000 and loads the TPL image to that address. The SRAM extends up to the top of 32-bit address space, but the last 2KB is the start16 region, so the TPL image must be 30KB at most, and CONFIG_TPL_TEXT_BASE must be ffff8000. Actually the start16 region is small and it could probably move from f800 to fe00, providing another 1.5KB, but TPL is only about 19KB so there is no need to change it at present. The size limit is enforced by CONFIG_TPL_SIZE_LIMIT to avoid producing images that won’t boot.
TPL (running from start.S) first sets up CAR (Cache-as-RAM) which provides larger area of RAM for use while booting. CAR is mapped at CONFIG_SYS_CAR_ADDR (fef00000) and is 768KB in size. It then sets up the stack in the botttom 64KB of this space (i.e. below fef10000). This means that the stack and early malloc() region in TPL can be 64KB at most.
TPL operates without CONFIG_TPL_PCI enabled so PCI config access must use the x86-specific functions pci_x86_write_config(), etc. SPL creates a simple-bus device so that PCI devices are bound by driver model. Then arch_cpu_init_tpl() is called to early init on various devices. This includes placing PCI devices at hard-coded addresses in the memory map. PCI auto-config is not used.
Most of the 16KB ROM is mapped into the very top of memory, except for the Intel descriptor (first 4KB) and the space for SRAM as above.
TPL does not set up a bloblist since at present it does not have anything to pass to SPL.
Once TPL is done it loads SPL from ROM using either the memory-mapped SPI or by using the Intel fast SPI driver. SPL is loaded into CAR, at the address given by CONFIG_SPL_TEXT_BASE, which is normally fef10000.
Note that booting using the SPI driver results in an TPL image that is about 26KB in size instead of 19KB. Also boot speed is worse by about 340ms. If you really want to use the driver, enable CONFIG_APL_SPI_FLASH_BOOT and set BOOT_FROM_FAST_SPI_FLASH to true.
Boot flow - SPL¶
SPL (running from start_from_tpl.S) continues to use the same stack as TPL. It calls arch_cpu_init_spl() to set up a few devices, then init_dram() loads the FSP-M binary into CAR and runs to, to set up SDRAM. The address of the output ‘HOB’ list (Hand-off-block) is stored into gd->arch.hob_list for parsing. There is a 2GB chunk of SDRAM starting at 0 and the rest is at 4GB.
PCI auto-config is not used in SPL either, but CONFIG_SPL_PCI is defined, so proper PCI access is available and normal dm_pci_read_config() calls can be used. However PCI auto-config is not used so the same static memory mapping set up by TPL is still active.
SPL on x86 always runs with CONFIG_SPL_SEPARATE_BSS=y and BSS is at 120000 (see u-boot-spl.lds). This works because SPL doesn’t access BSS until after board_init_r(), as per the rules, and DRAM is available then.
SPL sets up a bloblist and passes the SPL hand-off information to U-Boot proper. This includes a pointer to the HOB list as well as DRAM information. See struct arch_spl_handoff. The bloblist address is set by CONFIG_BLOBLIST_ADDR, normally 100000.
SPL uses SPI flash to update the MRC caches in ROM. This speeds up subsequent boots. Be warned that SPL can take 30 seconds without this cache! This is a known issue with Intel SoCs with modern DRAM and apparently cannot be improved. The MRC caches are used to work around this.
Once SPL is finished it loads U-Boot into SDRAM at CONFIG_SYS_TEXT_BASE, which is normally 1110000. Note that CAR is still active.
Boot flow - U-Boot pre-relocation¶
U-Boot (running from start_from_spl.S) starts running in RAM and uses the same stack as SPL. It does various init activities before relocation. Notably arch_cpu_init_dm() sets up the pin muxing for the chip using a very large table in the device tree.
PCI auto-config is not used before relocation, but CONFIG_PCI of course is defined, so proper PCI access is available. The same static memory mapping set up by TPL is still active until relocation.
As per usual, U-Boot allocates memory at the top of available RAM (a bit below 2GB in this case) and copies things there ready to relocate itself. Notably reserve_arch() does not reserve space for the HOB list returned by FSP-M since this is already located in RAM.
U-Boot then shuts down CAR and jumps to its relocated version.
Boot flow - U-Boot post-relocation¶
U-Boot starts up normally, running near the top of RAM. After driver model is running, arch_fsp_init_r() is called which loads and runs the FSP-S binary. This updates the HOB list to include graphics information, used by the fsp_video driver.
PCI autoconfig is done and a few devices are probed to complete init. Most others are started only when they are used.
Note that FSP-S is supposed to run after CAR has been shut down, which happens immediately before U-Boot starts up in its relocated position. Therefore we cannot run FSP-S before relocation. On the other hand we must run it before PCI auto-config is done, since FSP-S may show or hide devices. The first device that probes PCI after relocation is the serial port, in initr_serial(), so FSP-S must run before that. A corollary is that loading FSP-S must be done without using the SPI driver, to avoid probing PCI and causing an autoconfig, so memory-mapped reading is always used for FSP-S.
It would be possible to tear down CAR in SPL instead of U-Boot. The SPL handoff information could make sure it does not include any pointers into CAR (in fact it doesn’t). But tearing down CAR in U-Boot allows the initial state used by TPL and SPL to be read by U-Boot, which seems useful. It also matches how older platforms start up (those that don’t use SPL).
Bootstage is used through all phases of U-Boot to keep accurate timimgs for boot. Use ‘bootstage report’ in U-Boot to see the report, e.g.:
Timer summary in microseconds (16 records): Mark Elapsed Stage 0 0 reset 155,325 155,325 TPL 204,014 48,689 end TPL 204,385 371 SPL 738,633 534,248 end SPL 739,161 528 board_init_f 842,764 103,603 board_init_r 1,166,233 323,469 main_loop 1,166,283 50 id=175 Accumulated time: 62 fast_spi 202 dm_r 7,779 dm_spl 15,555 dm_f 208,357 fsp-m 239,847 fsp-s 292,143 mmap_spi
CPU performance is about 3500 DMIPS:
=> dhry 1000000 iterations in 161 ms: 6211180/s, 3535 DMIPS
Partial memory map¶
ffffffff Top of ROM (and last byte of 32-bit address space) ffff8000 TPL loaded here (from IFWI) ff000000 Bottom of ROM fefc0000 Top of CAR region fef96000 Stack for FSP-M fef40000 59000 FSP-M fef11000 SPL loaded here fef10000 CONFIG_BLOBLIST_ADDR fef10000 Stack top in TPL, SPL and U-Boot before relocation fef00000 1000 CONFIG_BOOTSTAGE_STASH_ADDR fef00000 Base of CAR region 30000 AP_DEFAULT_BASE (used to start up additional CPUs) f0000 CONFIG_ROM_TABLE_ADDR 120000 BSS (defined in u-boot-spl.lds) 200000 FSP-S (which is run after U-Boot is relocated) 1110000 CONFIG_SYS_TEXT_BASE
- SPI flash
- MMC (dev 0) and micro-SD (dev 1)
- Chrome OS EC
- Finish peripherals
- left-side USB
- Cr50 (security chip: a basic driver is running but not included here)
- Sound (Intel I2S support exists, but need da7219 driver)
- Various minor features supported by LPC, etc.
- Booting Chrome OS, e.g. with verified boot
- Integrate with Chrome OS vboot
- Improvements to booting from coreboot (i.e. as a coreboot target)
- Use FSP-T binary instead of our own CAR implementation
- Use the official FSP package instead of the coreboot one
- Enable all CPU cores
- Suspend / resume
This is a spare-time project conducted slowly over a long period of time.
Much of the code for this port came from Coreboot, an open-source firmware project similar to U-Boot’s SPL in terms of features.
Also see  for information about the boot flow used by coreboot. It is similar, but has an extra postcar stage. U-Boot doesn’t need this since it supports relocating itself in memory.
 Intel PDF https://www.coreboot.org/images/2/23/Apollolake_SoC.pdf