The ARM SoC landscape from Genode's perspective
We get repeatedly asked for our opinion about ARM system-on-chips (SoCs) suitable for the use with Genode. Even though Genode supports the ARM architecture since 2009, the answer has remained anything but simple. This article presents constraints faced by a small player attempting to realize an ARM-based product that deviates from the beaten track of using a Linux-based OS. It should not be regarded as ground truth but rather as the subjective perspective of Genode Labs.
X86 PC platforms are highly standardized and generally well documented. For example, all PC products - regardless of their form factor and set of peripheral devices - are based on general-purpose interfaces like PCI, USB, and ACPI, thereby providing generic means to enumerate and manage devices. Low-level devices like interrupt controllers, system timers, and IOMMUs are standardized too, even between competing CPU vendors like AMD and Intel. This situation is to the great benefit of OS vendors, which can target a vast variety of platforms with the same fundamental drivers and protocols. Even though the commonality of PC platforms ends at the level of peripheral drivers (wireless, networking, graphics), the common ground shared by all PC vendors is substantial. This common ground is generally documented, publicly discussed, open-source friendly, and well understood.
The ARM ecosystem is in stark contrast to the x86 world. Whereas the x86 chip market is dominated by very few vendors (Intel, AMD, VIA), ARM's business model is based on IP-core licensing to a vast number of chip (SoC) vendors on liberal terms. Even though SoC vendors use the same CPU core and ISA, their common ground ends at this point. Low-level concepts like booting the system, memory-layout detection, device discovery, cache management, the bootstrapping of secondary CPU cores, device power controlling, bus controllers, or interrupt routing are not standardized. The great benefit of this approach is the ability of SoC vendors to innovate independently from each other, and to target their chips extremely close to specific markets. For example, energy efficiency has been a major concern for low-power embedded systems and mobile devices, which prompted SoC vendors to explore vastly different energy-saving approaches. Energy efficiency has seen dramatic improvements in the ARM world that could not be matched by Intel, even given Intel's lead in manufacturing process technology.
This benefit, however, comes at the cost of fragmentation. In fact, each ARM SoC must be regarded as an island. ARM has tried to counter this fragmentation by extending their offering to supplemental IP cores such as a generalized interrupt controller (GIC), generalized SMP support and core-local timers, or generalized cache management. But since these supplements are not mandated but rather optional, the fragmentation could not be reverted. For example, the Broadcom SoC as used by the Raspberry Pi 3 still employs a custom interrupt controller. Besides the already fragmented SoC ecosystem around ARM, there are SoC vendors like Qualcomm that merely use ARM's instruction set architecture (ISA) but use to develop a completely custom implementation of the CPU core. For OS vendors, this situation is unfortunate because the investment required for supporting one SoC family can target only a small fraction of the ARM device market.
ARM-based mass-market products are most commonly based on open-source technology like the uboot boot loader and the Linux kernel. However, the operating-system support for ARM SoCs is generally not developed as a concerted community effort. Instead, the development costs are carried by the respective SoC vendors. There are dedicated Linux kernel teams at Samsung, Qualcomm, Mediatek, etc. As enforced by the GPL license of the Linux kernel, the work of those teams is eventually published, or even integrated into the mainline Linux kernel. But the development process usually remains behind closed doors of the respective SoC vendors.
This results in the following contradiction: In the perception of the public, ARM-based products are perceived as "open" because Linux is readily available for those products and the Linux-based OS can often be customized to a great degree. However, since the OS kernel and driver developments are pursued local to the SoC vendor, there is no incentive for the SoC vendor to publicly document the inner workings of the SoC. After all, the driver developers can simply call the chip designers in house when questions arise. For developers outside the organization, however, the operation of the SoC remains opaque. Often, the SoC vendor's source code is the only form of "documentation", which, however, is almost void of architectural clues.
The prime example of this contradiction is the Raspberry Pi. On the one hand, it is the favorite platform for tinkerers. On the other hand, there exists almost no hardware documentation outside the reverse-engineering efforts of a community of die-hard enthusiasts.
This status quo is not accidental but fostered by SoC vendors. The business of SoC vendors is not solely the sale of chips but also consulting services for OEMs! Here, we speak of an OEM as a company that integrates a SoC as a building block into a physical device. SoC vendors like Qualcomm or Mediatek provide their SoCs to the OEMs as so-called turn-key solutions that can be integrated into the physical product without any know-how needed about the inner working of the building block. The building block comprises the entire stack from the chip, over the OS kernel, SoC-specific drivers, up to the Android user land. The holistic know-how of the entire stack thus remains local to the SoC vendor. The OEM merely has the means to customize the Android user experience (on top of the turn-key solution) and has all information needed to physically integrate the SoC on a custom board (at the bottom of the turn-key solution) - usually in the form of publicly available reference-board schematics and Gerber files. Recently, SoC vendors try to provide additional hooks for customization in the form of (proprietary) APIs for their respective trusted execution environments (TEE).
However, for any customization of a turn-key solution beyond these blessed hooks, the OEM ultimately depends on the consultancy business of the SoC vendor. Consequently, the most prolific SoC vendors are in an extremely strong position and have no business incentive to weaken the exclusivity of their know how by publishing hardware documentation.
The power dynamics of the ARM SoC market described above put independent OS vendors that develop alternatives to Linux into a difficult situation because their goals are not aligned with the SoC vendors' interests. From what we learned from reaching out to various OEMs, even OEMs have no real leverage because - at the end of the day - they are mere consumers of SoCs as opaque turn-key solutions with little in-house expertise about SoC-related OS-kernel development. The SoC's inner working is not the OEM's business.
For the development of an ARM-based Genode product, there exist two principle options.
First, one may approach a SoC vendor in the role of an OEM and commission the SoC-specific OS development to the selected SoC vendor. This would be consistent with the SoC vendor's business model. That said, it seems unlikely that a small player would be able to present a strong-enough business case to a SoC vendor that normally deals with mass-market OEMs, or the costs may be prohibitive. Nevertheless, even if taken, this option has several downsides:
The operation of the hardware will remain opaque, which may impede the trustworthiness of the device.
One ultimately has to enter a dependency relationship with the SoC vendor regarding the system-software stack.
The consulting terms with the SoC vendor must be discussed with a large - and from a European's perspective foreign - corporation.
Genode would be an entirely new territory for the SoC vendor, which bears the risk of friction and high costs.
As the second option, one could work with a SoC vendor that is primarily focused on selling chips - not turn-key solutions - and thereby has a natural interest to make the chips useful to a broad community of customers by offering documentation. Such vendors can be found in industrial and automotive markets that demand a high level of transparency, long-term supportability, and freedom for customizations.
In particular NXP (formerly Freescale Semiconductor) falls into this category by offering an ARM SoC family (i.MX) geared towards multi-media products such as set-top boxes or in-car entertainment systems. The i.MX SoC family is also notably popular in e-book readers. The Genode OS framework supports i.MX since 2012. At that time, we explored ARM TrustZone and secure boot mechanisms and were pleasantly surprised about the openness of the i.MX SoC and the good state of documentation, which covers even NXP's custom high-assurance boot mechanism and TrustZone. Furthermore, NXP appears to us as an attractive partner because it is a European company that lives and breaths the small-business economy as present among, e.g., German industrial or automotive supply chains.
With this background, we successively consolidated Genode's ARM support towards the i.MX family, after trying to accommodate Samsung Exynos, Texas Instruments OMAP, and Broadcom SoCs in the past. The i.MX family entered the spotlight lately with the Librem 5 project, which is a crowd-funded effort to manufacture a completely open-source smart phone. The Librem 5 project selected i.MX SoC likely for the same reasons that attracted us. Other notable users of the SoC family are the USB Armory and the MNT reform laptop.
Granted, NXP i.MX SoCs may rightfully be criticized as not being bleeding edge when compared to the latest high-end smartphone chips. But from our perspective, this SoC family is the most attractive. Hence, when working on Genode's 64-bit ARM support, or device drivers, or ARM virtualization, we'll keep our focus on the i.MX series.