Porting Linux to the DEC Alpha: The Kernel and Shell
Achieving the shell prompt was only the beginning of the trial, not the end. Now we had to get other utilities working so as to allow us to debug more of the system and to make it into a real, usable UNIX-like system.
I set Brian Nelson to the task of porting more utilities, starting with the “fileutils” and “shellutils” subsets of the MCC 1.0+ distribution. Meanwhile, I realized that better debugging tools would expedite the debug process, and I started thinking about implementing some kind of remote debugger support for gdb. My first implementation was under the ISP CPU simulator. The reason for this is that I could add code to ISP to examine and modify the state of the machine and communicate with a debugger without having to code a breakpoint handler into the kernel itself.
GDB has a remote debugging protocol built into it; all I needed to do was to add code to ISP to respond to gdb commands and to encode the simulated machine state for GDB's consumption. Getting this all working was only a few days' work.
While all this was happening we realized that Linux/Alpha was turning into a serious project and that we could use some help in the device driver space. While console callback drivers had served admirably to get us up and running, they were not equal to the task of supporting a production system. We recruited Jay Estabrook from the Digital UNIX group in this capacity, and he has proved to be an immensely valuable addition to our team. Within his first two weeks with the group he produced a native text-mode VGA driver and a native keyboard driver for Linux on the DEC 2000 AXP (“Jensen”) series.
Porting device drivers to Linux/Alpha presents some interesting challenges. Fortunately, many of the problems only needed to be solved once and the results would be applicable to many different drivers.
The Alpha CPU has no concept of I/O bus access; there is no Alpha equivalent of the Intel inb/outb instructions for communicating with the I/O bus. In order to implement a PCI- or EISA-based Alpha system, some sort of “glue” logic is needed to translate Alpha load/store accesses into I/O bus accesses. On systems based on the DECchip 21064 and 21164 CPUs, this glue logic is implemented in an external chipset (DECchip 21071 series). On systems based on the DECchip 21066, this glue logic is built into the CPU. This glue logic sets aside certain areas of the system physical address space for communicating with the I/O busses. To perform a bus access, one takes such information as the I/O port number and the size of the transfer and encodes it into a special memory address (This encoding is different for the different glue logic implementations). One may then load the data from or store the data to this address.
Interrupt handling is also different on Alpha systems. On most Intel-based systems, the mapping between bus IRQs and interrupt vectors is fixed and straightforward, and the hardware dispatches directly to the interrupt vector associated with a particular IRQ. On Alpha, all interrupts are dispatched by PALcode. The Digital UNIX PALcode vectors all device interrupts to a single routine. One of the arguments to this routine is a number (called the “SCB vector” for reasons I don't need to go into here) indicating which interrupt was received. Unfortunately, the mapping between bus IRQs and SCB offsets is not the same across all platforms. This means that we need extra code in the interrupt handling path to map the SCB vector back to an IRQ number. In fact, there are different versions of the mapping routine for the different platforms.
It turns out that many Intel Linux device drivers rely on the fact that the BIOS puts the device into a known state before the operating system even sees it. We discovered this when we were debugging the interrupt handling for the keyboard driver (and, later on, the SCSI driver). Apparently the interrupt controller on Intel boxes is initialized by the BIOS to trigger on the transition of an IRQ line (edge triggering) rather than on the state of an IRQ line (level triggering). We were having no end of problems with “spurious” interrupts until we added code to the CPU initialization routine to set the interrupt controller properly.
While our early decision to use the SRM console was good to get the project off the ground, it turns out that the SRM console is not the best choice for Linux. First of all, the SRM console is memory-hungry because it must implement numerous functions required by OpenVMS and Digital UNIX. Many of these functions are not needed by Linux. Second, and more serious, is the fact that the SRM console is not freely redistributable. Digital charges a substantial license fee to third parties for resale rights to the SRM firmware, as well as a per-unit charge for each copy of the SRM firmware sold by Digital or a third party. While the end-user typically never sees these charges, they do serve to raise the price of the hardware when it is sold in a Digital UNIX or OpenVMS configuration. In addition, requiring the SRM console for Linux/Alpha would present a significant burden to clone vendors who wished to build and sell Alpha systems for the Linux marketplace.
For these reasons, we decided to investigate the possibility of developing a freeware “miniloader” for Linux/Alpha. The miniloader could be much smaller than the SRM console because it needs to implement only the functionality necessary to initialize the system, load the PALcode, and load Linux. It could also be freely redistributable in source and binary form.
Unfortunately, developing a replacement for the console firmware is a non-trivial undertaking. Fortunately, however, we had help in the person of Dave Rusling from the Digital Semiconductor facility in Reading, England. Dave had much experience in low-level hardware support on the evaluation boards produced by Digital Semiconductor, and he had already done significant work for the Linux/Alpha project in the area of PCI support. He eagerly took on the task of developing the miniloader.
The miniloader consists of system initialization code, an OSF-compliant PALcode library, a bootstrap loader, and console callback routines. It presents to the bootloader and kernel an interface similar to that seen in the SRM console, with stripped-down functionality. The miniloader only implements those SRM features and callbacks that are used by Linux. As of this writing, the miniloader has succeeded in booting Linux on several models of Digital Semiconductor evaluation boards, as well as on the low-cost AXPpci/33 “NoName” motherboard and the high-performance “Cabriolet” motherboard.
While we were working on our 32-bit port, Linus was toiling away in Helsinki on his 64-bit Alpha port. We knew that we would want to cut over to using his code base at some point so that we would be in sync with the rest of the Linux community. The main question was when we would do the cutover. Our port had more demonstrable functionality earlier (e.g. device support, networking, utilities), but Linus was catching up fast. We decided to continue working on the 32-bit port for demonstration purposes while keeping track of Linus' progress, and we would cut over to Linus' code base when doing so would yield a system of roughly equivalent functionality.
This point came in March, 1995 when Linus posted a message to the linux-alpha mailing list with the subject “self-hosting linux on ftp.cs.helsinki.fi”. While one of our own major goals was to have a self-hosting Linux/Alpha system, we had not been able to realize it due to the immense complexity of porting the GNU compiler suite in our cross-development environment. Linus very neatly sidestepped the entire cross-build issue by making his Linux/Alpha system calls compatible with their Digital UNIX counterparts. Therefore, he could achieve self-hosting simply by running the compilers from his Digital UNIX system on his Linux/Alpha system.
While this self-hosting environment did not meet our criterion of being 100% freeware, it was a useful starting point. Instead of building the GNU tools in a cantankerous cross-development environment and testing them on an immature operating system, we could prototype and debug our entire development environment on a Digital UNIX system. When we were satisfied with its functionality, we could then copy it over to a Linux/Alpha system with reasonable assurance that it would work. This, in fact, is exactly how we put together the self-hosting demo that we exhibited at DECUS in May, 1995 (“Linux at DECUS”, Linux Journal issue 15, July 1995).
An operating system is much more than just a kernel, as any of the creators of Linux distributions could tell you. In order to be able to provide all of Digital's Linux/Alpha contributions to the Linux community free of charge, we necessarily had to limit the investment we made in the project. As of this writing, Digital is funding three full-time engineers, a part-time product manager, a part-time technical writer, and several loaner Alpha systems (including the Alpha systems that Linus has been using). In my project plan outline for Linux/Alpha I pointed out that Linux was unique in that we could do the project with such limited resources. Given the history of Linux, I reasoned, once the Linux/Alpha code became available, developers all over the net would add functionality and fixes. My prediction turned out delightfully true. Several people, both inside and outside of Digital, made significant contributions to the project at no cost to Digital. The result is of enormous benefit to both Digital and the Linux community as a whole.
Linus Torvald's own contributions, of course, are legendary. I mention him here because without his tireless work the project would have taken a different turn and probably would not be as successful as it is today (Linus, if you're reading this, we could use a little breathing room between releases. At least let me finish compiling one release before you turn out the next!)
Another major champion and supporter of the Linux/Alpha project is David Mosberger-Tang of the University of Arizona. He was literally the first on his block to own an Alpha-based AXPpci/33 motherboard, and he provided all of the initial patches to enable both the 32-bit and 64-bit kernels to function on that platform. He has also been a valuable resource and a second set of eyes to assist in untangling sticky problems. In addition, he has ported numerous system and utility packages that would have taken us days or even weeks to do ourselves in our spare (ha!) time.
It has been said that “any sufficiently advanced technology is indistinguishable from a rigged demo,” and this could certainly be said of the DECUS demo that we staged. While the toolset was capable of building and linking the kernel, the 64-bit C runtime library was not yet stable enough to build user utilities. Fixing this was on our “to do” list along with a lot of other things, but it turned out we didn't have to. Shortly after we released the 64-bit development tools to Digital's FTP area, Bob Manson of Ohio State University released a working 64-bit library based on our earlier 32-bit work. Bob also released several useful sets of utilities that, again, it would have taken us weeks to get around to porting on our own. He is also rumored to be working on modifying gcc to generate floating-point code that is capable of recovering from exceptions.
After showing off Linux/Alpha at DECUS, it became clear that some kind of end-user-installable distribution was needed. At that point, Linux/Alpha resembled in some ways the early days of Intel Linux: the “system” consisted of a motley collection of source and binary archives scattered over several FTP sites on different continents. Putting together a running system out of these pieces was a job only a dedicated hacker would be willing to go through with.
One difference between then and now is that now there are high-quality commercial Linux distributions (Plug-And-Play, Red Hat, and Slackware, to name just a few) that can serve as the basis for equivalent Linux/Alpha distributions. We knew, though that it would be some time before these distributions were ported and qualified for Linux/Alpha. In oder to sustain the momentum built up at DECUS, we needed some kind of Linux/Alpha distribution sooner than that. That's when we decided to embark on a project designed to become obsolete: BLADE. BLADE stands for Basic Linux Alpha Distribution Expletive (I picked the name by starting with “LAD” for “Linux Alpha Distribution” and grepping through /usr/dict/words for this combination and playing acronym games with some of the results). BLADE is a Linux/Alpha kit that can be installed without needing to build kernels or use a host development system.
BLADE was designed to be deployed quickly, and it's pretty rough around the edges. There's only one automatic installation script, called install_subset. A lot of steps that are done automatically by other distributions must be handled manually in BLADE. We provide full step-by-step instructions, though, so the user knows what steps need to be taken.
The first release of BLADE (V0.1) provided basic functionality and a development system in character-cell mode. There was no networking, no GUI, and a limited utility set. However, it was self-hosting, and it came with kernel sources and the gnu compiler suite. One could build the kernel or any utilities one desired using BLADE. In fact, we used BLADE V0.1 as our primary development system for BLADE V0.2. BLADE V0.1 was based on a modified 1.2.8 kernel and supported the AXPpci/33 only.
The second release of BLADE (V0.2) added more utilities and networking functionality. Graphics were still not available, but one could perform development and basic networking (ftp, telnet, rlogin) in character-cell mode. BLADE V0.2 was also the first release to support the Linux/Alpha Miniloader (aka MILO/Alpha) on the AXPpci/33. MILO is a drop-in system firmware replacement that allows the user to boot and run Linux without requiring the SRM console firmware. BLADE V0.2 also added a kernel boot disk for the Digital Semiconductor 275-MHz EBPC/64 evaluation motherboard. This is the fastest system to date that supports Linux. BLADE V0.2 is also based on the modified 1.2.8 kernel.
Currently under development is BLADE V0.3. BLADE V0.3 will be based on a 1.3 kernel and will add support for the X window system (see below). It should also support more system types.
Thanks to the tireless efforts of my colleague Jay Estabrook, as of this writing X11R6 is up and running on Linux/Alpha. Most of the standard libraries and client executables are in place. At the present time, only the S3 server has been ported, and it has only been qualified on a few video cards. The current plan is to let this be a sample server and solicit other parties (e.g. the XFree86 Consortium) to port other servers.
One major problem with supporting multiple video cards has to do with the on-board ROM BIOS that many cards have. This BIOS typically contains code to initialize the card and to set video modes. Unfortunately, this BIOS is nearly always written in 80x86 assembly code. To execute it on an Alpha system requires an Intel execution engine. We are investigating several strategies to provide this functionality as part of MILO/Alpha in source form, and rumor has it that our old friend David Mosberger-Tang has made good progress in this area.
As I write this, Linux/Alpha is being exhibited at UNIX Expo in New York City in all of its X-windows glory. We have ported the freeware web-browser chimera, and these systems are available for surfing the web and for connecting to remote systems on the Internet via rlogin, telnet, and ftp. In fact, on setup day several people from non-net-connected boths came by to use these systems to retrieve forgotten files from their home systems. This includes yours truly, embarrassingly enough. We needed to connect to the serial port on our PC64 system to use the ROM debug monitor, but the version of Linux/Alpha on the other system did not have kermit, cu, tip, or any other terminal emulation or serial connection program. No problem: I ftp'ed over to David Mosberger-Tang's archive at ftp.azstarnet.com and retrieved the Linux/Alpha version of kermit. We were in business in a few minutes.
In short, Linux/Alpha is starting to feel like a real Linux system!
Despite our great progress, much work remains to be done on Linux/Alpha:
As mentioned above, we need to deploy some sort of BIOS-emulation facility so that we can execute the proprietary initialization code on some expansion cards. While initial code exists and works, it does not support the real-mode 32-bit instructions that are used in the BIOSes of some cards.
We need to tackle the great unsolved problem of floating-point exception handling. Programs that are floating-point intensive are not likely to work until this is done.
We need to write a character-cell driver and an X server for the TGA graphics adapter that is provided with Multia and several other Digital Alpha systems.
We desperately need shared libraries! As of this writing, the statically-linked executables in Linux/Alpha are rather large (around 200Kb for a typical utility, several megabytes for the X server). Shared libraries will decrease both disk space requirements and virtual-memory usage.
We need to work on compiler optimizations. The Alpha support in gcc does very good optimizations in some places, not so good in others. In addition, the compiler does not yet take advantage of Alpha's multiple-instruction issue feature. This feature allows more than one instruction to be issued per clock cycle, but only certain combinations are allowed. By carefully rearranging the instructions in the executable, one can take advantage of this feature and achieve significant performance improvements.
All in all, we are excited about the future. Linux/Alpha, even in its relatively primitive state, feels like a real Linux system. Addressing the above areas can only make it better!
Jim Paradis works as a Principal Software Engineer for Digital Equipment Corporation as a member of the Alpha Migration Tools group. Ever since a mainframe system administrator yelled at him in college, he's wanted to have a multiuser, multitasking operating system on his own desktop system. To this end, he has tried nearly every UNIX variant ever produced for PCs, including PCNX, System V, Minix, BSD, and Linux. Needless to say, he likes Linux the best. Jim currently lives in Worcester, Massachusetts with his wife, eleven cats, and a house forever under renovation. He can be reached via e-mail at firstname.lastname@example.org and on the WWW at www.iii.net/users/jrp.html